N-BEATS Perceiver: A Novel Approach for Robust Cryptocurrency Portfolio Forecasting

Abstract Cryptocurrencies are well-known for their high volatility and unpredictability, posing a challenge for forecasting using traditional methods. To address this issue, we explore variations of the N-BEATS deep learning (DL) architecture by adding convolutional network layers, Transformer mechanisms, and the Mish activation function, and propose a novel approach for forecasting cryptocurrency portfolios. Our comprehensive evaluation demonstrates that our model variations outperform other DL and traditional forecasting methods in numerous evaluation metrics, making them powerful tools for predicting cryptocurrency prices and portfolios in the rapidly-evolving cryptocurrency market. Furthermore, our newly proposed N-BEATS Perceiver model, a Transformer-based N-BEATS variation, exhibits a robust risk profile with less downside compared to other models and performs exceptionally well when evaluated using the TOPSIS method across a wide range of portfolio evaluation parameters. These results underscore the potential of our approach and specifically highlight the N-BEATS Perceiver's potential for selecting portfolios and forecasting cryptocurrency prices, offering valuable insights into the development of more accurate and reliable models for cryptocurrency forecasting.