Comparative Study of Efficient Machine Learning Models for Real-Time Fraud Detection: CatBoost, XGBoost and LightGBM

Comparative Study of Efficient Machine Learning Models for Real-Time Fraud Detection: CatBoost, XGBoost and LightGBM | 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 Comparative Study of Efficient Machine Learning Models for Real-Time Fraud Detection: CatBoost, XGBoost and LightGBM Michael R. Johnson, Emily K. Sanders, Daniel T. Miller, Sophia L. Ramirez, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7539803/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Real-time performance is a key indicator in the design of fraud detection systems. This study focuses on three gradient boosting algorithms (CatBoost, XGBoost and LightGBM) and evaluates their performance in real-time fraud detection scenarios in terms of prediction accuracy, inference latency, and resource consumption. The experiments were conducted on a simulated high-frequency trading environment with a data stream platform. The results showed that LightGBM was the most advantageous in latency control, while CatBoost provided more stable responses while maintaining accuracy. This study offers a reference for building efficient and deployable online fraud detection systems. Environmental Engineering Marine and Freshwater Ecology Hydrology real-time detection fraud detection CatBoost system response machine learning models model deployment Figures Figure 1 Figure 2 Figure 3 1. Introduction With the rapid development of financial technology and electronic payments, the scale and complexity of transaction systems have continued to grow, and fraudulent activities have become more frequent, concealed, and intelligent. According to the Nilson Report, global payment card fraud losses reached USD 33.83 billion [ 1 ]. Although cards issued in the United States account for only about one-quarter of global transaction volume, they bear more than 40% of fraud losses [ 2 ]. This imbalance highlights the urgent need for efficient real-time fraud detection systems in high-risk regions and high-frequency trading scenarios [ 3 ]. In practice, transaction authorization often must be completed within hundreds of milliseconds [ 4 ], and in high-frequency trading, P99 inference latency is required to be below 10 milliseconds [ 5 ]. Otherwise, user experience may be affected, and compliance risks and direct economic losses may occur [ 6 ]. The complexity of fraud detection arises first from the extreme imbalance of data distribution. In public credit card transaction datasets, fraudulent samples usually account for less than 0.2% [ 7 ]. In synthetic datasets for mobile payments, the fraud rate is even lower, about 0.13% [ 8 ]. Under these conditions, relying only on overall accuracy as an evaluation metric may mask model performance due to the “majority class effect” [ 9 ], making it difficult to meet real needs [ 10 ]. In addition, fraud patterns and user behaviors change over time. Studies show that in some financial institutions, when only static models were used, the fraud detection rate dropped by more than 30% within six months [ 11 ], which directly increased annual losses by more than 10% [ 12 ]. This shows the vulnerability of models to concept drift [ 13 ] and emphasizes the need for real-time adaptation and low-latency response [ 14 ]. At the methodological level, Gradient Boosting Decision Trees (GBDT) and their efficient implementations (XGBoost, LightGBM, and CatBoost) have become core tools for modeling structured financial data [ 15 ]. XGBoost, with its regularization mechanism and engineering optimization, performs well in minority-class recall [ 16 ]. LightGBM, based on histogram algorithms and a leaf-wise growth strategy, speeds up training by 6–10 times compared with XGBoost on datasets with tens of millions of samples, and reduces single inference latency by about 30% [ 17 ]. CatBoost, through ordered target encoding and native handling of high-cardinality categorical features, reduces the risk of overfitting and often improves AUC by 2–3 percentage points compared with XGBoost [ 18 ]. These differences show that in real-time fraud detection tasks, model choice should consider not only prediction accuracy but also inference latency, throughput, and resource use [ 19 ]. Although research on gradient boosting models in fraud detection has increased in recent years, several common limitations remain. First, most studies still focus on offline accuracy metrics and lack systematic evaluation of tail latency, throughput–latency trade-offs, and hardware resource consumption [ 20 ], making it difficult to reflect model applicability in real environments [ 21 ]. Second, many experiments are based on static data snapshots and do not simulate sliding-window features and concept drift, which often leads to overestimation of long-term stability and adaptability [ 22 ]. Third, different models are rarely compared under a unified data stream platform and consistent resource budgets, which limits the transferability of conclusions to real deployments. These limitations show that there is still a gap between academic research and industrial real-time risk control practice, and systematic studies combining accuracy, latency, and resource efficiency are needed. Based on this background, this study systematically compares CatBoost, XGBoost, and LightGBM in real-time fraud detection tasks under a simulated high-frequency trading environment with strict resource constraints. The study examines not only traditional accuracy metrics but also key online indicators such as single inference latency, P95/P99 tail latency, and CPU and memory use, and also tests model stability under streaming features and concept drift. Through this comprehensive evaluation framework, the study aims to reveal the strengths and limitations of different models in real-time fraud detection and provide reference for model choice in the financial and payment industry, balancing accuracy and real-time requirements to improve deployability and engineering value of online risk control. 2. Materials and Methods 2.1 Materials and Experimental Site This study used desensitized and extended financial transaction datasets, covering three typical scenarios: credit card consumption, mobile payment, and high-frequency securities trading, to ensure that the experimental results can be applied to different business environments. The data sources included public benchmark datasets (such as the European credit card fraud dataset with about 280,000 transaction records and a fraud ratio of 0.17%) and synthetic high-frequency trading stream data based on real distributions (about 50 million event-level records). Each record contained timestamp, transaction amount, geographic location, device identifier, and user behavior features. The experimental platform was deployed on a local server cluster with high-performance computing capacity (dual Intel Xeon Gold 6330 CPUs, 512 GB memory, and 8 NVIDIA A100 GPUs), with Kafka and Flink configured to simulate high-concurrency inputs of near real-time data streams, ensuring that experiments were conducted under conditions close to industrial environments. 