A HYBRID STATE-SPACE AND ATTENTION FRAMEWORK FOR CRYPTOCURRENCY TIME-SERIES FORECASTING USING CRYPTOMAMBA AND ATTENTIONMLP

Abstract

Cryptocurrency markets are unpredictable, highly non-stationary, and subject to intricate time-series patterns, which significantly complicate the forecasting task. Historically, different variants of RNN such as LSTM and BiLSTM were heavily utilized for time-series prediction, yet they often do not effectively capture long-range dependencies, while possessing high computational costs. This research suggests a comparable framework utilizing the state-space sequence modeling architecture CryptoMamba to tackle these problems. We benchmark the CryptoMamba's performance against classic RNN models and also propose a new version of CryptoMamba combined with Attention-based Multi-Layer Perceptron (AttentionMLP) to further boost the prediction accuracy. Experiments were performed over a dataset of cryptocurrency with identical features and evaluation criteria. Based on the evaluation, it can be observed that the model CryptoMamba + AttentionMLP provides the best prediction with lowest RMSE, MAE and MAPE, while outperforming LSTM, BiLSTM and GRU.