Traditional Long Short-Term Memory (LSTM) networks often struggle to effectively capture long-range dependencies and achieve optimal performance when analyzing complex financial time series. Addressing these challenges, this study proposes an improved financial forecasting model based on the recently developed extended Long Short-Term Memory 1 (xLSTM) architecture. Our method adapts the xLSTMTime model 2, utilizing its enhanced memory mechanisms—specifically sLSTM (stabilized LSTM) and mLSTM (matrix LSTM)—alongside exponential gating to better handle volatile financial data. To significantly enhance predictive power, we integrate 18 technical indicators 3 derived from historical price data into the model, capturing essential aspects of price trends and momentum. We fine-tune the model through hyperparameter optimization and apply enhancements like series decomposition and Reversible Instance Normalization (RevIN) to ensure robustness against distribution shifts and non-stationarity. Extensive experiments conducted on historical S&P 500 index data sourced from Yahoo Finance demonstrate the efficacy of our approach. The enhanced xLSTMTime model consistently outperforms baseline models and configurations run without the technical indicators, achieving a lower Mean Absolute Percentage Error (MAPE) in both short-term and long-term forecasts. For instance, in one short-horizon setup, MAPE improved from 0.2492 to 0.2226 with the full set of features. Ablation studies confirm the crucial additive value of these technical features in boosting predictive accuracy. This work validates the potential of refined recurrent architectures in financial forecasting, offering a robust and computationally efficient alternative to transformer-based models for handling multivariate time series with long horizons.
Duy et al. (Wed,) studied this question.