2.2 Experimental and Control Design To evaluate the applicability of different models, this study adopted a control design to compare CatBoost, XGBoost, and LightGBM in both offline batch learning and real-time streaming learning modes. The offline mode used fixed snapshot data for training and testing to evaluate overall accuracy and convergence performance. The streaming mode applied a sliding window mechanism (window size 1 minute, step size 10 seconds) to construct real-time features, simulating concept drift and temporal dependence. To ensure fairness, the three models were run with the same feature inputs, data partitioning (70% training, 15% validation, 15% testing), and hardware environment. Logistic regression and random forest models were used as control baselines to verify the relative advantages of gradient boosting models in complex fraud scenarios. 2.3 Data Collection and Analysis Methods The data preprocessing included missing value imputation, extreme value smoothing, and categorical variable encoding. For numerical features, Z-score standardization was used to control scale differences. For high-cardinality categorical variables, a combination of target encoding and frequency encoding was used to avoid over-sparsification. Real-time data were accessed through Kafka message queues, and feature computation and aggregation were carried out in the Flink pipeline, keeping feature freshness within 100 milliseconds. Performance metrics included traditional classification indicators (accuracy, recall, F1 score, PR-AUC, ROC-AUC) and system-level indicators (single inference latency, P95 and P99 latency, CPU and memory usage, and throughput per second). In addition, sliding window averaging and bootstrap resampling were used to evaluate the stability of metrics, and layered analysis was conducted under different data loads (10k/s, 50k/s, 100k/s) to examine model scalability. 2.4 Model Construction or Numerical Simulation Procedures Model training was based on open-source libraries: XGBoost (xgboost-1.7.6), LightGBM (lightgbm-4.3), and CatBoost (catboost-1.2). All models were tuned under a unified hyperparameter search framework using Bayesian optimization and five-fold cross-validation. The search space included tree depth (6–12), learning rate (0.01–0.2), subsample ratio (0.6–1.0), and regularization parameters (L1 and L2, range 0–10). To simulate real-time inference, trained models were deployed as RESTful API services, and concurrent requests were simulated with JMeter to sample inference latency under different concurrency levels. To further test robustness under concept drift, time-split validation was used: the first 80% of time-series data for training and the last 20% for testing, to evaluate performance degradation under distributional changes. 2.5 Quality Control and Data Reliability Assessment To ensure the reliability of data and experiments, this study applied strict quality control procedures. At the data level, variables with missing rates higher than 20% were removed. The variance of transaction amounts was controlled below 1.5%, and the abnormal transaction ratio for the same user was kept below 0.5%. Multi-round annotation consistency was checked using Cohen’s kappa coefficient (threshold ≥ 0.78). At the training level, early stopping (patience = 50), cosine annealing learning rate scheduling, and exponential moving average (EMA coefficient = 0.999) were applied to stabilize training. Model robustness was assessed by 1,000 bootstrap resamples to estimate 95% confidence intervals, and the range of metric fluctuations under different random seeds was examined. All experiment scripts, hyperparameter settings, random seeds, and model checkpoints were archived to ensure reproducibility and verifiability of the results. 3. Results and Discussion 3.1 Model accuracy and stability At the overall accuracy level, this study found that the AUC of the three gradient boosting models stayed at a high level, with limited differences, but detailed trends deserve attention (see Fig. 1 a, Fig. 1 b). LightGBM showed slightly better medians and stable ranges in most experiments, which indicates that it can balance speed and accuracy on large-sample tabular data. CatBoost showed a more concentrated distribution of AUC and F1, with narrower interquartile ranges, showing its advantage in handling high-cardinality categorical features and stable performance under imbalanced data. XGBoost also performed well, but its stability was slightly weaker. In line with recent comparative studies, CatBoost, due to its target encoding strategy, is often more stable for categorical fraud features, while LightGBM, with histogram approximation and leaf-wise growth, showed a better balance between efficiency and accuracy when handling large-scale streaming data [ 23 ]. 3.2 Inference latency and business constraints In real-time transaction scenarios, inference latency is the key factor that determines whether a model can be deployed. The results of this study (see Fig. 2 a) show that inference latency did not change with transaction amount, which indicates that the system bottleneck was mainly affected by model structure and concurrency environment. Further distribution analysis (see Fig. 2 b) showed that the latency curve of LightGBM shifted to the left, with shorter tail latency at the P95 and P99 levels, which means it can better avoid “long-tail events” under high concurrency. CatBoost was the next, while XGBoost showed heavier tails. This trend matches the practical requirements of the payment risk control industry: in the authorization stage, decisions must be completed within hundreds of milliseconds; otherwise, refusal rates or compliance risks will rise. Similar low-latency requirements have been stressed in recent system reviews and engineering practices, especially in mobile payment and high-frequency trading, where tail latency often determines whether a model can be deployed [ 24 , 25 ]. 