Abstract
While Long Short-Term Memory (LSTM) networks are pivotal for time-series forecasting, their fixed recurrent structure struggles to capture the complex dynamics of financial markets, such as abrupt volatility shifts and persistent trends. This study proposes a physics-inspired recurrent architecture, named Differential Long Short-Term Memory (D-LSTM), that embeds diffusion and dissipation operators directly into the cell-state evolution. These operators jointly capture how information propagates across time, how volatility dissipates, and how memory decays in dynamic financial environments. The model integrates the interpretability of differential equations with the predictive power of deep learning, forming a mathematically grounded recurrent structure. To ensure rigorous validation, D-LSTM is benchmarked against standard LSTM and self-attention LSTM (SA-LSTM) models across three heterogeneous financial datasets: Crude Oil (WTI), Gold, and NASDAQ Composite Index, covering the period 2014–2024. The results demonstrate that D-LSTM achieves a statistically significant improvement in predictive accuracy across all assets, as measured by MAE, RMSE, and R2 scores. Critically, our extensive ablation studies reveal that the model’s components have a data-dependent efficacy. The full architecture excels on the volatile oil series, whereas more parsimonious configurations are superior for the strongly trending gold and NASDAQ data. This research establishes D-LSTM not as a monolithic model, but as a flexible and robust framework for financial forecasting. Our findings underscore the critical importance of aligning architectural complexity with the statistical properties of the time series, offering a new paradigm for designing more adaptive and effective predictive models.

This publication has 34 references indexed in Scilit: