Cryptocurrency Price Forecasting via Gate-Level Multi-Source Transfer Learning

Cryptocurrency price series are highly non-stationary, which hinders transfer across assets. This work presents a gate-level multi-source transfer scheme for Long Short-Term Memory (LSTM) networks that assembles the first layer of the target model by selecting, for each gate (input, forget, candidate, and output), the source that best meets our selection criteria within a set of related cryptocurrencies. The selection integrates (i) intrinsic activation quality—entropy, mean magnitude, and saturation—weighted with CRITIC (Criteria Importance Through Intercriteria Correlation), and (ii) domain compatibility estimated through Gramian Angular Fields (GAF) and Central Moment Discrepancy (CMD). The architecture is fixed using Differential Evolution based solely on the target asset; each source is then retrained with that same architecture to ensure parameter-by-parameter correspondence and enable gate-by-gate assembly. Demonstrating the effectiveness of the proposed methodology, one of the experiments achieved reductions of 94.96% in MSE, 71.95% in RMSE, 77.26% in MAE, and 95.81% in Log-CoshLoss, showing that gate-level transfer constitutes an interpretable and highly effective strategy for multi-asset financial forecasting.