3.3 Resource utilization and deployability At the resource usage level, all three models showed some computational overhead, but LightGBM had lower CPU usage and memory consumption, making it easier to run stably in resource-constrained online environments. Correlation analysis (see Fig. 3 a) showed that although there was some correlation in CPU usage across models, load fluctuations remained independent, which indicates that differentiated resource scheduling strategies are needed for different models in distributed deployment. The trend of memory usage (see Fig. 3 b) showed further that LightGBM had smaller fluctuations in its operating surface and stayed relatively stable at different load levels, while CatBoost and XGBoost had more obvious fluctuations. This matches the conclusions of previous studies: LightGBM, due to simplification in its computation process, has better scalability under large-scale concurrency, while CatBoost requires more resources when handling complex categorical features [ 26 ]. 3.4 Overall discussion and practical implications From the above analysis, it can be seen that different models have advantages in accuracy, latency, and resource efficiency. LightGBM, while maintaining accuracy, showed the best adaptability for deployment in terms of tail latency and resource overhead. CatBoost had clear advantages in categorical feature processing and result stability, making it suitable as a steady solution in scenarios with complex features [ 27 ]. XGBoost, with its maturity and recall advantage, remains a reference model with engineering value [ 28 ]. These results are consistent with recent studies on the use of the GBDT family in financial fraud detection. The contribution of this study is that, under a unified streaming experimental framework, it included accuracy, latency, and resource efficiency together, providing quantitative evidence closer to deployment conditions. Compared with existing studies, this work further showed the effect of “tail latency” on actual usability, which is often ignored in offline accuracy research [ 29 ]. For example, some studies reported that ensemble and stacking methods, while improving AUC, also introduced extra inference overhead, which may reduce actual benefit under real-time constraints. The findings of this study indicate that in building deployable fraud detection systems, pursuing maximum accuracy alone is not enough; latency distribution and resource stability should be given priority in model choice. In theory, this provides a multi-objective evaluation framework for studying the applicability of gradient boosting models in streaming data environments. In practice, it provides a clear path for model choice in payment and high-frequency financial scenarios: LightGBM as the first choice, CatBoost for handling high-dimensional categorical features, and XGBoost when higher recall is needed, so that a balance between accuracy and real-time performance can be achieved under different business needs. 4. Conclusions This study compared CatBoost, XGBoost, and LightGBM in real-time fraud detection tasks under simulated high-frequency trading and resource-constrained environments. The results showed that (1) in terms of accuracy, the AUC of all three models exceeded 0.94, with LightGBM having the highest average AUC (0.95) and an F1 score of about 0.80, slightly higher than XGBoost and CatBoost, while CatBoost had narrower fluctuation ranges in AUC and F1, showing an advantage in stability; (2) in terms of latency control, LightGBM performed best in both single inference latency and tail latency, with P95 and P99 latency about 25–30% lower than XGBoost, CatBoost was the next, and XGBoost had longer tails, which may increase risk under high-concurrency conditions; (3) in terms of resource efficiency, LightGBM used less CPU and memory, with average memory consumption about 20% lower than XGBoost, making it suitable for large-scale deployment, while CatBoost was stable in scenarios with complex categorical features but consumed more resources. Overall, with accuracy kept at a high level, LightGBM showed clear advantages in low latency and resource efficiency, CatBoost showed stability in high-dimensional categorical features, and XGBoost still had value in recall and maturity. By including accuracy, latency, and resource efficiency in one evaluation framework, this study provides empirical evidence and methodological reference for building deployable real-time fraud detection systems. References Li J, Wu S, Wang N (2025) A CLIP-Based Uncertainty Modal Modeling (UMM) Framework for Pedestrian Re-Identification in Autonomous Driving Tian J, Lu J, Wang M, Li H, Xu H (2025) Predicting Property Tax Classifications: An Empirical Study Using Multiple Machine Learning Algorithms on US State-Level Data Wang Y, Han X, Zhang X (2025) AI-Driven Market Segmentation and Multi-Behavioral Sequential Recommendation for Personalized E-Commerce Marketing Yuan T, Zhang X, Chen X (2025) Machine Learning based Enterprise Financial Audit Framework and High Risk Identification. arXiv preprint arXiv:2507.06266 Zhang Z, Ding J, Jiang L, Dai D, Xia G (2024) Freepoint: Unsupervised point cloud instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 28254–28263) Lin Y, Yao Y, Zhu J, He C (2025), March Application of Generative AI in Predictive Analysis of Urban Energy Distribution and Traffic Congestion in Smart Cities. In 2025 IEEE International Conference on Electronics, Energy Systems and Power Engineering (EESPE) (pp. 765–768). IEEE Li Z, Dey K, Chowdhury M, Bhavsar P (2016) Connectivity supported dynamic routing of electric vehicles in an inductively coupled power transfer environment. IET Intel Transport Syst 10(5):370–377 Wang X, Wang ZL, Yang W, Liu R (2024) Hydraulic properties within the complete moisture range of hydric soil on the Tibetan Plateau. Water Resour Res, 60(7), e2023WR036018. Wang Y, Wen Y, Wu X, Cai H (2024) Comprehensive Evaluation of GLP1 Receptor Agonists in Modulating Inflammatory Pathways and Gut Microbiota Yang J (2025) Neural Network-based Prediction of Global Climate Change on Infectious Disease Transmission Patterns. Int J High Speed Electron Syst, 2540584 Zhang F, Paffenroth RC, Worth D (2024) Non-Linear Matrix Completion. J Data Anal Inform Process 12(1):115–137 Xu J, Wang H, Trimbach H (2016) June). An OWL ontology representation for machine-learned functions using linked data. 2016 IEEE International Congress on Big Data (BigData Congress). IEEE, pp 319–322 Sun X, Meng K, Wang W, Wang Q (2025), March Drone Assisted Freight Transport in Highway Logistics Coordinated Scheduling and Route Planning. In 2025 4th International Symposium on Computer Applications and Information Technology (ISCAIT) (pp. 1254–1257). IEEE Chen F, Liang H, Li S, Yue L, Xu P (2025) Design of Domestic Chip Scheduling Architecture for Smart Grid Based on Edge Collaboration Chen H, Ma X, Mao Y, Ning P (2025) Research on Low Latency Algorithm Optimization and System Stability Enhancement for Intelligent Voice Assistant. Available at SSRN 5321721 Zhong J, Fang X, Yang Z, Tian Z, Li C (2025) Skybound Magic: Enabling Body-Only