Predictive Models for Inventory Optimization: a machine learning application for demand forecasting at a construction supplies distributor

Predictive Models for Inventory Optimization: a machine learning application for demand forecasting at a construction supplies distributor | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Predictive Models for Inventory Optimization: a machine learning application for demand forecasting at a construction supplies distributor Diocélio Dornela Goulart, Rodrigo Baroni de Carvalho, Mariana Almeida Henriques, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8502995/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 31 Mar, 2026 Read the published version in Future Business Journal → Version 1 posted You are reading this latest preprint version Abstract Inventory management is a critical task for distribution companies, as forecasting errors can lead to excess inventory, higher storage costs, or stockouts. This practice & policy paper describes the development and deployment of a machine learning-based predictive model designed to optimize inventory management for a medium-sized Brazilian construction supplies distributor. Using the Amazon Forecast and Amazon SageMaker Canvas platforms, the study demonstrated how advanced demand forecasting techniques reduce stockouts and enhance operational efficiency. The research compared various predictive algorithms, assessing their performance with metrics such as RMSE, MAPE, MASE, and WAPE. The results showed that the implemented model achieved approximately 99.31% forecast accuracy, offering significant benefits like fewer stockouts, optimized inventory levels, and improved purchasing planning. This practice study provides technical and managerial recommendations for companies seeking to predict their inventory needs accurately, demonstrating the machine learning capabilities for Supply Chain Management. Predictive models Artificial Intelligence Machine Learning Inventory Management Supply Chain Management (SCM) Figures Figure 1 Figure 2 Figure 3 Figure 4 1. CONTEXT Efficient inventory management is a vital challenge for distribution companies, as it directly affects customer satisfaction, operating costs, and overall business profitability. In the construction supplies industry, which features a diverse range of products and seasonal demand fluctuations, forecasting becomes even more difficult (Syntetos et al., 2016 ). Forecasting mistakes can result in excess inventory, raising storage costs and the risk of obsolescence, or lead to stockouts (product shortages), causing lost sales and customer dissatisfaction (Fildes et al., 2009 ). In this context, the objective of this practice paper is to develop and apply machine learning (ML) techniques to improve inventory management. The focus was on predicting product demand more accurately, lowering costs, and maximizing resource usage. The practical application was centered on Casa Cardão, a medium-sized Brazilian distributor operating in both wholesale and retail channels for construction supplies. This business context highlights the significant logistical and operational challenges, justifying the implementation of advanced predictive techniques to reduce disruptions and lower inventory holding costs. As is typical in medium-sized organizations, the company has a lean IT (Information Technology) structure and lacks analysts with advanced knowledge of Data Science and AI (Artificial Intelligence). Many available software solutions require significant investment and specialized staff. Currently, there is a shortage of these professionals in Brazil, so advanced applications of technologies such as AI and data analytics remain out of reach for many medium-sized Brazilian companies. The complexity inherent in Supply Chain Management (SCM) and demand forecasting, especially in scenarios with intermittent or irregular demand (Tian et al., 2021 ), justifies exploring more advanced methods than traditional ones, which are often limited in their ability to capture nonlinear patterns and multiple influencing variables (Makridakis et al., 2018 ). The use of ML in SCM has emerged as a promising approach, providing powerful tools for analyzing large datasets and identifying complex patterns (Tirkolaee et al., 2021 ). This practice paper outlines the process of developing, implementing, and evaluating ML-based predictive models using Amazon Web Services (AWS), specifically Amazon Forecast and Amazon SageMaker Canvas. This practical study demonstrates the feasibility and benefits of using these cloud-based technologies to improve decision-making in inventory management, supporting the digital transformation and operational efficiency of an organization with a less structured IT department. The study also offers a reproducible model for Small and Medium-sized Enterprises (SMEs), enabling the use of machine learning for inventory management at lower costs. The study was developed to create an ML-based solution that integrates internal data from ERP (Enterprise Resource Planning) and CRM (Customer Relationship Management) business systems, along with external data (public data and partner databases), in a Data Lake environment. This practical application aims to provide more accurate forecasts, supporting informed decision-making and reducing waste (Van Steenbergen & Mes, 2020 ; Naik et al., 2022 ). The project was selected and received support from a specific innovation call for proposals of the National Council for Scientific and Technological Development (CNPq), an entity associated with the Brazilian Ministry of Science, Technology and Innovation (MCTI). The purpose of the call for proposals was to promote postgraduate research collaborations between industry and universities. The research project team comprised the project coordinator (a professor in Management), a PhD researcher in Management, and two undergraduate research scholarship students in Software Engineering. Other partners are also involved in the project: i) a Brazilian business school, which provided access to its customer database to identify potential opportunities for tests and proof of concept; ii) a Paraguayan-Brazilian technological park associated with one of the largest hydroelectric power plants in the world, which supported submitting the research project to the FINEP (Brazilian Funding Agency for Studies and Projects, also associated with MCTI) call for innovation proposals, potentially enabling additional resources; iii) a cloud computing company based in São Paulo, which hosted the necessary resources for the project's development and infrastructure; iv) AWS - Amazon Web Services, which supplied an environment for testing over 12 months, enabling the execution of the POC - Proof of Concept; and v) a consulting and development firm in Minas Gerais, which provided access and resources for AI development tools such as ChatGPT Pro, Manus AI, and Base44. The overall goal of the project was to research, identify, and develop ML-based models and technological solutions that can serve as predictors and automate data collection and analysis using Artificial Intelligence. The project aimed to create an intuitive interface for managers that enables accurate simulations and forecasts for inventory management. A specific goal is to make the application development easy to duplicate at a low cost, using open-source software or no-code or low-code platforms (with little or no coding or development), enabling its adoption by companies with reduced IT budget and personnel. Beyond this introduction, the next section of this report comprises the scientific framework that guides the study's development. Section 3 details the application development and reports the experiment conducted, presenting the achieved results, and discussing the application of the tools. Section 4 evaluates the study across the following dimensions: a) how well it aligns with related fields and areas of Management; b) the potential environmental and stakeholder impacts of the project; c) the feasibility of applying the proposed technology and processes; d) the innovation and level of knowledge involved in product development; and e) the complexity of the network and relationships necessary for its design and execution. Section 5 discusses the results, while Section 6 concludes the practical with recommendations for future work and addresses its limitations. 2 THEORETICAL BACKGROUND This section offers an overview of the theoretical framework that guided the application development. Inventory management plays a key role in logistics and operations management, balancing inventory costs with the risk of product shortages. Accurate demand forecasting is essential for effective inventory control (Hyndman & Athanasopoulos, 2021 ). Traditional forecasting techniques, such as moving averages, exponential smoothing, and ARIMA (Autoregressive Integrated Moving Average) models, are commonly used but may have limitations when handling complex, seasonal, or multi-variable data (Bowerman et al., 2005 ). Demand in the retail and distribution industry, such as construction supplies, often exhibits characteristics like intermittency (periods without demand) and seasonality. Traditional models may struggle to capture these patterns accurately (Syntetos et al., 2016 ; Tian et al., 2021 ). Additionally, external factors like promotions, economic conditions, changes in tax laws, political events, competitor activities, and even weather can influence demand, requiring models that can incorporate multiple sources of information. Predictive models using Artificial Intelligence (AI) for inventory management offer significant advantages over traditional methods. ML algorithms process many factors and identify complex patterns that traditional techniques overlook (Ampazis, 2015 ), resulting in more accurate demand forecasts and enabling quick adaptation to changing market conditions (Likhar et al., 2023 ). As a subfield of AI, ML provides alternative and complementary methods to traditional statistical forecasting techniques. ML algorithms can identify complex patterns from historical data, including nonlinear relationships and interactions among multiple variables, without relying on strict statistical assumptions (Makridakis et al., 2018 ). Product demand forecasting is a vital technique for inventory management, allowing companies to optimize their stock levels, lower costs, and enhance customer service. One of the machine learning methods that has gained prominence in this field is the use of Random Forest for forecasting (Vairagade et al., 2019 ). Other machine learning algorithms have been effectively used for demand forecasting, including Artificial Neural Networks (ANN), Support Vector Machines (SVM), and decision tree-based methods like Random Forest and Gradient Boosting (Tirkolaee et al., 2021 ; Vairagade et al., 2019 ). These approaches are especially valuable when large amounts of data (Big Data) are available and demand is driven by a complex mix of factors (Carbonneau et al., 2009 ). Multiple studies show that this technique is highly effective for inventory management. Van Steenbergen and Mes ( 2020 ) introduced DemandForest, an innovative approach that combines K-means clustering, Random Forests, and Quantile Regression Forests to forecast demand patterns for new products. This method was tested with real data sets and outperformed several reference methods, leading to around 15% savings in inventory (Van Steenbergen & Mes, 2020 ). Vairagade et al. ( 2019 ) compared various demand forecasting models, including Random Forest and Artificial Neural Networks. They found that Random Forest was more accurate, making it better suited for predicting product demand in complex supply chains. On the other hand, Tian et al. ( 2021 ) tackled intermittent demand forecasting, a common challenge in retail, and introduced a combined approach that accounts for current inventory and past sales. This method, known as the Markov-combined method (MCM), proved to be more accurate than traditional techniques, leading to significant improvements in inventory management. Punia et al. ( 2020 ) introduced another combined approach that utilizes Long Short-Term Memory (LSTM) and Random Forest models to capture complex temporal and regression relationships. This method greatly outperformed other forecasting techniques in accuracy and robustness within multichannel retail settings. Naik et al. ( 2022 ) examined sales forecasting in the food industry using Random Forest Regression. This model was notable for its high accuracy compared to other machine learning methods, making it ideal for predicting sales and optimizing inventory management. In summary, the reviewed literature highlights three key pillars that guided the application design: (i) the superiority of ML approaches over traditional statistical methods in volatile demand scenarios, due to their ability to capture nonlinear relationships and multiple variables (Makridakis et al., 2018 ; Hyndman & Athanasopoulos, 2021 ); (ii) the need to combine algorithmic forecasting with human judgment to enhance operational decision-making, as supported by studies on managerial adjustments in forecasting (Goodwin & Fildes, 2019); and (iii) the importance of explainability mechanisms that clarify the most influential factors behind the forecast, thereby increasing managers' confidence in the system's recommendations (Sokol & Flach, 2020 ). These findings informed the selection of AutoML platforms capable of integrating various algorithms (e.g., DeepAR+, Random Forest) and robust evaluation metrics. As detailed in the next section, these theoretical guidelines were put into practice in the application development, defining the data pipeline, model selection criteria, and human-machine interaction protocol that form the foundation of the implemented solution. 3. APPLICATION DEVELOPMENT This section describes the process of application development, detailing the IT tools, the underlying architecture, and the performance assessment. The context of the application was Casa Cardão, a hundred-year-old Brazilian distributor operating in both wholesale and retail channels for construction supplies. It employs around 220 staff members and 200 sales representatives. It runs two regional distribution centers covering 12,000 m² of storage and markets a portfolio of over 10,500 SKUs (Stock Keeping Units), with strong demand for 150 items from the ABC curve. In 2023, the company generated R $ 310 million (about US $ 60 million) in revenue, holding a 4% market share in Minas Gerais state within the cement, mortar, and ceramic tile segments (Sinduscon-MG, 2024 ). 