BITCOIN RETURN FORECASTING USING INTRINSIC MODE FUNCTION AND DEEP NEURAL NETWORK WITH CONFIDENCE-BASED TRADING SIGNAL

Abstract

Bitcoin is a highly volatile asset that poses considerable investment risk. This study proposes a trading strategy that integrates signal decomposition with deep learning to enhance forecasting accuracy, reduce the likelihood of losses, and minimize trades triggered by inaccurate predictions. Daily closing prices of Bitcoin are decomposed using Empirical Mode Decomposition (EMD) to extract Intrinsic Mode Functions (IMFs). IMFs with significant temporal correlation to returns are selected through cross-correlation analysis and used as input features for a stacked deep neural network comprising Bi-LSTM, LSTM, and Dense layers to forecast daily returns. The forecast outputs are transformed into Z-scores and p-values to evaluate confidence levels, which are then used to generate trading signals based on thresholds ranging from 0% to 90%. The model achieves an R² of 0.55 and directional accuracy of 80.32%. Without applying confidence thresholds, the strategy yields an average daily return of 1.50%, clearly outperforming the buy-and-hold approach (0.163%). Even when using a high confidence threshold of 90%, which reduces the number of trades and average returns, the strategy still delivers better returns than the buy-and-hold strategy at all confidence levels.