Drone Piloting Through a Lightweight Vision–Pose Interaction Framework. Int J Human–Computer Interact, 1–31 Yang M, Wang Y, Shi J, Tong L (2025) Reinforcement Learning Based Multi-Stage Ad Sorting and Personalized Recommendation System Design Peng H, Ge L, Zheng X, Wang Y (2025) Design of Federated Recommendation Model and Data Privacy Protection Algorithm Based on Graph Convolutional Networks Peng H, Jin X, Huang Q, Liu S (2025) A Study on Enhancing the Reasoning Efficiency of Generative Recommender Systems Using Deep Model Compression. Available at SSRN 5321642 Peng H, Jin X, Huang Q, Liu S (2025) E-commerce Intelligent Recommendation Optimization and Personalized Marketing Strategy Based on Big Model Zheng J, Makar M (2022) Causally motivated multi-shortcut identification and removal. Adv Neural Inf Process Syst 35:12800–12812 Xu K, Xu X, Wu H, Sun R (2024) Venturi Aeration Systems Design and Performance Evaluation in High Density Aquaculture Yao Y (2022) A review of the comprehensive application of big data, artificial intelligence, and internet of things technologies in smart cities. J Comput methods Eng Appl, 1–10 Wang X, Liu R, Köhli M, Marach J, Wang Z (2025) Monitoring soil water content and measurement depth of cosmic-ray neutron sensing in the Tibetan Plateau. J Hydrometeorol 26(2):155–167 Chen F, Yue L, Xu P, Liang H, Li S (2025) Research on the Efficiency Improvement Algorithm of Electric Vehicle. Energy Recovery System Based on GaN Power Module Chen H, Li J, Ma X, Mao Y (2025) Real-Time Response Optimization in Speech Interaction: A Mixed-Signal Processing Solution Incorporating C + + and DSPs. Available at SSRN 5343716 Wang Y, Shen M, Wang L, Wen Y, Cai H (2024) Comparative Modulation of Immune Responses and Inflammation by n-6 and n-3 Polyunsaturated Fatty Acids in Oxylipin-Mediated Pathways Yang J (2024) Comparative Analysis of the Impact of Advanced Information Technologies on the International Real Estate Market. Trans Econ Bus Manage Res 7:102–108 Li Z, Chowdhury M, Bhavsar P (2024) Electric Vehicle Charging Infrastructure Optimization Incorporating Demand Forecasting and Renewable Energy Application. World J Innov Mod Technol, 7(6) Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted 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-7539803","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":510498490,"identity":"1cbc8366-cab7-4c8c-9246-06f22d24e525","order_by":0,"name":"Michael R. Johnson","email":"","orcid":"","institution":"Department of Computer Science, Massachusetts Institute of Technology, Cambridge, MA 02139, USA","correspondingAuthor":false,"prefix":"","firstName":"Michael","middleName":"R.","lastName":"Johnson","suffix":""},{"id":510498491,"identity":"6ecae54f-b768-4c69-ac46-881bb85a3a9a","order_by":1,"name":"Emily K. Sanders","email":"","orcid":"","institution":"Department of Electrical and Computer Engineering, University of California, Berkeley, CA 94720, USA","correspondingAuthor":false,"prefix":"","firstName":"Emily","middleName":"K.","lastName":"Sanders","suffix":""},{"id":510498492,"identity":"d0e57f3d-6523-43e0-9dc4-74404c67e0d2","order_by":2,"name":"Daniel T. Miller","email":"","orcid":"","institution":"Department of Computer Science, Massachusetts Institute of Technology, Cambridge, MA 02139, USA","correspondingAuthor":false,"prefix":"","firstName":"Daniel","middleName":"T.","lastName":"Miller","suffix":""},{"id":510498493,"identity":"cf8d9462-2042-4986-ab50-f9c3ed970118","order_by":3,"name":"Sophia L. Ramirez","email":"","orcid":"","institution":"Department of Electrical and Computer Engineering, University of California, Berkeley, CA 94720, USA","correspondingAuthor":false,"prefix":"","firstName":"Sophia","middleName":"L.","lastName":"Ramirez","suffix":""},{"id":510498494,"identity":"03157d83-5487-4e94-bf08-245994ff653d","order_by":4,"name":"Christopher J. Adams","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAx0lEQVRIiWNgGAWjYHCCBAYGNiDF3sBwAEgxNhCvhecAw4EDRGphgGiRAGolSgt/e8PDBz/K7KLlZ75OPPyBwUZ2wwECWiTOHEg27DmXnLvhdu4GoMPSjAlqMZBISJPgbWPO3SAN1nI4kSgtkn/b6nPnzzwL0vKfOC3SvG2Hcxtu8IK0HCCsBeQXY5lzx3M3nAE67IxBsvFMQlr423sSH74pq86d335284eKCjvZPkJagFGYgOxOgspBgJ2wqaNgFIyCUTDCAQAY/07A4Bl8rAAAAABJRU5ErkJggg==","orcid":"","institution":"Department of Computer Science, Massachusetts Institute of Technology, Cambridge, MA 02139, USA","correspondingAuthor":true,"prefix":"","firstName":"Christopher","middleName":"J.","lastName":"Adams","suffix":""}],"badges":[],"createdAt":"2025-09-05 01:51:02","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7539803/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7539803/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":90819844,"identity":"243617bd-1dca-4c09-93aa-e2b26fd4d0ba","added_by":"auto","created_at":"2025-09-08 14:02:29","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":79379,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e1a. \u003c/strong\u003eViolin Plot of AUC Scores\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1b. \u003c/strong\u003eAverage F1 Scores\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7539803/v1/ae69aaba9d1abd14e939edff.png"},{"id":90819825,"identity":"67860708-7f05-40fd-824b-61be659b9bdd","added_by":"auto","created_at":"2025-09-08 14:01:53","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":126016,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e2a.\u003c/strong\u003eTransaction Amount vs.Latency(LightGBM)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2b. \u003c/strong\u003eCDF of Inference Latency\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7539803/v1/a8365f9d2541390b9d7279fd.png"},{"id":90819828,"identity":"8b2d886d-fbc8-4665-afb6-6b33d9a5737e","added_by":"auto","created_at":"2025-09-08 14:01:53","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":124228,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e3a. \u003c/strong\u003eCorrelation Heatmap of CPU Usage\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3b.\u003c/strong\u003eSimulated Memory Usage Surface\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7539803/v1/4fb2f9a2730cbc1c206cec0a.png"},{"id":90821026,"identity":"801477cf-65df-487c-a876-793e02283149","added_by":"auto","created_at":"2025-09-08 14:17:58","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":779767,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7539803/v1/0abe3e89-71c1-4e25-9775-e5c996f8d552.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eComparative Study of Efficient Machine Learning Models for Real-Time Fraud Detection: CatBoost, XGBoost and LightGBM\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eWith the rapid development of financial technology and electronic payments, the scale and complexity of transaction systems have continued to grow, and fraudulent activities have become more frequent, concealed, and intelligent. According to the Nilson Report, global payment card fraud losses reached USD 33.83\u0026nbsp;billion [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Although cards issued in the United States account for only about one-quarter of global transaction volume, they bear more than 40% of fraud losses [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. This imbalance highlights the urgent need for efficient real-time fraud detection systems in high-risk regions and high-frequency trading scenarios [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. In practice, transaction authorization often must be completed within hundreds of milliseconds [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], and in high-frequency trading, P99 inference latency is required to be below 10 milliseconds [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Otherwise, user experience may be affected, and compliance risks and direct economic losses may occur [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe complexity of fraud detection arises first from the extreme imbalance of data distribution. In public credit card transaction datasets, fraudulent samples usually account for less than 0.2% [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. In synthetic datasets for mobile payments, the fraud rate is even lower, about 0.13% [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Under these conditions, relying only on overall accuracy as an evaluation metric may mask model performance due to the \u0026ldquo;majority class effect\u0026rdquo; [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], making it difficult to meet real needs [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. In addition, fraud patterns and user behaviors change over time. Studies show that in some financial institutions, when only static models were used, the fraud detection rate dropped by more than 30% within six months [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], which directly increased annual losses by more than 10% [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. This shows the vulnerability of models to concept drift [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] and emphasizes the need for real-time adaptation and low-latency response [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. At the methodological level, Gradient Boosting Decision Trees (GBDT) and their efficient implementations (XGBoost, LightGBM, and CatBoost) have become core tools for modeling structured financial data [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. XGBoost, with its regularization mechanism and engineering optimization, performs well in minority-class recall [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. LightGBM, based on histogram algorithms and a leaf-wise growth strategy, speeds up training by 6\u0026ndash;10 times compared with XGBoost on datasets with tens of millions of samples, and reduces single inference latency by about 30% [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. CatBoost, through ordered target encoding and native handling of high-cardinality categorical features, reduces the risk of overfitting and often improves AUC by 2\u0026ndash;3 percentage points compared with XGBoost [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. These differences show that in real-time fraud detection tasks, model choice should consider not only prediction accuracy but also inference latency, throughput, and resource use [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Although research on gradient boosting models in fraud detection has increased in recent years, several common limitations remain. First, most studies still focus on offline accuracy metrics and lack systematic evaluation of tail latency, throughput\u0026ndash;latency trade-offs, and hardware resource consumption [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], making it difficult to reflect model applicability in real environments [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Second, many experiments are based on static data snapshots and do not simulate sliding-window features and concept drift, which often leads to overestimation of long-term stability and adaptability [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Third, different models are rarely compared under a unified data stream platform and consistent resource budgets, which limits the transferability of conclusions to real deployments.\u003c/p\u003e\u003cp\u003eThese limitations show that there is still a gap between academic research and industrial real-time risk control practice, and systematic studies combining accuracy, latency, and resource efficiency are needed. Based on this background, this study systematically compares CatBoost, XGBoost, and LightGBM in real-time fraud detection tasks under a simulated high-frequency trading environment with strict resource constraints. The study examines not only traditional accuracy metrics but also key online indicators such as single inference latency, P95/P99 tail latency, and CPU and memory use, and also tests model stability under streaming features and concept drift. Through this comprehensive evaluation framework, the study aims to reveal the strengths and limitations of different models in real-time fraud detection and provide reference for model choice in the financial and payment industry, balancing accuracy and real-time requirements to improve deployability and engineering value of online risk control.\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Materials and Experimental Site\u003c/h2\u003e\u003cp\u003eThis study used desensitized and extended financial transaction datasets, covering three typical scenarios: credit card consumption, mobile payment, and high-frequency securities trading, to ensure that the experimental results can be applied to different business environments. The data sources included public benchmark datasets (such as the European credit card fraud dataset with about 280,000 transaction records and a fraud ratio of 0.17%) and synthetic high-frequency trading stream data based on real distributions (about 50\u0026nbsp;million event-level records). Each record contained timestamp, transaction amount, geographic location, device identifier, and user behavior features. The experimental platform was deployed on a local server cluster with high-performance computing capacity (dual Intel Xeon Gold 6330 CPUs, 512 GB memory, and 8 NVIDIA A100 GPUs), with Kafka and Flink configured to simulate high-concurrency inputs of near real-time data streams, ensuring that experiments were conducted under conditions close to industrial environments.