3.1 Development tools Developing and deploying ML models can be complex and demand substantial computational resources. Cloud platforms like Amazon Web Services (AWS), Google Cloud Platform (GCP), and Microsoft Azure provide managed services that potentially make this process easier. In this study, two AWS tools were utilized. Amazon Forecast is a fully managed service that uses ML to deliver highly accurate time-series forecasts. It automates much of the ML process, from data loading to model training and deployment, automatically testing various algorithms (including ARIMA, ETS, Prophet, DeepAR+, and others) and choosing the best one for the data provided (Amazon Web Services, 2025a ). The second tool, Amazon SageMaker Canvas, is a no-code visual interface that enables business analysts to build ML models without writing code. It streamlines data preparation, model training, and forecast generation, making ML more accessible to non-expert users (Amazon Web Services, 2025b ). These platforms facilitated the efficient experimentation and implementation of predictive models in the Casa Cardão environment. The project resulted in a dashboard or visual layer that displays structured data collected through an ETL (Extract, Transform, and Load) process using AWS SageMaker Data Wrangler, covering both external and internal data. This dashboard includes a series of graphs showing inventory forecasts (Fig. 1 ). With a prompt feature available on the generative AI-powered dashboard interface, the application provides detailed explanations of the methodology and graphs used, making the tool more accessible and understandable for managers. Note Screen of the application developed by the research group. Additionally, the interface enables managers to run simulations as needed, combine different data sources, and leverage this information to enhance inventory management. The project is delivered in a tailored manner, using company-specific data along with information from public sources. The conversational management dashboard can be continuously improved by adding new data sources to develop a more robust and effective predictive model for inventory management. The developed product is a demand prediction system that utilizes predictor results generated through AWS SageMaker Canvas, featuring an interactive dashboard that provides product demand forecasts across different scenarios. These scenarios include optimistic, neutral, and pessimistic. In the optimistic scenario, inventory levels are low, and the forecast accounts for seasonality and promotions. In the pessimistic scenario, inventories are high, and there is no forecasted demand for the products. The system's features include time series analysis, integration of internal and external data, customized simulations, and a console explaining the methodology and graphs used. The project can help the company by lowering costs related to excess inventory or product shortages and by making inventory management more efficient, which allows for more confident decision-making. Specifically, the goal was to create AI-based inventory management software that could predict the necessary inventory levels with 90% accuracy, and this target was successfully achieved with a 99.31% accuracy rate. Measurable objectives include reducing excess inventory by 20% and shortages by 15% within one year of implementation. Several machine learning techniques were tested and used to develop the project, including AWS's proprietary algorithm, DeepAR. It is based on Recurrent Neural Networks (RNN) with an LSTM (Long Short-Term Memory) model and is often used to generate probabilistic forecasts for multiple series (various products). DeepAR demonstrated the best accuracy among the models available on the AWS platform, such as DeepAR+, CNN-QR, Prophet, ARIMA, and ETS. 3.2 Technology architecture and phases Application development began in August 2024, and by April 2025, the proof-of-concept and validation of the results were completed. The system was fully operational within the organization by July 2025. Figure 2 illustrates the technological architecture used to develop the application, outlining its technology pipeline organized into four functional layers in the AWS environment. In the ingestion layer, transactional data extracted from Casa Cardão's ERP and CRM systems, along with external data such as weather, holidays, and macroeconomic indices, are stored in a data lake on Amazon S3. The preparation layer uses SageMaker Data Wrangler for ETL processes, ensuring the cleaning, aggregation, and enrichment of time-series data. Next, the modeling layer employs Amazon SageMaker Canvas for exploratory AutoML and the Forecast service (DeepAR+, CNN-QR, Prophet, among others) for final training, automatically selecting the algorithm with the lowest back-test error (Amazon Web Services, 2025a ; 2025b ). The generated forecasts are stored in a versioned S3 bucket, while AWS Lambda functions manage the daily model updates, providing serverless scalability. Finally, at the application layer, the results are fed into an interactive dashboard (developed in React using the base44 platform) that displays P35 (pessimistic), P50 (neutral), P60, and P80 (optimistic) scenarios, along with a conversational interface powered by GPT O3-mini. In the pessimistic scenario, there is little demand and a lot of stock; in the optimistic scenario, there is a lot of demand and little stock. In the neutral scenario, there is no apparent trend in demand, and the stock is balanced. This modular and fully managed workflow enables rapid iteration, cost efficiency, and seamless integration between algorithmic output and human decision-making. A set of requirements, categorized as technological, functional, and legal, was considered for the project's development. Technologically, an adequate IT infrastructure was necessary to support the collection, storage, and processing of large data volumes. AWS was selected for its resources, widespread availability, and ease of implementation. Functionally, the developed interface—dashboard—is intuitive and user-friendly, capable of integrating with results generated by AWS and providing scenario analysis features for users. Legally, the project complies with data protection and privacy regulations and does not handle any sensitive data. The stakeholders involved included both internal and external parties. Internally, the research team and IT team at Casa Cardão, along with the project manager at the cloud computing company, were essential. Externally, AWS served as a key technology provider. AWS provided the credits and resources needed for testing through its cloud computing and AI infrastructure. The Proof Of Concept (POC) was conducted at Casa Cardão, providing a real environment to test and assess the technology's effectiveness. The application development was structured into five distinct phases: Planning, Development, Testing and Validation, Implementation, and Evaluation and Improvement. The Planning phase included defining requirements and scope, as well as a detailed project schedule to ensure clarity regarding the overall objectives. In the Development phase, the predictor and an interactive dashboard were built. Next, during the Testing and Validation phase, software testing, predictive model validation, and POCs were conducted. These first three phases occurred between August 2024 and April 2025. Subsequent phases include Implementation, with implementation in the company, user training, and initial support; and finally, Evaluation and Improvement, which involves performance monitoring, user feedback collection, and continuous improvement. These last two phases were completed by July 2025. Figure 3 details each of these phases. The main challenge was to improve the accuracy of demand forecasting at Casa Cardão to enhance inventory management. The initial phase involved gathering historical sales data from the company, covering a relevant period (usually 3 to 5 years, depending on availability and detail) to identify seasonal patterns and trends. The data included, at a minimum, product identifiers (SKUs), transaction dates, and quantities sold. Additional information, such as details on promotions, holidays, or product characteristics (metadata), was considered to improve the models. The collected data was cleaned and preprocessed. This procedure involved handling missing values, aggregating the data to the required forecast level (e.g., daily or weekly sales per SKU), and identifying and addressing outliers as needed. The target time series (sales per SKU per period) and related datasets (item metadata, related time series like promotion data) were created. The Amazon Forecast and Amazon SageMaker Canvas platforms were used to develop and train the predictive models. The prepared data was uploaded to the service. Amazon Forecast automatically trained and evaluated multiple time-series forecasting algorithms, including DeepAR+, CNN-QR, Prophet, ARIMA, and ETS. The service selected the best-performing algorithm based on predefined evaluation metrics (back testing). Predictors (trained models) were generated for the SKU set. However, unfortunately, in the second half of 2024, AWS discontinued the platform. A complementary or alternative approach was explored using SageMaker Canvas as a natural replacement for Amazon Forecast. The data was imported into the visual interface, and predictive models focused on time-series forecasting were built using the platform's automated features, enabling users with less coding expertise to conduct feasibility assessments quickly. 3.3 Performance Evaluation The developed model's performance was assessed using standard time-series forecasting metrics, with back testing conducted on AWS. The following metrics were considered: RMSE (Root Mean Squared Error), which gauges the average size of errors and penalizes larger deviations; MAPE (Mean Absolute Percentage Error), which expresses errors in percentage terms relative to actual demand and is sensitive to low values; MASE (Mean Absolute Scaled Error), which compares the model's performance against a naive model and is especially useful for intermittent data; and WAPE (Weighted Absolute Percentage Error), which adjusts errors by total demand, providing greater robustness in scenarios with items of varying sales volumes. The WAPE metric was chosen as the primary indicator for overall accuracy because it best reflects the impact on total sales volume, making it particularly relevant in the retail context. After selecting the best-performing model(s), demand forecasts were created for a relevant future period (e.g., the next 4–12 weeks). These forecasts were then provided to Casa Cardão's Commercial Department, through reports and the dashboard, to aid inventory replenishment and purchasing planning decisions, as well as to validate the results obtained. The implementation of predictive models using Amazon Forecast and SageMaker Canvas at Casa Cardão yielded significant results. A comparative analysis of the different algorithms trained by the AWS platforms showed that ML-based models, such as DeepAR+ (a recurrent neural network algorithm from Amazon), often outperformed traditional statistical models (such as ARIMA and ETS) for most of the SKUs analyzed, especially those with more complex or seasonal demand patterns. To demonstrate the model's performance on a high-turnover item, the SKU "Amanco 6 m × 100 mm drain pipe" was selected, classified as "A" for value and "B" for volume in the Casa Cardão portfolio. The historical data comprised 16 months of daily records (July 2023 – November 2024), totaling 11,074 observations, with 80% used for training and 20% for backtesting. The forecast horizon was set to 60 days. In the optimistic scenario (P80), chosen because the factory delayed passing on the exchange rate increase and kept prices stable despite the US dollar’s appreciation, the DeepAR + algorithm estimated 2,340 units for December 2023, while actual sales were 2,255 units, resulting in a WAPE of 0.7%. The deviation of less than 1% reinforces the model’s accuracy and highlights the synergy between statistical forecasting and managerial judgment in selecting the appropriate quantile for market conditions. Developed through optimization and selection by Amazon Forecast, the final predictive model achieved an overall forecast accuracy of approximately 99.31%. This metric was determined using WAPE, showing that the weighted average absolute error was about 0.69% of the total demand during the evaluation period. The flowchart (Fig. 4 ) illustrates the entire cycle of using the predictive demand model, outlining seven macro-stages integrated within a cloud computing environment. The process begins with the automated extraction of transactional data from the company's ERP and CRM systems, along with external variables—macroeconomic indicators, holiday calendars, and weather data—which are stored in a data lake on Amazon S3. Next, using SageMaker Data Wrangler, the ETL pipeline runs, responsible for cleaning, aggregating overtime, and enriching historical data to ensure its quality for modeling. In the third step, this structured data feeds into SageMaker Canvas, where multiple time series algorithms are automatically trained and validated through back testing; the model with the lowest weighted absolute error (WAPE) is then versioned as the active predictor. The next step involves periodically generating forecasts, which are stored in a separate S3 bucket and simultaneously sent to a relational database via AWS Lambda functions for real-time access. In the fifth step, these forecasts are used by an interactive dashboard built with React (an open-source JavaScript library for building user interfaces), where the manager selects P50, P60, or P80 scenarios and reviews the key factors influencing each forecast horizon. The resulting decisions—such as adjusting orders, promotions, or inventory policies—are fed back into the system in the sixth step by recording the actions taken and their outcomes, forming an operational feedback database. The seventh step involves continuous accuracy monitoring: a daily verification process compares actual sales with predictions; if the WAPE exceeds the 5% threshold, a new retraining cycle is automatically initiated, ensuring adaptive learning and predictive robustness over time. This modular and coordinated chain offers scalability, data governance, and alignment between algorithmic intelligence and managerial judgment, which are essential for the successful adoption of demand models in small and medium-sized enterprises. Note Flowchart to be used during the inventory replenishment process. During the POC, the main results observed with implementing predictive models showed significant potential benefits for the business, especially the reduction of stockouts, enabled by more accurate forecasts that improved replenishment planning. Additionally, inventory levels were optimized by adjusting safety stocks and reorder points, leading to better allocation of working capital through balancing availability and excess inventory. Purchasing planning also improved, with forecasts providing strategic support for supplier negotiations and greater alignment among procurement, inventory, and demand. The adoption of AWS platforms helped streamline the development and implementation of models, enabling the team to focus on analyzing results and integrating forecasts into business processes, despite initial limitations in deep machine learning. 4. ANALYSIS OF THE RESULTS This section provides detailed information on key aspects of the application, including its compliance with management areas, its influence on various organizational and social aspects, and its usefulness in different business environments. Additionally, it examines the advancements brought by integrating predictive models with ML tools, emphasizing how these innovative solutions can change decision-making and improve processes. Lastly, it discusses the complexity and importance of these models for small and medium-sized businesses. 4.1 Adherence The application of predictive models and inventory management is closely connected to various areas of Management and related disciplines, with a special focus on Management of Information Systems. Predictive models align with theme 1 on Decision Making (theories, modeling, support systems, and technologies), promoting tools that support and facilitate decision-making in complex environments. Additionally, the application relates to digital transformation and innovation, enabling greater efficiency in inventory and operations management. The application also engages with Big Data, Data Science, and AI for Strategic Intelligence, as ML applications require large volumes of data. Besides technology and information, the application links with other Management fields such as Operations and Logistics, Innovation, Marketing, and Finance. 4.2 Impact Developing a machine-learning-based project to predict product demand can significantly improve the way companies manage complex inventories. This influence reaches various parts of the organization, community, and local area, helping to improve operational efficiency, cut costs, and boost customer satisfaction. The application aimed to promote AI adoption among small and medium-sized enterprises (SMEs) through a range of strategies. First, the application focused on providing user-friendly, pre-made AI tools tailored to meet the specific needs of SMEs, which can be easily integrated into existing systems and workflows. This feature reduces the need for specialized AI technical expertise. Additionally, the solution automates routine and repetitive tasks, allowing companies to allocate time and resources to activities of greater strategic importance, thereby enhancing operational efficiency and reducing costs. The present application can positively impact the environment by preventing the loss of perishable items and reducing unnecessary consumption for companies in the food supply chain, making resource use more efficient and helping to lower carbon emissions. For example, it can help companies better manage their inventories and supply chains, preventing material waste through demand forecasting, delivery route optimization, and real-time inventory monitoring. The application can also have a significant positive social impact by promoting greater efficiency and growth for companies. The automation of routine tasks and the improved decision-making enabled by AI can boost business productivity. 4.3 Applicability The replicability of this type of project in other contexts is high due to the adaptable nature of ML algorithms. Small and medium-sized (SMEs) businesses can also use cloud-based ML services such as AWS, Google AI, and Azure, which offer scalable and low-cost solutions. Especially considering SMEs, considerable potential for the application of the proposed model is observed since low-code platforms do not demand highly skilled teams in machine learning. ML algorithms can be integrated with ERP systems and other existing business management platforms, enabling companies to utilize their historical and real-time data to train models and enhance forecast accuracy. The scalability of these models is a key benefit, as they can adapt to the data volume and complexity of inventory operations. ML's capacity to detect intricate patterns and generate precise predictions greatly reduces human error and streamlines inventory and replenishment processes, especially in business environments where demand is volatile and unpredictable. ML algorithms are adaptable enough to function in different settings and can be tailored to manage inventory across various industries such as retail, manufacturing, and distribution. Thorough documentation of models, training procedures, and integrations supports replication in new environments without having to start from scratch. The proposed application is a modular solution, enabling specific components to be reused or adapted for various applications, such as modifying a demand forecasting model to predict sales across different markets or products. The potential market for the project is extensive, as any company with critical inventory management needs can benefit from predictive models. In the industrial sector, these models can forecast demand for specific products, optimize production, and reduce excess inventory. Additionally, they can identify products at high risk of becoming obsolete, enabling preventive measures, and predict equipment failures to facilitate maintenance and avoid production downtime. Distributors and wholesalers can utilize these models to optimize inventory levels, prevent shortages, and minimize storage costs. They can also forecast demand for seasonal products to ensure adequate stock and identify slow-moving items to inform discontinuation decisions. In retail, predictive models help forecast demand for particular products in specific stores, preventing shortages and maximizing sales. They also identify high-turnover products to optimize store displays and predict the end of product life cycles, aiding promotions and reducing waste. For service companies, predictive models help forecast demand for specific services, optimize resource allocation, and prevent overload. They can identify customers at high risk of default, enabling preventive measures, and predict equipment maintenance needs to avoid service interruptions. 4.4 Innovation The level of innovation of the application is justified by making it easier for SMEs to use ML to develop predictive models that forecast product demand. In demand forecasting, ML can analyze historical data on sales, prices, promotions, and other factors to identify patterns that help predict future demand. The use of ML for demand forecasting offers several advantages over traditional methods. First, it provides greater accuracy by considering a wide range of factors and identifying patterns that are invisible to the human eye. Additionally, machine learning is faster, enabling companies to make decisions quickly and respond promptly to market changes. An analysis of predictive model functionality shows that this feature is missing in current enterprise resource planning (ERP) systems, which only generate reports based on past data. Likewise, business intelligence (BI) systems only display historical data, not data for predictive analysis. Although specialized systems that cross-reference data and offer predictive models exist, their high development costs make this technology inaccessible to small and medium-sized enterprises (SMEs). Implementing such technologies requires experts like data scientists, AI specialists, and cloud computing professionals, highly paid professionals that many SMEs do not have. Carvalho et al. ( 2023 ) highlight the challenge of finding these professionals in Brazil, along with the shortage of trainees. The application described in this practical paper aims to address these gaps by offering a simple, user-friendly interface that provides advanced computational resources to SMEs. This feature will enable these companies to leverage advanced AI to solve major business problems, democratizing Data Analytics and making it accessible to more companies that might not afford high implementation costs. The solution's intuitive and straightforward interface enables managers without extensive technical knowledge to use the tool effectively, conducting simulations and understanding forecasts clearly. Another key benefit is the system's customization and flexibility, which can be tailored with company-specific data and continuously improved using new data sources. Combining data from various internal and external sources through an ETL (Extract, Transform, Load) process offers a comprehensive and precise view for decision-making. 4.5 Complexity Creating a project that uses machine learning to predict product demand involves a high level of complexity, requiring close collaboration among different participants and a wide range of specialized knowledge. This type of project demands teamwork from multidisciplinary groups, including data scientists, software engineers, supply chain experts, and project managers. Engaging with external stakeholders, such as suppliers and customers, is crucial to gather relevant data and validate predictive models, offering valuable insights that are not available internally (Zhang & Thomson, 2019 ). The project requires advanced technical knowledge in machine learning, including supervised and unsupervised learning algorithms, data preprocessing techniques, feature engineering, and model validation (Ko et al., 2019 ). Furthermore, using collaboration and project management tools can help improve communication and cooperation among the different participants, aiding in the identification of uncertainties, risks, and critical factors for the project's success (Book et al., 2022 ). When implementing the project in other companies, some risks should be considered such as poor planning, a lack of technical skills, difficulties with system integration, and employee resistance to change. Integrating with existing ones can lead to problems like incompatibility, failures, and errors. Additionally, employee resistance to adopting the new ML application can lead to mistakes, misuse, and low adoption. Specific technological risks include data quality, as incomplete, inaccurate, or biased data can lead to inefficient or incorrect predictions. Information overload can make data management and interpretation difficult. Data security is essential to protect against breaches. The AI model performance may not reach the desired accuracy due to technical limitations or insufficient data. Rapid technological advances can quickly make the system obsolete, requiring continuous updates and ongoing investments. Additionally, reliance on appropriate technological infrastructure is necessary to support data processing and storage. A dependency on IT vendors, such as AWS, that can migrate to another platform or even render their own platform obsolete, could increase costs and pose a risk to the project. Regarding regulatory risks, compliance with data protection laws such as the Brazilian LGPD (General Data Protection Law) is critical to avoid fines and reputational damage. There are also specific regulations on AI use and industry standards that must be followed to ensure security and compliance. To address these risks, strategies such as proper planning, training and capacity building, planned integration, stakeholder engagement, data governance, technological updates, and cybersecurity measures should be adopted. 5. DISCUSSION OF RESULTS 5.1 Theoretical implications The results confirmed the application's goal that using machine learning techniques, enabled by cloud platforms, could greatly enhance demand forecasting and inventory management at a building materials distributor. The achieved accuracy (about 99.31% via WAPE) marked a significant improvement over the simpler methods previously used by the company and over naive model benchmarks (as evidenced by the better performance, also measured by MASE, if applicable). This finding aligns with the literature that highlights the benefits of ML for capturing complex patterns in demand time series (Makridakis et al., 2018 ; Tirkolaee et al., 2021 ). The ability of the algorithms used (such as DeepAR+) to automatically incorporate features like seasonality, trends, and related metadata (if provided) was crucial to the performance achieved. Using platforms like Amazon Forecast and SageMaker Canvas proved an effective way to overcome barriers to ML adoption, especially for SMEs such as Casa Cardão. Automating tasks such as algorithm selection, hyperparameter tuning, and model evaluation (AutoML) sped up development and enabled strong results without requiring a large team of data scientists (Amazon Web Services, 2025a ; Amazon Web Services, 2025b ). The tangible benefits observed—reduction in stockouts, inventory optimization, and better planning—corroborate the positive impact of technology on operational efficiency and potential cost savings (Carbonneau et al., 2009 ; Fildes et al., 2009 ). By minimizing waste (associated with excess inventory) and lost sales (due to disruptions), the solution directly contributed to the company's competitiveness. 5.2. Managerial recommendations From a decision-making perspective, companies must balance the trade-offs between processing costs, predictive performance, and the level of explainability when selecting cloud platforms. Services like Amazon SageMaker Canvas offer competitive pay-as-you-go pricing and require minimal knowledge for setting up data pipelines. In contrast, open-source solutions (e.g., Prophet) offer greater algorithmic transparency but depend on an internal data science team. For SMEs with limited IT resources, Casa Cardão's experience indicates that outsourcing infrastructure is an efficient and replicable way to adopt AI in operations gradually. However, some limitations and challenges arose. The quality and availability of historical data were essential; gaps or inconsistencies required significant effort during pre-processing. Additionally, the interpretability of some more complex ML models, known as "black box" models, can be difficult, though platforms like SageMaker Canvas offer tools to help explain their predictions. Successful implementation also depended on a cultural shift and adapting internal company processes to trust and effectively use the predictions generated by the system. Despite these challenges, the experience demonstrates that integrating artificial AI into SCM processes is practical and beneficial. 6. CONCLUSION This work summarized the development and deployment of machine-learning-based predictive models for inventory management at Casa Cardão, a medium-sized Brazilian distributor of building materials. Using Amazon Forecast and SageMaker Canvas, the project implemented a demand-forecasting system that achieved an overall accuracy of approximately 99.31% (WAPE), substantially outperforming the simpler forecasting procedures previously used by the company. The solution helped reduce stockouts, optimize inventory levels, and improve purchasing planning, thereby generating direct operational and economic benefits. At the same time, the study illustrated how cloud-based AutoML platforms can lower the entry barriers for Small and Medium-sized Enterprises (SMEs) that lack specialized data science teams, providing a concrete example of how AI-driven demand forecasting can support digital transformation in supply chain management. From a theoretical and practical standpoint, the study contributes by compiling, in detail, an end-to-end implementation of ML in an SME supply-chain context, including the technology architecture, data pipeline, and human–machine interaction protocols. It reinforces the evidence that advanced ML models, when combined with managerial judgment and appropriate performance monitoring, can capture nonlinear and seasonal patterns that traditional statistical methods may miss, especially in volatile demand environments such as construction supplies distribution. The project also highlights the importance of governance mechanisms—such as continuous accuracy monitoring, retraining triggers, and feedback loops between forecasts and operational decisions—to sustain predictive performance over time. However, as a practice and policy paper based on a single case, this study has several limitations that should be acknowledged. First, the proof of concept and subsequent deployment were conducted in one company, operating in a specific regional and industry context (construction supplies distribution in Minas Gerais, Brazil). This context dependence constrains the generalizability of the results to other industries, countries, or supply-chain configurations. Second, access to data was limited in both breadth and depth: some SKUs had relatively short or irregular historical time series; several potentially relevant external variables (e.g., detailed competitor actions, granular macroeconomic indicators, and micro-regional construction activity) were either unavailable or only partially integrated; and data quality issues required extensive pre-processing. These constraints may have influenced model performance and reduced the ability to test more sophisticated feature-engineering strategies. Further limitations arise from the technological choices. The solution relied primarily on AWS managed services (Amazon Forecast and SageMaker Canvas) and benefited from temporary cloud credits, which may not reflect the long-term cost structure faced by all SMEs. There is also a degree of vendor lock-in and opacity associated with proprietary algorithms, which can limit transparency, reproducibility, and portability to other environments. Additionally, the evaluation period for business outcomes was relatively short and focused mainly on forecasting metrics and qualitative managerial perceptions, rather than on a full cost–benefit analysis including financial indicators such as return on investment, working-capital reduction, or service-level improvements over multiple years. Finally, the study did not systematically assess organizational change dimensions, such as user adoption, learning curves, and the impact of the system on roles, routines, and decision-making culture. These limitations open several promising avenues for future work. One important direction is the design, implementation, and evaluation of an open-source reference model that can replicate the core functionalities of the solution with even lower marginal costs and greater transparency. Such a model could be built using open-source libraries and frameworks and released under a permissive license, enabling replication, adaptation, and peer review by researchers, practitioners, and policy-makers. Comparative studies could then benchmark proprietary cloud services against fully open-source stacks—considering not only predictive accuracy, but also total cost of ownership, explainability, maintainability, and ease of integration with existing ERP and CRM systems. Future research should also extend the empirical scope beyond a single company and sector. Multi-case or cross-industry studies—including food retail, pharmaceuticals, automotive parts, and other sectors with complex, intermittent demand—could test the robustness of the proposed architecture and identify contingencies that moderate its effectiveness (e.g., product variety, demand volatility, lead times, or data availability). Longitudinal designs could assess the long-term financial impact of ML-based demand forecasting on inventory turnover, stockout rates, service levels, and profitability, ideally using quasi-experimental designs or A/B tests comparing business units with and without the predictive system. Another promising stream of work involves integrating demand forecasting with optimization models and prescriptive analytics. Rather than stopping at forecasting, future applications could combine predictions with optimization routines for order quantities, safety stocks, and replenishment policies, thus forming closed-loop decision-support systems. Research could also explore sociotechnical aspects in greater depth, including user trust in algorithmic recommendations, the design of explainable interfaces, training programs for managers, and governance frameworks for AI in SMEs. Finally, policy-oriented studies could examine how public agencies, development banks, and innovation programs might support the diffusion of low-cost, open, and responsible AI solutions for inventory management among small and medium-sized firms. This study concludes that advanced cloud-based machine-learning techniques can substantially improve demand forecasting and inventory management in a medium-sized construction supplies distributor, but the study also warns that such solutions are not off-the-shelf panaceas. Their success depends on data availability and quality, careful technological choices, sustained organizational engagement, and appropriate governance. Recognizing these limitations and exploring the proposed possibilities for future work may help researchers, practitioners, and policy-makers design more robust, transparent, and accessible predictive systems—contributing to the broader goal of democratizing AI-enabled supply-chain optimization for SMEs. Declarations Ethical Approval and Consent to Participate This study did not involve human participants, personal data, or interventions requiring ethical approval. The research was conducted using organizational and operational data provided by a private company under a confidentiality agreement. Therefore, ethical approval and consent to participate are not applicable . Consent for Publication All authors have reviewed and approved the final version of the manuscript and consent to its publication. Funding This research was supported by the MAI/DAI Call No. 68/2022 , funded by the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) , under the Ministry of Science, Technology and Innovation of Brazil . Author Contribution The research group's coordination and the primary manuscript's composition were undertaken by D.D.G. R.B.C. served as the doctoral supervisor and performed a comprehensive review of the document. M.A.H., in conjunction with B.K.G.N.C., bore responsibility for the application's development and ongoing support during the Proof of Concept phase, in addition to authoring the front-end interface. Acknowledgement We want to thank the Brazilian Ministry of Science and Technology, primarily through CNPq and the MAI/DAI 68 call for proposals in 2022, for providing funding for the development of this research. Data Availability Data Availability StatementThe data that support the findings of this study are not publicly available due to confidentiality agreements with the participating organization and the presence of commercially sensitive information. The datasets were used under license for the current study and are therefore not available for public disclosure. Aggregated and anonymized data, as well as methodological details necessary to replicate the analytical procedures, are included within the article. Additional information may be made available from the corresponding author upon reasonable request, subject to approval by the participating organization. References Amazon Web Services. (2025a). Amazon Forecast: Developer guide . https://docs.aws.amazon.com/forecast/ Amazon Web Services. (2025b). Amazon SageMaker Canvas: User guide . https://docs.aws.amazon.com/sagemaker/latest/dg/canvas.html Ampazis, N. (2015). Forecasting demand in supply chain using machine learning algorithms. International Journal of Artificial Life Research , 5 (1), 56–73. Book, M., Riedel, M., Neukirchen, H., & Erlingsson, E. (2022). Facilitating collaboration in machine learning and high-performance computing projects with an interaction room. In 2022 IEEE 18th International Conference on e-Science (e-Science) (pp. 529–538). IEEE. https://doi.org/10.1109/eScience55777.2022.00093 ResearchGate + 1 Bowerman, B. L., O’Connell, R. T., & Koehler, A. B. (2005). Forecasting, time series, and regression: An applied approach . Cengage Learning. Carbonneau, R., Laframboise, K., & Vahidov, R. (2009). Application of machine learning techniques for supply chain demand forecasting. European Journal of Operational Research , 184 (3), 1140–1154. https://doi.org/10.1016/j.ejor.2006.12.004 Carvalho, F. P., dos Santos, R. C., Nascimento, S. M., da Silva Coutinho, J. C., & de Sousa, R. R. (2023). Investigating the relationship between academia and the information technology industry: A systematic literature review. Concilium , 23 (21), 11–35. Fildes, R., Goodwin, P., Lawrence, M., & Nikolopoulos, K. (2009). Effective forecasting and judgmental adjustments: An empirical evaluation and strategies for improvement in supply-chain planning. International Journal of Forecasting , 25 (1), 3–23. https://doi.org/10.1016/j.ijforecast.2008.11.010 Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and practice (3rd ed.). OTexts. https://otexts.com/fpp3/ Ko, H., Witherell, P., Ndiaye, N. Y., & Lu, Y. (2019). Machine learning based continuous knowledge engineering for additive manufacturing. In 2019 IEEE 15th International Conference on Automation Science and Engineering (CASE) (pp. 648–654). IEEE. https://doi.org/10.1109/COASE.2019.8843316 ACM Digital Library + 1 Likhar, P. K., Jha, A., Tiwari, S., Sunar, A., Shahu, S., & Thate, S. (2023). Machine learning-based sales prediction and inventory management for grocery stores. International Journal of Advanced Research in Science, Communication and Technology. https://doi.org/10.48175/ijarsct-13605 Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2018). Statistical and machine learning forecasting methods: Concerns and ways forward. PLOS ONE , 13 (3), e0194889. https://doi.org/10.1371/journal.pone.0194889 Naik, H., Yashwanth, K., Suraj, P., & Jayapandian, N. (2022). Machine learning based food sales prediction using random forest regression. In 2022 6th International Conference on Electronics, Communication and Aerospace Technology (pp. 998–1004). Punia, S., Nikolopoulos, K., Singh, S. P., Madaan, J. K., & Litsiou, K. (2020). Deep learning with long short-term memory networks and random forests for demand forecasting in multi-channel retail. International Journal of Production Research , 58 (16), 4964–4979. Sinduscon-MG. (2024). Sector report on the construction industry in Minas Gerais 2023–2024: Market for cement, mortar, and ceramic tiles . Minas Gerais Construction Industry Union. Sokol, K., & Flach, P. (2020). One explanation does not fit all: The promise of interactive explanations for machine learning transparency. KI – Künstliche Intelligenz , 34 (2), 235–250. https://doi.org/10.1007/s13218-020-00637-y Syntetos, A. A., Babai, Z., Boylan, J. E., Kolassa, S., & Nikolopoulos, K. (2016). Supply chain forecasting: Theory, practice, their gap and the future. European Journal of Operational Research , 252 (1), 1–26. https://doi.org/10.1016/j.ejor.2015.11.010 Tian, X., Wang, H., & Erjiang, E. (2021). Forecasting intermittent demand for inventory management by retailers: A new approach. Journal of Retailing and Consumer Services , 62 , 102662. https://doi.org/10.1016/j.jretconser.2021.102662 Tirkolaee, E. B., Sadeghi, S., Mooseloo, F. M., Vandchali, H. R., & Aeini, S. (2021). Application of machine learning in supply chain management: A comprehensive overview of the main areas. Mathematical Problems in Engineering, 2021 , 5566842. https://doi.org/10.1155/2021/5566842 Van Steenbergen, R. M., & Mes, M. R. K. (2020). Forecasting demand profiles of new products. Decision Support Systems , 139 , 113401. https://doi.org/10.1016/j.dss.2020.113401 Vairagade, N., Logofatu, D., Leon, F., & Muharemi, F. (2019). Demand forecasting using random forest and artificial neural network for supply chain management. In Computational Collective Intelligence: 11th International Conference, ICCCI 2019, Proceedings, Part I (pp. 328–339). Springer. https://doi.org/10.1007/978-3-030-28374-0_27 Zhang, X., & Thomson, V. (2019). Modelling the development of complex products using a knowledge perspective. Research in Engineering Design , 30 , 203–226. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 31 Mar, 2026 Read the published version in Future Business Journal → Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8502995","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":574798087,"identity":"8f11295f-3e61-47dd-876e-60581ed4d749","order_by":0,"name":"Diocélio Dornela Goulart","email":"data:image/png;base64,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","orcid":"","institution":"Fundação Dom Cabral","correspondingAuthor":true,"prefix":"","firstName":"Diocélio","middleName":"Dornela","lastName":"Goulart","suffix":""},{"id":574798089,"identity":"6cc22909-bbfa-4ee9-af1c-61a58304289f","order_by":1,"name":"Rodrigo Baroni de Carvalho","email":"","orcid":"","institution":"Pontifícia Universidade Católica de Minas Gerais","correspondingAuthor":false,"prefix":"","firstName":"Rodrigo","middleName":"Baroni","lastName":"de Carvalho","suffix":""},{"id":574798090,"identity":"dc1a7aa8-e6a7-419c-9bce-f28dcab623c8","order_by":2,"name":"Mariana Almeida Henriques","email":"","orcid":"","institution":"Pontifícia Universidade Católica de Minas Gerais","correspondingAuthor":false,"prefix":"","firstName":"Mariana","middleName":"Almeida","lastName":"Henriques","suffix":""},{"id":574798091,"identity":"e2cc6223-d420-490e-943d-f37243923258","order_by":3,"name":"Bárbara Kawany Gonçalves Nunes Carvalho","email":"","orcid":"","institution":"Pontifícia Universidade Católica de Minas Gerais","correspondingAuthor":false,"prefix":"","firstName":"Bárbara","middleName":"Kawany Gonçalves Nunes","lastName":"Carvalho","suffix":""}],"badges":[],"createdAt":"2026-01-02 19:23:40","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8502995/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8502995/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s43093-026-00807-8","type":"published","date":"2026-03-31T15:58:09+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":100414697,"identity":"e3830de0-8f01-40bc-b50e-26825e020dd5","added_by":"auto","created_at":"2026-01-16 13:19:55","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4236949,"visible":true,"origin":"","legend":"","description":"","filename":"FBUJPredictiveModelsConstructionSupplies.docx","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/3b5a0818bcf5243d96b2ded2.docx"},{"id":100414923,"identity":"4b64757e-1f72-4c6d-9ba9-e9e529ba1532","added_by":"auto","created_at":"2026-01-16 13:20:19","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6397,"visible":true,"origin":"","legend":"","description":"","filename":"3a50fe1660564e3d99418c67fa8078c6.json","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/7c9714b2d585bf173d387a90.json"},{"id":100414847,"identity":"935f729d-c325-4d9d-8672-2ed5d1b62035","added_by":"auto","created_at":"2026-01-16 13:20:09","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":91944,"visible":true,"origin":"","legend":"","description":"","filename":"3a50fe1660564e3d99418c67fa8078c61enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/181132cec94fd51eb86c1bf4.xml"},{"id":100414549,"identity":"4eca744d-2da6-4be4-b6b5-60fa39a0de22","added_by":"auto","created_at":"2026-01-16 13:19:33","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":78457,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/b14a4a8b8a2f3761f1a9445f.png"},{"id":100414747,"identity":"0be44d7d-c6d6-4192-9022-98d9a53558bb","added_by":"auto","created_at":"2026-01-16 13:19:58","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":59882,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/5a7d24997bc1d4845aa17527.png"},{"id":100414547,"identity":"1c041e96-497f-4ac7-90a3-afd6ebd29e0c","added_by":"auto","created_at":"2026-01-16 13:19:33","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":38203,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/838bddebea56e53194369533.png"},{"id":100414695,"identity":"ec049046-88ae-43e2-9d5b-d6e82070cbec","added_by":"auto","created_at":"2026-01-16 13:19:54","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13550,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/19408e37e98f9eb78438f05f.png"},{"id":100414760,"identity":"051de34b-1f90-4675-8f28-fc5d3fea8d17","added_by":"auto","created_at":"2026-01-16 13:20:00","extension":"xml","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":89780,"visible":true,"origin":"","legend":"","description":"","filename":"3a50fe1660564e3d99418c67fa8078c61structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/64cf97f61f6e20e741b56647.xml"},{"id":100414586,"identity":"7bdd606c-ee0b-4049-a955-3a8ffd12daa9","added_by":"auto","created_at":"2026-01-16 13:19:43","extension":"html","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":98192,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/384127cba7c7691c1aa0f8d1.html"},{"id":100414844,"identity":"5d63e25d-00ba-4a50-bb94-11145b11cb0d","added_by":"auto","created_at":"2026-01-16 13:20:08","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1172524,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eInventory forecast graphs\u003c/em\u003e\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/6f26c1f286c826b2dd6e2ee0.png"},{"id":100414838,"identity":"c7130ad8-63a4-4b16-83e6-e82ed2dcae85","added_by":"auto","created_at":"2026-01-16 13:20:07","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1750208,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eTechnological Architecture for Development\u003c/em\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/5ccc91a9dc68fb73f3007f8d.png"},{"id":100414696,"identity":"c7d99c5c-4169-4224-97a1-0ed36accdaf7","added_by":"auto","created_at":"2026-01-16 13:19:55","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1212802,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003ePhases for development\u003c/em\u003e\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/32ada8e585b44eefaa7c94ad.png"},{"id":100414661,"identity":"645853ad-6760-41d6-95c7-db143502e438","added_by":"auto","created_at":"2026-01-16 13:19:50","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":39963,"visible":true,"origin":"","legend":"\u003cp\u003eDemand Predictor Model Flowchart\u003c/p\u003e","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/fb24a6148354b853d624e852.jpeg"},{"id":106344738,"identity":"c2b34134-20bc-4ac3-8018-c8faf325f316","added_by":"auto","created_at":"2026-04-07 16:16:28","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4274778,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8502995/v1/06738c6c-6379-484f-ab7c-2cebb0997081.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003ePredictive Models for Inventory Optimization: a machine learning application for demand forecasting at a construction supplies distributor\u003c/p\u003e","fulltext":[{"header":"1. CONTEXT","content":"\u003cp\u003eEfficient inventory management is a vital challenge for distribution companies, as it directly affects customer satisfaction, operating costs, and overall business profitability. In the construction supplies industry, which features a diverse range of products and seasonal demand fluctuations, forecasting becomes even more difficult (Syntetos et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Forecasting mistakes can result in excess inventory, raising storage costs and the risk of obsolescence, or lead to stockouts (product shortages), causing lost sales and customer dissatisfaction (Fildes et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2009\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn this context, the objective of this practice paper is to develop and apply machine learning (ML) techniques to improve inventory management. The focus was on predicting product demand more accurately, lowering costs, and maximizing resource usage. The practical application was centered on Casa Card\u0026atilde;o, a medium-sized Brazilian distributor operating in both wholesale and retail channels for construction supplies. This business context highlights the significant logistical and operational challenges, justifying the implementation of advanced predictive techniques to reduce disruptions and lower inventory holding costs.\u003c/p\u003e \u003cp\u003eAs is typical in medium-sized organizations, the company has a lean IT (Information Technology) structure and lacks analysts with advanced knowledge of Data Science and AI (Artificial Intelligence). Many available software solutions require significant investment and specialized staff. Currently, there is a shortage of these professionals in Brazil, so advanced applications of technologies such as AI and data analytics remain out of reach for many medium-sized Brazilian companies.