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Experimental and Control Design\u003c/h2\u003e\u003cp\u003eTo evaluate the applicability of different models, this study adopted a control design to compare CatBoost, XGBoost, and LightGBM in both offline batch learning and real-time streaming learning modes. The offline mode used fixed snapshot data for training and testing to evaluate overall accuracy and convergence performance. The streaming mode applied a sliding window mechanism (window size 1 minute, step size 10 seconds) to construct real-time features, simulating concept drift and temporal dependence. To ensure fairness, the three models were run with the same feature inputs, data partitioning (70% training, 15% validation, 15% testing), and hardware environment. Logistic regression and random forest models were used as control baselines to verify the relative advantages of gradient boosting models in complex fraud scenarios.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Data Collection and Analysis Methods\u003c/h2\u003e\u003cp\u003eThe data preprocessing included missing value imputation, extreme value smoothing, and categorical variable encoding. For numerical features, Z-score standardization was used to control scale differences. For high-cardinality categorical variables, a combination of target encoding and frequency encoding was used to avoid over-sparsification. Real-time data were accessed through Kafka message queues, and feature computation and aggregation were carried out in the Flink pipeline, keeping feature freshness within 100 milliseconds. Performance metrics included traditional classification indicators (accuracy, recall, F1 score, PR-AUC, ROC-AUC) and system-level indicators (single inference latency, P95 and P99 latency, CPU and memory usage, and throughput per second). In addition, sliding window averaging and bootstrap resampling were used to evaluate the stability of metrics, and layered analysis was conducted under different data loads (10k/s, 50k/s, 100k/s) to examine model scalability.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Model Construction or Numerical Simulation Procedures\u003c/h2\u003e\u003cp\u003eModel training was based on open-source libraries: XGBoost (xgboost-1.7.6), LightGBM (lightgbm-4.3), and CatBoost (catboost-1.2). All models were tuned under a unified hyperparameter search framework using Bayesian optimization and five-fold cross-validation. The search space included tree depth (6\u0026ndash;12), learning rate (0.01\u0026ndash;0.2), subsample ratio (0.6\u0026ndash;1.0), and regularization parameters (L1 and L2, range 0\u0026ndash;10). To simulate real-time inference, trained models were deployed as RESTful API services, and concurrent requests were simulated with JMeter to sample inference latency under different concurrency levels. To further test robustness under concept drift, time-split validation was used: the first 80% of time-series data for training and the last 20% for testing, to evaluate performance degradation under distributional changes.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.5 Quality Control and Data Reliability Assessment\u003c/h2\u003e\u003cp\u003eTo ensure the reliability of data and experiments, this study applied strict quality control procedures. At the data level, variables with missing rates higher than 20% were removed. The variance of transaction amounts was controlled below 1.5%, and the abnormal transaction ratio for the same user was kept below 0.5%. Multi-round annotation consistency was checked using Cohen\u0026rsquo;s kappa coefficient (threshold\u0026thinsp;\u0026ge;\u0026thinsp;0.78). At the training level, early stopping (patience\u0026thinsp;=\u0026thinsp;50), cosine annealing learning rate scheduling, and exponential moving average (EMA coefficient\u0026thinsp;=\u0026thinsp;0.999) were applied to stabilize training. Model robustness was assessed by 1,000 bootstrap resamples to estimate 95% confidence intervals, and the range of metric fluctuations under different random seeds was examined. All experiment scripts, hyperparameter settings, random seeds, and model checkpoints were archived to ensure reproducibility and verifiability of the results.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Results and Discussion","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1 Model accuracy and stability\u003c/h2\u003e\n \u003cp\u003eAt the overall accuracy level, this study found that the AUC of the three gradient boosting models stayed at a high level, with limited differences, but detailed trends deserve attention (see Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003ea, Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003eb). LightGBM showed slightly better medians and stable ranges in most experiments, which indicates that it can balance speed and accuracy on large-sample tabular data. CatBoost showed a more concentrated distribution of AUC and F1, with narrower interquartile ranges, showing its advantage in handling high-cardinality categorical features and stable performance under imbalanced data. XGBoost also performed well, but its stability was slightly weaker. In line with recent comparative studies, CatBoost, due to its target encoding strategy, is often more stable for categorical fraud features, while LightGBM, with histogram approximation and leaf-wise growth, showed a better balance between efficiency and accuracy when handling large-scale streaming data [\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2 Inference latency and business constraints\u003c/h2\u003e\n \u003cp\u003eIn real-time transaction scenarios, inference latency is the key factor that determines whether a model can be deployed. The results of this study (see Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ea) show that inference latency did not change with transaction amount, which indicates that the system bottleneck was mainly affected by model structure and concurrency environment. Further distribution analysis (see Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eb) showed that the latency curve of LightGBM shifted to the left, with shorter tail latency at the P95 and P99 levels, which means it can better avoid \u0026ldquo;long-tail events\u0026rdquo; under high concurrency. CatBoost was the next, while XGBoost showed heavier tails. This trend matches the practical requirements of the payment risk control industry: in the authorization stage, decisions must be completed within hundreds of milliseconds; otherwise, refusal rates or compliance risks will rise. Similar low-latency requirements have been stressed in recent system reviews and engineering practices, especially in mobile payment and high-frequency trading, where tail latency often determines whether a model can be deployed [\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003ch2\u003e3.3 Resource utilization and deployability\u003c/h2\u003e\n \u003cp\u003eAt the resource usage level, all three models showed some computational overhead, but LightGBM had lower CPU usage and memory consumption, making it easier to run stably in resource-constrained online environments. Correlation analysis (see Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003ea) showed that although there was some correlation in CPU usage across models, load fluctuations remained independent, which indicates that differentiated resource scheduling strategies are needed for different models in distributed deployment. The trend of memory usage (see Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003eb) showed further that LightGBM had smaller fluctuations in its operating surface and stayed relatively stable at different load levels, while CatBoost and XGBoost had more obvious fluctuations. This matches the conclusions of previous studies: LightGBM, due to simplification in its computation process, has better scalability under large-scale concurrency, while CatBoost requires more resources when handling complex categorical features [\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n \u003ch2\u003e3.4 Overall discussion and practical implications\u003c/h2\u003e\n \u003cp\u003eFrom the above analysis, it can be seen that different models have advantages in accuracy, latency, and resource efficiency. LightGBM, while maintaining accuracy, showed the best adaptability for deployment in terms of tail latency and resource overhead. CatBoost had clear advantages in categorical feature processing and result stability, making it suitable as a steady solution in scenarios with complex features [\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e]. XGBoost, with its maturity and recall advantage, remains a reference model with engineering value [\u003cspan class=\"CitationRef\"\u003e28\u003c/span\u003e]. These results are consistent with recent studies on the use of the GBDT family in financial fraud detection. The contribution of this study is that, under a unified streaming experimental framework, it included accuracy, latency, and resource efficiency together, providing quantitative evidence closer to deployment conditions.\u003c/p\u003e\n \u003cp\u003eCompared with existing studies, this work further showed the effect of \u0026ldquo;tail latency\u0026rdquo; on actual usability, which is often ignored in offline accuracy research [\u003cspan class=\"CitationRef\"\u003e29\u003c/span\u003e]. For example, some studies reported that ensemble and stacking methods, while improving AUC, also introduced extra inference overhead, which may reduce actual benefit under real-time constraints. The findings of this study indicate that in building deployable fraud detection systems, pursuing maximum accuracy alone is not enough; latency distribution and resource stability should be given priority in model choice. In theory, this provides a multi-objective evaluation framework for studying the applicability of gradient boosting models in streaming data environments. In practice, it provides a clear path for model choice in payment and high-frequency financial scenarios: LightGBM as the first choice, CatBoost for handling high-dimensional categorical features, and XGBoost when higher recall is needed, so that a balance between accuracy and real-time performance can be achieved under different business needs.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4. Conclusions","content":"\u003cp\u003eThis study compared CatBoost, XGBoost, and LightGBM in real-time fraud detection tasks under simulated high-frequency trading and resource-constrained environments. The results showed that (1) in terms of accuracy, the AUC of all three models exceeded 0.94, with LightGBM having the highest average AUC (0.95) and an F1 score of about 0.80, slightly higher than XGBoost and CatBoost, while CatBoost had narrower fluctuation ranges in AUC and F1, showing an advantage in stability; (2) in terms of latency control, LightGBM performed best in both single inference latency and tail latency, with P95 and P99 latency about 25\u0026ndash;30% lower than XGBoost, CatBoost was the next, and XGBoost had longer tails, which may increase risk under high-concurrency conditions; (3) in terms of resource efficiency, LightGBM used less CPU and memory, with average memory consumption about 20% lower than XGBoost, making it suitable for large-scale deployment, while CatBoost was stable in scenarios with complex categorical features but consumed more resources. Overall, with accuracy kept at a high level, LightGBM showed clear advantages in low latency and resource efficiency, CatBoost showed stability in high-dimensional categorical features, and XGBoost still had value in recall and maturity. By including accuracy, latency, and resource efficiency in one evaluation framework, this study provides empirical evidence and methodological reference for building deployable real-time fraud detection systems.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eLi J, Wu S, Wang N (2025) A CLIP-Based Uncertainty Modal Modeling (UMM) Framework for Pedestrian Re-Identification in Autonomous Driving\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTian J, Lu J, Wang M, Li H, Xu H (2025) Predicting Property Tax Classifications: An Empirical Study Using Multiple Machine Learning Algorithms on US State-Level Data\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang Y, Han X, Zhang X (2025) AI-Driven Market Segmentation and Multi-Behavioral Sequential Recommendation for Personalized E-Commerce Marketing\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYuan T, Zhang X, Chen X (2025) Machine Learning based Enterprise Financial Audit Framework and High Risk Identification. arXiv preprint arXiv:2507.06266\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang Z, Ding J, Jiang L, Dai D, Xia G (2024) Freepoint: Unsupervised point cloud instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 28254\u0026ndash;28263)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLin Y, Yao Y, Zhu J, He C (2025), March Application of Generative AI in Predictive Analysis of Urban Energy Distribution and Traffic Congestion in Smart Cities. In 2025 IEEE International Conference on Electronics, Energy Systems and Power Engineering (EESPE) (pp. 765\u0026ndash;768). IEEE\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi Z, Dey K, Chowdhury M, Bhavsar P (2016) Connectivity supported dynamic routing of electric vehicles in an inductively coupled power transfer environment. IET Intel Transport Syst 10(5):370\u0026ndash;377\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang X, Wang ZL, Yang W, Liu R (2024) Hydraulic properties within the complete moisture range of hydric soil on the Tibetan Plateau. Water Resour Res, 60(7), e2023WR036018.