\u003c/p\u003e \u003cp\u003eThe complexity inherent in Supply Chain Management (SCM) and demand forecasting, especially in scenarios with intermittent or irregular demand (Tian et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), justifies exploring more advanced methods than traditional ones, which are often limited in their ability to capture nonlinear patterns and multiple influencing variables (Makridakis et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). The use of ML in SCM has emerged as a promising approach, providing powerful tools for analyzing large datasets and identifying complex patterns (Tirkolaee et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis practice paper outlines the process of developing, implementing, and evaluating ML-based predictive models using Amazon Web Services (AWS), specifically Amazon Forecast and Amazon SageMaker Canvas. This practical study demonstrates the feasibility and benefits of using these cloud-based technologies to improve decision-making in inventory management, supporting the digital transformation and operational efficiency of an organization with a less structured IT department. The study also offers a reproducible model for Small and Medium-sized Enterprises (SMEs), enabling the use of machine learning for inventory management at lower costs.\u003c/p\u003e \u003cp\u003eThe study was developed to create an ML-based solution that integrates internal data from ERP (Enterprise Resource Planning) and CRM (Customer Relationship Management) business systems, along with external data (public data and partner databases), in a Data Lake environment. This practical application aims to provide more accurate forecasts, supporting informed decision-making and reducing waste (Van Steenbergen \u0026amp; Mes, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Naik et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe project was selected and received support from a specific innovation call for proposals of the National Council for Scientific and Technological Development (CNPq), an entity associated with the Brazilian Ministry of Science, Technology and Innovation (MCTI). The purpose of the call for proposals was to promote postgraduate research collaborations between industry and universities. The research project team comprised the project coordinator (a professor in Management), a PhD researcher in Management, and two undergraduate research scholarship students in Software Engineering.\u003c/p\u003e \u003cp\u003eOther partners are also involved in the project: i) a Brazilian business school, which provided access to its customer database to identify potential opportunities for tests and proof of concept; ii) a Paraguayan-Brazilian technological park associated with one of the largest hydroelectric power plants in the world, which supported submitting the research project to the FINEP (Brazilian Funding Agency for Studies and Projects, also associated with MCTI) call for innovation proposals, potentially enabling additional resources; iii) a cloud computing company based in S\u0026atilde;o Paulo, which hosted the necessary resources for the project's development and infrastructure; iv) AWS - Amazon Web Services, which supplied an environment for testing over 12 months, enabling the execution of the POC - Proof of Concept; and v) a consulting and development firm in Minas Gerais, which provided access and resources for AI development tools such as ChatGPT Pro, Manus AI, and Base44.\u003c/p\u003e \u003cp\u003eThe overall goal of the project was to research, identify, and develop ML-based models and technological solutions that can serve as predictors and automate data collection and analysis using Artificial Intelligence. The project aimed to create an intuitive interface for managers that enables accurate simulations and forecasts for inventory management. A specific goal is to make the application development easy to duplicate at a low cost, using open-source software or no-code or low-code platforms (with little or no coding or development), enabling its adoption by companies with reduced IT budget and personnel.\u003c/p\u003e \u003cp\u003eBeyond this introduction, the next section of this report comprises the scientific framework that guides the study's development. Section \u003cspan refid=\"Sec3\" class=\"InternalRef\"\u003e3\u003c/span\u003e details the application development and reports the experiment conducted, presenting the achieved results, and discussing the application of the tools. Section \u003cspan refid=\"Sec7\" class=\"InternalRef\"\u003e4\u003c/span\u003e evaluates the study across the following dimensions: a) how well it aligns with related fields and areas of Management; b) the potential environmental and stakeholder impacts of the project; c) the feasibility of applying the proposed technology and processes; d) the innovation and level of knowledge involved in product development; and e) the complexity of the network and relationships necessary for its design and execution. Section \u003cspan refid=\"Sec13\" class=\"InternalRef\"\u003e5\u003c/span\u003e discusses the results, while Section \u003cspan refid=\"Sec16\" class=\"InternalRef\"\u003e6\u003c/span\u003e concludes the practical with recommendations for future work and addresses its limitations.\u003c/p\u003e"},{"header":"2 THEORETICAL BACKGROUND","content":"\u003cp\u003eThis section offers an overview of the theoretical framework that guided the application development. Inventory management plays a key role in logistics and operations management, balancing inventory costs with the risk of product shortages. Accurate demand forecasting is essential for effective inventory control (Hyndman \u0026amp; Athanasopoulos, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Traditional forecasting techniques, such as moving averages, exponential smoothing, and ARIMA (Autoregressive Integrated Moving Average) models, are commonly used but may have limitations when handling complex, seasonal, or multi-variable data (Bowerman et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2005\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDemand in the retail and distribution industry, such as construction supplies, often exhibits characteristics like intermittency (periods without demand) and seasonality. Traditional models may struggle to capture these patterns accurately (Syntetos et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Tian et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Additionally, external factors like promotions, economic conditions, changes in tax laws, political events, competitor activities, and even weather can influence demand, requiring models that can incorporate multiple sources of information.\u003c/p\u003e \u003cp\u003ePredictive models using Artificial Intelligence (AI) for inventory management offer significant advantages over traditional methods. ML algorithms process many factors and identify complex patterns that traditional techniques overlook (Ampazis, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2015\u003c/span\u003e), resulting in more accurate demand forecasts and enabling quick adaptation to changing market conditions (Likhar et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). As a subfield of AI, ML provides alternative and complementary methods to traditional statistical forecasting techniques. ML algorithms can identify complex patterns from historical data, including nonlinear relationships and interactions among multiple variables, without relying on strict statistical assumptions (Makridakis et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eProduct demand forecasting is a vital technique for inventory management, allowing companies to optimize their stock levels, lower costs, and enhance customer service. One of the machine learning methods that has gained prominence in this field is the use of Random Forest for forecasting (Vairagade et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Other machine learning algorithms have been effectively used for demand forecasting, including Artificial Neural Networks (ANN), Support Vector Machines (SVM), and decision tree-based methods like Random Forest and Gradient Boosting (Tirkolaee et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Vairagade et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). These approaches are especially valuable when large amounts of data (Big Data) are available and demand is driven by a complex mix of factors (Carbonneau et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2009\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eMultiple studies show that this technique is highly effective for inventory management. Van Steenbergen and Mes (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) introduced DemandForest, an innovative approach that combines K-means clustering, Random Forests, and Quantile Regression Forests to forecast demand patterns for new products. This method was tested with real data sets and outperformed several reference methods, leading to around 15% savings in inventory (Van Steenbergen \u0026amp; Mes, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Vairagade et al. (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) compared various demand forecasting models, including Random Forest and Artificial Neural Networks. They found that Random Forest was more accurate, making it better suited for predicting product demand in complex supply chains.\u003c/p\u003e \u003cp\u003eOn the other hand, Tian et al. (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) tackled intermittent demand forecasting, a common challenge in retail, and introduced a combined approach that accounts for current inventory and past sales. This method, known as the Markov-combined method (MCM), proved to be more accurate than traditional techniques, leading to significant improvements in inventory management.\u003c/p\u003e \u003cp\u003ePunia et al. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) introduced another combined approach that utilizes Long Short-Term Memory (LSTM) and Random Forest models to capture complex temporal and regression relationships. This method greatly outperformed other forecasting techniques in accuracy and robustness within multichannel retail settings.\u003c/p\u003e \u003cp\u003eNaik et al. (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) examined sales forecasting in the food industry using Random Forest Regression. This model was notable for its high accuracy compared to other machine learning methods, making it ideal for predicting sales and optimizing inventory management.\u003c/p\u003e \u003cp\u003eIn summary, the reviewed literature highlights three key pillars that guided the application design: (i) the superiority of ML approaches over traditional statistical methods in volatile demand scenarios, due to their ability to capture nonlinear relationships and multiple variables (Makridakis et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Hyndman \u0026amp; Athanasopoulos, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2021\u003c/span\u003e); (ii) the need to combine algorithmic forecasting with human judgment to enhance operational decision-making, as supported by studies on managerial adjustments in forecasting (Goodwin \u0026amp; Fildes, 2019); and (iii) the importance of explainability mechanisms that clarify the most influential factors behind the forecast, thereby increasing managers' confidence in the system's recommendations (Sokol \u0026amp; Flach, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). These findings informed the selection of AutoML platforms capable of integrating various algorithms (e.g., DeepAR+, Random Forest) and robust evaluation metrics. As detailed in the next section, these theoretical guidelines were put into practice in the application development, defining the data pipeline, model selection criteria, and human-machine interaction protocol that form the foundation of the implemented solution.\u003c/p\u003e"},{"header":"3. APPLICATION DEVELOPMENT","content":"\u003cp\u003eThis section describes the process of application development, detailing the IT tools, the underlying architecture, and the performance assessment. The context of the application was Casa Card\u0026atilde;o, a hundred-year-old Brazilian distributor operating in both wholesale and retail channels for construction supplies. It employs around 220 staff members and 200 sales representatives. It runs two regional distribution centers covering 12,000 m\u0026sup2; of storage and markets a portfolio of over 10,500 SKUs (Stock Keeping Units), with strong demand for 150 items from the ABC curve. In 2023, the company generated R\u003cspan\u003e$\u003c/span\u003e310\u0026nbsp;million (about US\u003cspan\u003e$\u003c/span\u003e60\u0026nbsp;million) in revenue, holding a 4% market share in Minas Gerais state within the cement, mortar, and ceramic tile segments (Sinduscon-MG, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Development tools\u003c/h2\u003e \u003cp\u003eDeveloping and deploying ML models can be complex and demand substantial computational resources. Cloud platforms like Amazon Web Services (AWS), Google Cloud Platform (GCP), and Microsoft Azure provide managed services that potentially make this process easier. In this study, two AWS tools were utilized. Amazon Forecast is a fully managed service that uses ML to deliver highly accurate time-series forecasts. It automates much of the ML process, from data loading to model training and deployment, automatically testing various algorithms (including ARIMA, ETS, Prophet, DeepAR+, and others) and choosing the best one for the data provided (Amazon Web Services, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e). The second tool, Amazon SageMaker Canvas, is a no-code visual interface that enables business analysts to build ML models without writing code. It streamlines data preparation, model training, and forecast generation, making ML more accessible to non-expert users (Amazon Web Services, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2025b\u003c/span\u003e). These platforms facilitated the efficient experimentation and implementation of predictive models in the Casa Card\u0026atilde;o environment.\u003c/p\u003e \u003cp\u003eThe project resulted in a dashboard or visual layer that displays structured data collected through an ETL (Extract, Transform, and Load) process using AWS SageMaker Data Wrangler, covering both external and internal data. This dashboard includes a series of graphs showing inventory forecasts (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). With a prompt feature available on the generative AI-powered dashboard interface, the application provides detailed explanations of the methodology and graphs used, making the tool more accessible and understandable for managers.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eNote\u003c/strong\u003e \u003cp\u003eScreen of the application developed by the research group.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eAdditionally, the interface enables managers to run simulations as needed, combine different data sources, and leverage this information to enhance inventory management. The project is delivered in a tailored manner, using company-specific data along with information from public sources. The conversational management dashboard can be continuously improved by adding new data sources to develop a more robust and effective predictive model for inventory management.\u003c/p\u003e \u003cp\u003eThe developed product is a demand prediction system that utilizes predictor results generated through AWS SageMaker Canvas, featuring an interactive dashboard that provides product demand forecasts across different scenarios. These scenarios include optimistic, neutral, and pessimistic. In the optimistic scenario, inventory levels are low, and the forecast accounts for seasonality and promotions. In the pessimistic scenario, inventories are high, and there is no forecasted demand for the products. The system's features include time series analysis, integration of internal and external data, customized simulations, and a console explaining the methodology and graphs used.\u003c/p\u003e \u003cp\u003eThe project can help the company by lowering costs related to excess inventory or product shortages and by making inventory management more efficient, which allows for more confident decision-making. Specifically, the goal was to create AI-based inventory management software that could predict the necessary inventory levels with 90% accuracy, and this target was successfully achieved with a 99.31% accuracy rate. Measurable objectives include reducing excess inventory by 20% and shortages by 15% within one year of implementation.