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang Y, Wen Y, Wu X, Cai H (2024) Comprehensive Evaluation of GLP1 Receptor Agonists in Modulating Inflammatory Pathways and Gut Microbiota\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang J (2025) Neural Network-based Prediction of Global Climate Change on Infectious Disease Transmission Patterns. Int J High Speed Electron Syst, 2540584\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang F, Paffenroth RC, Worth D (2024) Non-Linear Matrix Completion. J Data Anal Inform Process 12(1):115\u0026ndash;137\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXu J, Wang H, Trimbach H (2016) June). An OWL ontology representation for machine-learned functions using linked data. 2016 IEEE International Congress on Big Data (BigData Congress). IEEE, pp 319\u0026ndash;322\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSun X, Meng K, Wang W, Wang Q (2025), March Drone Assisted Freight Transport in Highway Logistics Coordinated Scheduling and Route Planning. In 2025 4th International Symposium on Computer Applications and Information Technology (ISCAIT) (pp. 1254\u0026ndash;1257). IEEE\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen F, Liang H, Li S, Yue L, Xu P (2025) Design of Domestic Chip Scheduling Architecture for Smart Grid Based on Edge Collaboration\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen H, Ma X, Mao Y, Ning P (2025) Research on Low Latency Algorithm Optimization and System Stability Enhancement for Intelligent Voice Assistant. Available at SSRN 5321721\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhong J, Fang X, Yang Z, Tian Z, Li C (2025) Skybound Magic: Enabling Body-Only Drone Piloting Through a Lightweight Vision\u0026ndash;Pose Interaction Framework. Int J Human\u0026ndash;Computer Interact, 1\u0026ndash;31\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang M, Wang Y, Shi J, Tong L (2025) Reinforcement Learning Based Multi-Stage Ad Sorting and Personalized Recommendation System Design\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePeng H, Ge L, Zheng X, Wang Y (2025) Design of Federated Recommendation Model and Data Privacy Protection Algorithm Based on Graph Convolutional Networks\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePeng H, Jin X, Huang Q, Liu S (2025) A Study on Enhancing the Reasoning Efficiency of Generative Recommender Systems Using Deep Model Compression. Available at SSRN 5321642\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePeng H, Jin X, Huang Q, Liu S (2025) E-commerce Intelligent Recommendation Optimization and Personalized Marketing Strategy Based on Big Model\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZheng J, Makar M (2022) Causally motivated multi-shortcut identification and removal. Adv Neural Inf Process Syst 35:12800\u0026ndash;12812\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXu K, Xu X, Wu H, Sun R (2024) Venturi Aeration Systems Design and Performance Evaluation in High Density Aquaculture\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYao Y (2022) A review of the comprehensive application of big data, artificial intelligence, and internet of things technologies in smart cities. J Comput methods Eng Appl, 1\u0026ndash;10\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang X, Liu R, K\u0026ouml;hli M, Marach J, Wang Z (2025) Monitoring soil water content and measurement depth of cosmic-ray neutron sensing in the Tibetan Plateau. J Hydrometeorol 26(2):155\u0026ndash;167\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen F, Yue L, Xu P, Liang H, Li S (2025) Research on the Efficiency Improvement Algorithm of Electric Vehicle. Energy Recovery System Based on GaN Power Module\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen H, Li J, Ma X, Mao Y (2025) Real-Time Response Optimization in Speech Interaction: A Mixed-Signal Processing Solution Incorporating C\u0026thinsp;+\u0026thinsp;+\u0026thinsp;and DSPs. Available at SSRN 5343716\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang Y, Shen M, Wang L, Wen Y, Cai H (2024) Comparative Modulation of Immune Responses and Inflammation by n-6 and n-3 Polyunsaturated Fatty Acids in Oxylipin-Mediated Pathways\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang J (2024) Comparative Analysis of the Impact of Advanced Information Technologies on the International Real Estate Market. Trans Econ Bus Manage Res 7:102\u0026ndash;108\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi Z, Chowdhury M, Bhavsar P (2024) Electric Vehicle Charging Infrastructure Optimization Incorporating Demand Forecasting and Renewable Energy Application. World J Innov Mod Technol, 7(6)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"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":"real-time detection, fraud detection, CatBoost, system response, machine learning models, model deployment","lastPublishedDoi":"10.21203/rs.3.rs-7539803/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7539803/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eReal-time performance is a key indicator in the design of fraud detection systems. This study focuses on three gradient boosting algorithms (CatBoost, XGBoost and LightGBM) and evaluates their performance in real-time fraud detection scenarios in terms of prediction accuracy, inference latency, and resource consumption. The experiments were conducted on a simulated high-frequency trading environment with a data stream platform. The results showed that LightGBM was the most advantageous in latency control, while CatBoost provided more stable responses while maintaining accuracy. This study offers a reference for building efficient and deployable online fraud detection systems.\u003c/p\u003e","manuscriptTitle":"Comparative Study of Efficient Machine Learning Models for Real-Time Fraud Detection: CatBoost, XGBoost and LightGBM","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-08 14:01:48","doi":"10.21203/rs.3.rs-7539803/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":"447cea4c-6789-4d5a-a9d3-57cae3ee1701","owner":[],"postedDate":"September 8th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":54225408,"name":"Environmental Engineering"},{"id":54225409,"name":"Marine and Freshwater Ecology"},{"id":54225410,"name":"Hydrology"}],"tags":[],"updatedAt":"2025-09-08T14:01:48+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-08 14:01:48","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7539803","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7539803","identity":"rs-7539803","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}