\u003c/p\u003e \u003cp\u003eSeveral machine learning techniques were tested and used to develop the project, including AWS's proprietary algorithm, DeepAR. It is based on Recurrent Neural Networks (RNN) with an LSTM (Long Short-Term Memory) model and is often used to generate probabilistic forecasts for multiple series (various products). DeepAR demonstrated the best accuracy among the models available on the AWS platform, such as DeepAR+, CNN-QR, Prophet, ARIMA, and ETS.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Technology architecture and phases\u003c/h2\u003e \u003cp\u003eApplication development began in August 2024, and by April 2025, the proof-of-concept and validation of the results were completed. The system was fully operational within the organization by July 2025. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e illustrates the technological architecture used to develop the application, outlining its technology pipeline organized into four functional layers in the AWS environment. In the ingestion layer, transactional data extracted from Casa Card\u0026atilde;o's ERP and CRM systems, along with external data such as weather, holidays, and macroeconomic indices, are stored in a data lake on Amazon S3. The preparation layer uses SageMaker Data Wrangler for ETL processes, ensuring the cleaning, aggregation, and enrichment of time-series data. Next, the modeling layer employs Amazon SageMaker Canvas for exploratory AutoML and the Forecast service (DeepAR+, CNN-QR, Prophet, among others) for final training, automatically selecting the algorithm with the lowest back-test error (Amazon Web Services, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e; \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2025b\u003c/span\u003e). The generated forecasts are stored in a versioned S3 bucket, while AWS Lambda functions manage the daily model updates, providing serverless scalability.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFinally, at the application layer, the results are fed into an interactive dashboard (developed in React using the base44 platform) that displays P35 (pessimistic), P50 (neutral), P60, and P80 (optimistic) scenarios, along with a conversational interface powered by GPT O3-mini. In the pessimistic scenario, there is little demand and a lot of stock; in the optimistic scenario, there is a lot of demand and little stock. In the neutral scenario, there is no apparent trend in demand, and the stock is balanced. This modular and fully managed workflow enables rapid iteration, cost efficiency, and seamless integration between algorithmic output and human decision-making.\u003c/p\u003e \u003cp\u003eA set of requirements, categorized as technological, functional, and legal, was considered for the project's development. Technologically, an adequate IT infrastructure was necessary to support the collection, storage, and processing of large data volumes. AWS was selected for its resources, widespread availability, and ease of implementation. Functionally, the developed interface\u0026mdash;dashboard\u0026mdash;is intuitive and user-friendly, capable of integrating with results generated by AWS and providing scenario analysis features for users. Legally, the project complies with data protection and privacy regulations and does not handle any sensitive data.\u003c/p\u003e \u003cp\u003eThe stakeholders involved included both internal and external parties. Internally, the research team and IT team at Casa Card\u0026atilde;o, along with the project manager at the cloud computing company, were essential. Externally, AWS served as a key technology provider. AWS provided the credits and resources needed for testing through its cloud computing and AI infrastructure. The Proof Of Concept (POC) was conducted at Casa Card\u0026atilde;o, providing a real environment to test and assess the technology's effectiveness. The application development was structured into five distinct phases: Planning, Development, Testing and Validation, Implementation, and Evaluation and Improvement. The Planning phase included defining requirements and scope, as well as a detailed project schedule to ensure clarity regarding the overall objectives.\u003c/p\u003e \u003cp\u003eIn the Development phase, the predictor and an interactive dashboard were built. Next, during the Testing and Validation phase, software testing, predictive model validation, and POCs were conducted. These first three phases occurred between August 2024 and April 2025.\u003c/p\u003e \u003cp\u003eSubsequent phases include Implementation, with implementation in the company, user training, and initial support; and finally, Evaluation and Improvement, which involves performance monitoring, user feedback collection, and continuous improvement. These last two phases were completed by July 2025. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e details each of these phases.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe main challenge was to improve the accuracy of demand forecasting at Casa Card\u0026atilde;o to enhance inventory management. The initial phase involved gathering historical sales data from the company, covering a relevant period (usually 3 to 5 years, depending on availability and detail) to identify seasonal patterns and trends. The data included, at a minimum, product identifiers (SKUs), transaction dates, and quantities sold. Additional information, such as details on promotions, holidays, or product characteristics (metadata), was considered to improve the models.\u003c/p\u003e \u003cp\u003eThe collected data was cleaned and preprocessed. This procedure involved handling missing values, aggregating the data to the required forecast level (e.g., daily or weekly sales per SKU), and identifying and addressing outliers as needed. The target time series (sales per SKU per period) and related datasets (item metadata, related time series like promotion data) were created.\u003c/p\u003e \u003cp\u003eThe Amazon Forecast and Amazon SageMaker Canvas platforms were used to develop and train the predictive models. The prepared data was uploaded to the service. Amazon Forecast automatically trained and evaluated multiple time-series forecasting algorithms, including DeepAR+, CNN-QR, Prophet, ARIMA, and ETS. The service selected the best-performing algorithm based on predefined evaluation metrics (back testing). Predictors (trained models) were generated for the SKU set. However, unfortunately, in the second half of 2024, AWS discontinued the platform. A complementary or alternative approach was explored using SageMaker Canvas as a natural replacement for Amazon Forecast. The data was imported into the visual interface, and predictive models focused on time-series forecasting were built using the platform's automated features, enabling users with less coding expertise to conduct feasibility assessments quickly.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Performance Evaluation\u003c/h2\u003e \u003cp\u003eThe developed model's performance was assessed using standard time-series forecasting metrics, with back testing conducted on AWS. The following metrics were considered: RMSE (Root Mean Squared Error), which gauges the average size of errors and penalizes larger deviations; MAPE (Mean Absolute Percentage Error), which expresses errors in percentage terms relative to actual demand and is sensitive to low values; MASE (Mean Absolute Scaled Error), which compares the model's performance against a naive model and is especially useful for intermittent data; and WAPE (Weighted Absolute Percentage Error), which adjusts errors by total demand, providing greater robustness in scenarios with items of varying sales volumes. The WAPE metric was chosen as the primary indicator for overall accuracy because it best reflects the impact on total sales volume, making it particularly relevant in the retail context.\u003c/p\u003e \u003cp\u003eAfter selecting the best-performing model(s), demand forecasts were created for a relevant future period (e.g., the next 4\u0026ndash;12 weeks). These forecasts were then provided to Casa Card\u0026atilde;o's Commercial Department, through reports and the dashboard, to aid inventory replenishment and purchasing planning decisions, as well as to validate the results obtained.\u003c/p\u003e \u003cp\u003eThe implementation of predictive models using Amazon Forecast and SageMaker Canvas at Casa Card\u0026atilde;o yielded significant results. A comparative analysis of the different algorithms trained by the AWS platforms showed that ML-based models, such as DeepAR+ (a recurrent neural network algorithm from Amazon), often outperformed traditional statistical models (such as ARIMA and ETS) for most of the SKUs analyzed, especially those with more complex or seasonal demand patterns.\u003c/p\u003e \u003cp\u003eTo demonstrate the model's performance on a high-turnover item, the SKU \"Amanco 6 m \u0026times; 100 mm drain pipe\" was selected, classified as \"A\" for value and \"B\" for volume in the Casa Card\u0026atilde;o portfolio. The historical data comprised 16 months of daily records (July 2023 \u0026ndash; November 2024), totaling 11,074 observations, with 80% used for training and 20% for backtesting. The forecast horizon was set to 60 days. In the optimistic scenario (P80), chosen because the factory delayed passing on the exchange rate increase and kept prices stable despite the US dollar\u0026rsquo;s appreciation, the DeepAR\u0026thinsp;+\u0026thinsp;algorithm estimated 2,340 units for December 2023, while actual sales were 2,255 units, resulting in a WAPE of 0.7%. The deviation of less than 1% reinforces the model\u0026rsquo;s accuracy and highlights the synergy between statistical forecasting and managerial judgment in selecting the appropriate quantile for market conditions.\u003c/p\u003e \u003cp\u003eDeveloped through optimization and selection by Amazon Forecast, the final predictive model achieved an overall forecast accuracy of approximately 99.31%. This metric was determined using WAPE, showing that the weighted average absolute error was about 0.69% of the total demand during the evaluation period.\u003c/p\u003e \u003cp\u003eThe flowchart (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) illustrates the entire cycle of using the predictive demand model, outlining seven macro-stages integrated within a cloud computing environment. The process begins with the automated extraction of transactional data from the company's ERP and CRM systems, along with external variables\u0026mdash;macroeconomic indicators, holiday calendars, and weather data\u0026mdash;which are stored in a data lake on Amazon S3. Next, using SageMaker Data Wrangler, the ETL pipeline runs, responsible for cleaning, aggregating overtime, and enriching historical data to ensure its quality for modeling. In the third step, this structured data feeds into SageMaker Canvas, where multiple time series algorithms are automatically trained and validated through back testing; the model with the lowest weighted absolute error (WAPE) is then versioned as the active predictor.\u003c/p\u003e \u003cp\u003eThe next step involves periodically generating forecasts, which are stored in a separate S3 bucket and simultaneously sent to a relational database via AWS Lambda functions for real-time access. In the fifth step, these forecasts are used by an interactive dashboard built with React (an open-source JavaScript library for building user interfaces), where the manager selects P50, P60, or P80 scenarios and reviews the key factors influencing each forecast horizon. The resulting decisions\u0026mdash;such as adjusting orders, promotions, or inventory policies\u0026mdash;are fed back into the system in the sixth step by recording the actions taken and their outcomes, forming an operational feedback database.\u003c/p\u003e \u003cp\u003eThe seventh step involves continuous accuracy monitoring: a daily verification process compares actual sales with predictions; if the WAPE exceeds the 5% threshold, a new retraining cycle is automatically initiated, ensuring adaptive learning and predictive robustness over time. This modular and coordinated chain offers scalability, data governance, and alignment between algorithmic intelligence and managerial judgment, which are essential for the successful adoption of demand models in small and medium-sized enterprises.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eNote\u003c/strong\u003e \u003cp\u003eFlowchart to be used during the inventory replenishment process.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eDuring the POC, the main results observed with implementing predictive models showed significant potential benefits for the business, especially the reduction of stockouts, enabled by more accurate forecasts that improved replenishment planning. Additionally, inventory levels were optimized by adjusting safety stocks and reorder points, leading to better allocation of working capital through balancing availability and excess inventory. Purchasing planning also improved, with forecasts providing strategic support for supplier negotiations and greater alignment among procurement, inventory, and demand.\u003c/p\u003e \u003cp\u003eThe adoption of AWS platforms helped streamline the development and implementation of models, enabling the team to focus on analyzing results and integrating forecasts into business processes, despite initial limitations in deep machine learning.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. ANALYSIS OF THE RESULTS","content":"\u003cp\u003eThis section provides detailed information on key aspects of the application, including its compliance with management areas, its influence on various organizational and social aspects, and its usefulness in different business environments. Additionally, it examines the advancements brought by integrating predictive models with ML tools, emphasizing how these innovative solutions can change decision-making and improve processes. Lastly, it discusses the complexity and importance of these models for small and medium-sized businesses.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Adherence\u003c/h2\u003e \u003cp\u003eThe application of predictive models and inventory management is closely connected to various areas of Management and related disciplines, with a special focus on Management of Information Systems. Predictive models align with theme 1 on Decision Making (theories, modeling, support systems, and technologies), promoting tools that support and facilitate decision-making in complex environments. Additionally, the application relates to digital transformation and innovation, enabling greater efficiency in inventory and operations management. The application also engages with Big Data, Data Science, and AI for Strategic Intelligence, as ML applications require large volumes of data. Besides technology and information, the application links with other Management fields such as Operations and Logistics, Innovation, Marketing, and Finance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Impact\u003c/h2\u003e \u003cp\u003eDeveloping a machine-learning-based project to predict product demand can significantly improve the way companies manage complex inventories. This influence reaches various parts of the organization, community, and local area, helping to improve operational efficiency, cut costs, and boost customer satisfaction.\u003c/p\u003e \u003cp\u003eThe application aimed to promote AI adoption among small and medium-sized enterprises (SMEs) through a range of strategies. First, the application focused on providing user-friendly, pre-made AI tools tailored to meet the specific needs of SMEs, which can be easily integrated into existing systems and workflows. This feature reduces the need for specialized AI technical expertise. Additionally, the solution automates routine and repetitive tasks, allowing companies to allocate time and resources to activities of greater strategic importance, thereby enhancing operational efficiency and reducing costs.\u003c/p\u003e \u003cp\u003eThe present application can positively impact the environment by preventing the loss of perishable items and reducing unnecessary consumption for companies in the food supply chain, making resource use more efficient and helping to lower carbon emissions. For example, it can help companies better manage their inventories and supply chains, preventing material waste through demand forecasting, delivery route optimization, and real-time inventory monitoring. The application can also have a significant positive social impact by promoting greater efficiency and growth for companies. The automation of routine tasks and the improved decision-making enabled by AI can boost business productivity.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Applicability\u003c/h2\u003e \u003cp\u003eThe replicability of this type of project in other contexts is high due to the adaptable nature of ML algorithms. Small and medium-sized (SMEs) businesses can also use cloud-based ML services such as AWS, Google AI, and Azure, which offer scalable and low-cost solutions. Especially considering SMEs, considerable potential for the application of the proposed model is observed since low-code platforms do not demand highly skilled teams in machine learning.\u003c/p\u003e \u003cp\u003eML algorithms can be integrated with ERP systems and other existing business management platforms, enabling companies to utilize their historical and real-time data to train models and enhance forecast accuracy. The scalability of these models is a key benefit, as they can adapt to the data volume and complexity of inventory operations. ML's capacity to detect intricate patterns and generate precise predictions greatly reduces human error and streamlines inventory and replenishment processes, especially in business environments where demand is volatile and unpredictable.\u003c/p\u003e \u003cp\u003eML algorithms are adaptable enough to function in different settings and can be tailored to manage inventory across various industries such as retail, manufacturing, and distribution. Thorough documentation of models, training procedures, and integrations supports replication in new environments without having to start from scratch.\u003c/p\u003e \u003cp\u003eThe proposed application is a modular solution, enabling specific components to be reused or adapted for various applications, such as modifying a demand forecasting model to predict sales across different markets or products. The potential market for the project is extensive, as any company with critical inventory management needs can benefit from predictive models. In the industrial sector, these models can forecast demand for specific products, optimize production, and reduce excess inventory. Additionally, they can identify products at high risk of becoming obsolete, enabling preventive measures, and predict equipment failures to facilitate maintenance and avoid production downtime. Distributors and wholesalers can utilize these models to optimize inventory levels, prevent shortages, and minimize storage costs. They can also forecast demand for seasonal products to ensure adequate stock and identify slow-moving items to inform discontinuation decisions. In retail, predictive models help forecast demand for particular products in specific stores, preventing shortages and maximizing sales. They also identify high-turnover products to optimize store displays and predict the end of product life cycles, aiding promotions and reducing waste.\u003c/p\u003e \u003cp\u003eFor service companies, predictive models help forecast demand for specific services, optimize resource allocation, and prevent overload. They can identify customers at high risk of default, enabling preventive measures, and predict equipment maintenance needs to avoid service interruptions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Innovation\u003c/h2\u003e \u003cp\u003eThe level of innovation of the application is justified by making it easier for SMEs to use ML to develop predictive models that forecast product demand. In demand forecasting, ML can analyze historical data on sales, prices, promotions, and other factors to identify patterns that help predict future demand.\u003c/p\u003e \u003cp\u003eThe use of ML for demand forecasting offers several advantages over traditional methods. First, it provides greater accuracy by considering a wide range of factors and identifying patterns that are invisible to the human eye. Additionally, machine learning is faster, enabling companies to make decisions quickly and respond promptly to market changes.\u003c/p\u003e \u003cp\u003eAn analysis of predictive model functionality shows that this feature is missing in current enterprise resource planning (ERP) systems, which only generate reports based on past data. Likewise, business intelligence (BI) systems only display historical data, not data for predictive analysis. Although specialized systems that cross-reference data and offer predictive models exist, their high development costs make this technology inaccessible to small and medium-sized enterprises (SMEs). Implementing such technologies requires experts like data scientists, AI specialists, and cloud computing professionals, highly paid professionals that many SMEs do not have.\u003c/p\u003e \u003cp\u003eCarvalho et al. (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) highlight the challenge of finding these professionals in Brazil, along with the shortage of trainees. The application described in this practical paper aims to address these gaps by offering a simple, user-friendly interface that provides advanced computational resources to SMEs. This feature will enable these companies to leverage advanced AI to solve major business problems, democratizing Data Analytics and making it accessible to more companies that might not afford high implementation costs. The solution's intuitive and straightforward interface enables managers without extensive technical knowledge to use the tool effectively, conducting simulations and understanding forecasts clearly.\u003c/p\u003e \u003cp\u003eAnother key benefit is the system's customization and flexibility, which can be tailored with company-specific data and continuously improved using new data sources. Combining data from various internal and external sources through an ETL (Extract, Transform, Load) process offers a comprehensive and precise view for decision-making.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Complexity\u003c/h2\u003e \u003cp\u003eCreating a project that uses machine learning to predict product demand involves a high level of complexity, requiring close collaboration among different participants and a wide range of specialized knowledge. This type of project demands teamwork from multidisciplinary groups, including data scientists, software engineers, supply chain experts, and project managers. Engaging with external stakeholders, such as suppliers and customers, is crucial to gather relevant data and validate predictive models, offering valuable insights that are not available internally (Zhang \u0026amp; Thomson, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe project requires advanced technical knowledge in machine learning, including supervised and unsupervised learning algorithms, data preprocessing techniques, feature engineering, and model validation (Ko et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Furthermore, using collaboration and project management tools can help improve communication and cooperation among the different participants, aiding in the identification of uncertainties, risks, and critical factors for the project's success (Book et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWhen implementing the project in other companies, some risks should be considered such as poor planning, a lack of technical skills, difficulties with system integration, and employee resistance to change. Integrating with existing ones can lead to problems like incompatibility, failures, and errors. Additionally, employee resistance to adopting the new ML application can lead to mistakes, misuse, and low adoption.\u003c/p\u003e \u003cp\u003eSpecific technological risks include data quality, as incomplete, inaccurate, or biased data can lead to inefficient or incorrect predictions. Information overload can make data management and interpretation difficult. Data security is essential to protect against breaches. The AI model performance may not reach the desired accuracy due to technical limitations or insufficient data. Rapid technological advances can quickly make the system obsolete, requiring continuous updates and ongoing investments.\u003c/p\u003e \u003cp\u003eAdditionally, reliance on appropriate technological infrastructure is necessary to support data processing and storage. A dependency on IT vendors, such as AWS, that can migrate to another platform or even render their own platform obsolete, could increase costs and pose a risk to the project. Regarding regulatory risks, compliance with data protection laws such as the Brazilian LGPD (General Data Protection Law) is critical to avoid fines and reputational damage. There are also specific regulations on AI use and industry standards that must be followed to ensure security and compliance. To address these risks, strategies such as proper planning, training and capacity building, planned integration, stakeholder engagement, data governance, technological updates, and cybersecurity measures should be adopted.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. DISCUSSION OF RESULTS","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Theoretical implications\u003c/h2\u003e \u003cp\u003eThe results confirmed the application's goal that using machine learning techniques, enabled by cloud platforms, could greatly enhance demand forecasting and inventory management at a building materials distributor. The achieved accuracy (about 99.31% via WAPE) marked a significant improvement over the simpler methods previously used by the company and over naive model benchmarks (as evidenced by the better performance, also measured by MASE, if applicable).\u003c/p\u003e \u003cp\u003eThis finding aligns with the literature that highlights the benefits of ML for capturing complex patterns in demand time series (Makridakis et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Tirkolaee et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The ability of the algorithms used (such as DeepAR+) to automatically incorporate features like seasonality, trends, and related metadata (if provided) was crucial to the performance achieved.\u003c/p\u003e \u003cp\u003eUsing platforms like Amazon Forecast and SageMaker Canvas proved an effective way to overcome barriers to ML adoption, especially for SMEs such as Casa Card\u0026atilde;o. Automating tasks such as algorithm selection, hyperparameter tuning, and model evaluation (AutoML) sped up development and enabled strong results without requiring a large team of data scientists (Amazon Web Services, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e; Amazon Web Services, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2025b\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe tangible benefits observed\u0026mdash;reduction in stockouts, inventory optimization, and better planning\u0026mdash;corroborate the positive impact of technology on operational efficiency and potential cost savings (Carbonneau et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2009\u003c/span\u003e; Fildes et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). By minimizing waste (associated with excess inventory) and lost sales (due to disruptions), the solution directly contributed to the company's competitiveness.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e5.2. Managerial recommendations\u003c/h2\u003e \u003cp\u003eFrom a decision-making perspective, companies must balance the trade-offs between processing costs, predictive performance, and the level of explainability when selecting cloud platforms. Services like Amazon SageMaker Canvas offer competitive pay-as-you-go pricing and require minimal knowledge for setting up data pipelines. In contrast, open-source solutions (e.g., Prophet) offer greater algorithmic transparency but depend on an internal data science team. For SMEs with limited IT resources, Casa Card\u0026atilde;o's experience indicates that outsourcing infrastructure is an efficient and replicable way to adopt AI in operations gradually.\u003c/p\u003e \u003cp\u003eHowever, some limitations and challenges arose. The quality and availability of historical data were essential; gaps or inconsistencies required significant effort during pre-processing. Additionally, the interpretability of some more complex ML models, known as \"black box\" models, can be difficult, though platforms like SageMaker Canvas offer tools to help explain their predictions. Successful implementation also depended on a cultural shift and adapting internal company processes to trust and effectively use the predictions generated by the system. Despite these challenges, the experience demonstrates that integrating artificial AI into SCM processes is practical and beneficial.\u003c/p\u003e \u003c/div\u003e"},{"header":"6. CONCLUSION","content":"\u003cp\u003eThis work summarized the development and deployment of machine-learning-based predictive models for inventory management at Casa Card\u0026atilde;o, a medium-sized Brazilian distributor of building materials. Using Amazon Forecast and SageMaker Canvas, the project implemented a demand-forecasting system that achieved an overall accuracy of approximately 99.31% (WAPE), substantially outperforming the simpler forecasting procedures previously used by the company. The solution helped reduce stockouts, optimize inventory levels, and improve purchasing planning, thereby generating direct operational and economic benefits. At the same time, the study illustrated how cloud-based AutoML platforms can lower the entry barriers for Small and Medium-sized Enterprises (SMEs) that lack specialized data science teams, providing a concrete example of how AI-driven demand forecasting can support digital transformation in supply chain management.\u003c/p\u003e \u003cp\u003eFrom a theoretical and practical standpoint, the study contributes by compiling, in detail, an end-to-end implementation of ML in an SME supply-chain context, including the technology architecture, data pipeline, and human\u0026ndash;machine interaction protocols. It reinforces the evidence that advanced ML models, when combined with managerial judgment and appropriate performance monitoring, can capture nonlinear and seasonal patterns that traditional statistical methods may miss, especially in volatile demand environments such as construction supplies distribution. The project also highlights the importance of governance mechanisms\u0026mdash;such as continuous accuracy monitoring, retraining triggers, and feedback loops between forecasts and operational decisions\u0026mdash;to sustain predictive performance over time.\u003c/p\u003e \u003cp\u003eHowever, as a practice and policy paper based on a single case, this study has several limitations that should be acknowledged. First, the proof of concept and subsequent deployment were conducted in one company, operating in a specific regional and industry context (construction supplies distribution in Minas Gerais, Brazil). This context dependence constrains the generalizability of the results to other industries, countries, or supply-chain configurations. Second, access to data was limited in both breadth and depth: some SKUs had relatively short or irregular historical time series; several potentially relevant external variables (e.g., detailed competitor actions, granular macroeconomic indicators, and micro-regional construction activity) were either unavailable or only partially integrated; and data quality issues required extensive pre-processing. These constraints may have influenced model performance and reduced the ability to test more sophisticated feature-engineering strategies.\u003c/p\u003e \u003cp\u003eFurther limitations arise from the technological choices. The solution relied primarily on AWS managed services (Amazon Forecast and SageMaker Canvas) and benefited from temporary cloud credits, which may not reflect the long-term cost structure faced by all SMEs. There is also a degree of vendor lock-in and opacity associated with proprietary algorithms, which can limit transparency, reproducibility, and portability to other environments. Additionally, the evaluation period for business outcomes was relatively short and focused mainly on forecasting metrics and qualitative managerial perceptions, rather than on a full cost\u0026ndash;benefit analysis including financial indicators such as return on investment, working-capital reduction, or service-level improvements over multiple years. Finally, the study did not systematically assess organizational change dimensions, such as user adoption, learning curves, and the impact of the system on roles, routines, and decision-making culture.\u003c/p\u003e \u003cp\u003eThese limitations open several promising avenues for future work. One important direction is the design, implementation, and evaluation of an open-source reference model that can replicate the core functionalities of the solution with even lower marginal costs and greater transparency. Such a model could be built using open-source libraries and frameworks and released under a permissive license, enabling replication, adaptation, and peer review by researchers, practitioners, and policy-makers. Comparative studies could then benchmark proprietary cloud services against fully open-source stacks\u0026mdash;considering not only predictive accuracy, but also total cost of ownership, explainability, maintainability, and ease of integration with existing ERP and CRM systems.\u003c/p\u003e \u003cp\u003eFuture research should also extend the empirical scope beyond a single company and sector. Multi-case or cross-industry studies\u0026mdash;including food retail, pharmaceuticals, automotive parts, and other sectors with complex, intermittent demand\u0026mdash;could test the robustness of the proposed architecture and identify contingencies that moderate its effectiveness (e.g., product variety, demand volatility, lead times, or data availability). Longitudinal designs could assess the long-term financial impact of ML-based demand forecasting on inventory turnover, stockout rates, service levels, and profitability, ideally using quasi-experimental designs or A/B tests comparing business units with and without the predictive system.\u003c/p\u003e \u003cp\u003eAnother promising stream of work involves integrating demand forecasting with optimization models and prescriptive analytics. Rather than stopping at forecasting, future applications could combine predictions with optimization routines for order quantities, safety stocks, and replenishment policies, thus forming closed-loop decision-support systems. Research could also explore sociotechnical aspects in greater depth, including user trust in algorithmic recommendations, the design of explainable interfaces, training programs for managers, and governance frameworks for AI in SMEs. Finally, policy-oriented studies could examine how public agencies, development banks, and innovation programs might support the diffusion of low-cost, open, and responsible AI solutions for inventory management among small and medium-sized firms.\u003c/p\u003e \u003cp\u003eThis study concludes that advanced cloud-based machine-learning techniques can substantially improve demand forecasting and inventory management in a medium-sized construction supplies distributor, but the study also warns that such solutions are not off-the-shelf panaceas. Their success depends on data availability and quality, careful technological choices, sustained organizational engagement, and appropriate governance. Recognizing these limitations and exploring the proposed possibilities for future work may help researchers, practitioners, and policy-makers design more robust, transparent, and accessible predictive systems\u0026mdash;contributing to the broader goal of democratizing AI-enabled supply-chain optimization for SMEs.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eEthical Approval and Consent to Participate\u003c/h2\u003e\n\u003cp\u003eThis study did not involve human participants, personal data, or interventions requiring ethical approval. The research was conducted using organizational and operational data provided by a private company under a confidentiality agreement. Therefore, ethical approval and consent to participate are \u003cstrong\u003enot applicable\u003c/strong\u003e.\u003c/p\u003e\n\u003ch2\u003eConsent for Publication\u003c/h2\u003e\n\u003cp\u003eAll authors have reviewed and approved the final version of the manuscript and consent to its publication.\u003c/p\u003e\n\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eThis research was supported by the \u003cstrong\u003eMAI/DAI Call No. 68/2022\u003c/strong\u003e, funded by the \u003cstrong\u003eConselho Nacional de Desenvolvimento Cient\u0026iacute;fico e Tecnol\u0026oacute;gico (CNPq)\u003c/strong\u003e, under the \u003cstrong\u003eMinistry of Science, Technology and Innovation of Brazil\u003c/strong\u003e.\u003c/p\u003e\n\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\n\u003cp\u003eThe research group\u0026apos;s coordination and the primary manuscript\u0026apos;s composition were undertaken by D.D.G. R.B.C. served as the doctoral supervisor and performed a comprehensive review of the document. M.A.H., in conjunction with B.K.G.N.C., bore responsibility for the application\u0026apos;s development and ongoing support during the Proof of Concept phase, in addition to authoring the front-end interface.\u003c/p\u003e\n\u003ch2\u003eAcknowledgement\u003c/h2\u003e\n\u003cp\u003eWe want to thank the Brazilian Ministry of Science and Technology, primarily through CNPq and the MAI/DAI 68 call for proposals in 2022, for providing funding for the development of this research.\u003c/p\u003e\n\u003ch2\u003eData Availability\u003c/h2\u003e\n\u003cp\u003eData Availability StatementThe data that support the findings of this study are not publicly available due to confidentiality agreements with the participating organization and the presence of commercially sensitive information. The datasets were used under license for the current study and are therefore not available for public disclosure. Aggregated and anonymized data, as well as methodological details necessary to replicate the analytical procedures, are included within the article. Additional information may be made available from the corresponding author upon reasonable request, subject to approval by the participating organization.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAmazon Web Services. (2025a). \u003cem\u003eAmazon Forecast: Developer guide\u003c/em\u003e. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://docs.aws.amazon.com/forecast/\u003c/span\u003e\u003cspan address=\"https://docs.aws.amazon.com/forecast/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmazon Web Services. (2025b). \u003cem\u003eAmazon SageMaker Canvas: User guide\u003c/em\u003e. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://docs.aws.amazon.com/sagemaker/latest/dg/canvas.html\u003c/span\u003e\u003cspan address=\"https://docs.aws.amazon.com/sagemaker/latest/dg/canvas.html\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmpazis, N. (2015). Forecasting demand in supply chain using machine learning algorithms. \u003cem\u003eInternational Journal of Artificial Life Research\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e(1), 56\u0026ndash;73.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBook, M., Riedel, M., Neukirchen, H., \u0026amp; Erlingsson, E. (2022). Facilitating collaboration in machine learning and high-performance computing projects with an interaction room. In \u003cem\u003e2022 IEEE 18th International Conference on e-Science (e-Science)\u003c/em\u003e (pp. 529\u0026ndash;538). IEEE. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/eScience55777.2022.00093 ResearchGate\u0026thinsp;+\u0026thinsp;1\u003c/span\u003e\u003cspan address=\"10.1109/eScience55777.2022.00093 ResearchGate\u0026thinsp;+\u0026thinsp;1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBowerman, B. L., O\u0026rsquo;Connell, R. T., \u0026amp; Koehler, A. B. (2005). \u003cem\u003eForecasting, time series, and regression: An applied approach\u003c/em\u003e. Cengage Learning.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCarbonneau, R., Laframboise, K., \u0026amp; Vahidov, R. (2009). Application of machine learning techniques for supply chain demand forecasting. \u003cem\u003eEuropean Journal of Operational Research\u003c/em\u003e, \u003cem\u003e184\u003c/em\u003e(3), 1140\u0026ndash;1154. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ejor.2006.12.004\u003c/span\u003e\u003cspan address=\"10.1016/j.ejor.2006.12.004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCarvalho, F. P., dos Santos, R. C., Nascimento, S. M., da Silva Coutinho, J. C., \u0026amp; de Sousa, R. R. (2023). Investigating the relationship between academia and the information technology industry: A systematic literature review. \u003cem\u003eConcilium\u003c/em\u003e, \u003cem\u003e23\u003c/em\u003e(21), 11\u0026ndash;35.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFildes, R., Goodwin, P., Lawrence, M., \u0026amp; Nikolopoulos, K. (2009). Effective forecasting and judgmental adjustments: An empirical evaluation and strategies for improvement in supply-chain planning. \u003cem\u003eInternational Journal of Forecasting\u003c/em\u003e, \u003cem\u003e25\u003c/em\u003e(1), 3\u0026ndash;23. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ijforecast.2008.11.010\u003c/span\u003e\u003cspan address=\"10.1016/j.ijforecast.2008.11.010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHyndman, R. J., \u0026amp; Athanasopoulos, G. (2021). \u003cem\u003eForecasting: Principles and practice\u003c/em\u003e (3rd ed.). OTexts. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://otexts.com/fpp3/\u003c/span\u003e\u003cspan address=\"https://otexts.com/fpp3/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKo, H., Witherell, P., Ndiaye, N. Y., \u0026amp; Lu, Y. (2019). Machine learning based continuous knowledge engineering for additive manufacturing. In \u003cem\u003e2019 IEEE 15th International Conference on Automation Science and Engineering (CASE)\u003c/em\u003e (pp. 648\u0026ndash;654). IEEE. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/COASE.2019.8843316\u003c/span\u003e\u003cspan address=\"10.1109/COASE.2019.8843316\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e ACM Digital Library\u0026thinsp;+\u0026thinsp;1\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLikhar, P. K., Jha, A., Tiwari, S., Sunar, A., Shahu, S., \u0026amp; Thate, S. (2023). Machine learning-based sales prediction and inventory management for grocery stores. \u003cem\u003eInternational Journal of Advanced Research in Science, Communication and Technology.\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48175/ijarsct-13605\u003c/span\u003e\u003cspan address=\"10.48175/ijarsct-13605\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMakridakis, S., Spiliotis, E., \u0026amp; Assimakopoulos, V. (2018). Statistical and machine learning forecasting methods: Concerns and ways forward. \u003cem\u003ePLOS ONE\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(3), e0194889. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0194889\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0194889\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNaik, H., Yashwanth, K., Suraj, P., \u0026amp; Jayapandian, N. (2022). Machine learning based food sales prediction using random forest regression. In \u003cem\u003e2022 6th International Conference on Electronics, Communication and Aerospace Technology\u003c/em\u003e (pp. 998\u0026ndash;1004).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePunia, S., Nikolopoulos, K., Singh, S. P., Madaan, J. K., \u0026amp; Litsiou, K. (2020). Deep learning with long short-term memory networks and random forests for demand forecasting in multi-channel retail. \u003cem\u003eInternational Journal of Production Research\u003c/em\u003e, \u003cem\u003e58\u003c/em\u003e(16), 4964\u0026ndash;4979.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSinduscon-MG. (2024). \u003cem\u003eSector report on the construction industry in Minas Gerais 2023\u0026ndash;2024: Market for cement, mortar, and ceramic tiles\u003c/em\u003e. Minas Gerais Construction Industry Union.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSokol, K., \u0026amp; Flach, P. (2020). One explanation does not fit all: The promise of interactive explanations for machine learning transparency. \u003cem\u003eKI \u0026ndash; K\u0026uuml;nstliche Intelligenz\u003c/em\u003e, \u003cem\u003e34\u003c/em\u003e(2), 235\u0026ndash;250. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s13218-020-00637-y\u003c/span\u003e\u003cspan address=\"10.1007/s13218-020-00637-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSyntetos, A. A., Babai, Z., Boylan, J. E., Kolassa, S., \u0026amp; Nikolopoulos, K. (2016). Supply chain forecasting: Theory, practice, their gap and the future. \u003cem\u003eEuropean Journal of Operational Research\u003c/em\u003e, \u003cem\u003e252\u003c/em\u003e(1), 1\u0026ndash;26. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ejor.2015.11.010\u003c/span\u003e\u003cspan address=\"10.1016/j.ejor.2015.11.010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTian, X., Wang, H., \u0026amp; Erjiang, E. (2021). Forecasting intermittent demand for inventory management by retailers: A new approach. \u003cem\u003eJournal of Retailing and Consumer Services\u003c/em\u003e, \u003cem\u003e62\u003c/em\u003e, 102662. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jretconser.2021.102662\u003c/span\u003e\u003cspan address=\"10.1016/j.jretconser.2021.102662\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTirkolaee, E. B., Sadeghi, S., Mooseloo, F. M., Vandchali, H. R., \u0026amp; Aeini, S. (2021). Application of machine learning in supply chain management: A comprehensive overview of the main areas. \u003cem\u003eMathematical Problems in Engineering, 2021\u003c/em\u003e, 5566842. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1155/2021/5566842\u003c/span\u003e\u003cspan address=\"10.1155/2021/5566842\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Steenbergen, R. M., \u0026amp; Mes, M. R. K. (2020). Forecasting demand profiles of new products. \u003cem\u003eDecision Support Systems\u003c/em\u003e, \u003cem\u003e139\u003c/em\u003e, 113401. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.dss.2020.113401\u003c/span\u003e\u003cspan address=\"10.1016/j.dss.2020.113401\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVairagade, N., Logofatu, D., Leon, F., \u0026amp; Muharemi, F. (2019). Demand forecasting using random forest and artificial neural network for supply chain management. In \u003cem\u003eComputational Collective Intelligence: 11th International Conference, ICCCI 2019, Proceedings, Part I\u003c/em\u003e (pp. 328\u0026ndash;339). Springer. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/978-3-030-28374-0_27\u003c/span\u003e\u003cspan address=\"10.1007/978-3-030-28374-0_27\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, X., \u0026amp; Thomson, V. (2019). Modelling the development of complex products using a knowledge perspective. \u003cem\u003eResearch in Engineering Design\u003c/em\u003e, \u003cem\u003e30\u003c/em\u003e, 203\u0026ndash;226.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Predictive models, Artificial Intelligence, Machine Learning, Inventory Management, Supply Chain Management (SCM)","lastPublishedDoi":"10.21203/rs.3.rs-8502995/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8502995/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eInventory management is a critical task for distribution companies, as forecasting errors can lead to excess inventory, higher storage costs, or stockouts. This practice \u0026amp; policy paper describes the development and deployment of a machine learning-based predictive model designed to optimize inventory management for a medium-sized Brazilian construction supplies distributor. Using the Amazon Forecast and Amazon SageMaker Canvas platforms, the study demonstrated how advanced demand forecasting techniques reduce stockouts and enhance operational efficiency. The research compared various predictive algorithms, assessing their performance with metrics such as RMSE, MAPE, MASE, and WAPE. The results showed that the implemented model achieved approximately 99.31% forecast accuracy, offering significant benefits like fewer stockouts, optimized inventory levels, and improved purchasing planning. This practice study provides technical and managerial recommendations for companies seeking to predict their inventory needs accurately, demonstrating the machine learning capabilities for Supply Chain Management.\u003c/p\u003e","manuscriptTitle":"Predictive Models for Inventory Optimization: a machine learning application for demand forecasting at a construction supplies distributor","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-16 11:47:29","doi":"10.21203/rs.3.rs-8502995/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"1c2feb21-04ba-4d4d-9a37-113dab35357e","owner":[],"postedDate":"January 16th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-07T16:12:31+00:00","versionOfRecord":{"articleIdentity":"rs-8502995","link":"https://doi.org/10.1186/s43093-026-00807-8","journal":{"identity":"future-business-journal","isVorOnly":false,"title":"Future Business Journal"},"publishedOn":"2026-03-31 15:58:09","publishedOnDateReadable":"March 31st, 2026"},"versionCreatedAt":"2026-01-16 11:47:29","video":"","vorDoi":"10.1186/s43093-026-00807-8","vorDoiUrl":"https://doi.org/10.1186/s43093-026-00807-8","workflowStages":[]},"version":"v1","identity":"rs-8502995","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8502995","identity":"rs-8502995","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}