LSTM-Based Financial Forecasting with Interpretability:Evidence from US Equity Markets

This paper examines the predictive performance of Long Short-Term Memory (LSTM) neural networks in forecasting daily equity prices, using Apple Inc. (AAPL, 2010–2025) as the primary case study and Microsoft Corporation (MSFT) as a cross-asset generalization test through transfer learning. The proposed architecture consists of a three-layer stacked LSTM model optimized for financial time-series prediction. Empirical results demonstrate strong out-of-sample forecasting accuracy, with RMSE = $7.03, MAE = $5.50, R² = 0.9537, and MAPE = 3.03%, outperforming traditional ARIMA and Support Vector Regression (SVR) benchmarks across both statistical and financial performance measures. Additional evaluation through investment-oriented indicators yields a Sharpe ratio of 1.23 and a directional hit ratio of 0.68, confirming the economic relevance of the generated forecasts for trading and portfolio management applications.

Model interpretability is enhanced through SHAP (SHapley Additive exPlanations) analysis, which identifies the 10–15 most recent price lags as the dominant contributors to predictions, providing evidence consistent with short-term momentum dynamics in equity markets. To assess robustness under extreme market conditions, the model is stress-tested during the COVID-19 crisis period (March 2020), where forecasting accuracy deteriorates significantly, with MAPE increasing from 2.9% to 4.8% (p < 0.05), revealing sensitivity to structural breaks and heightened volatility regimes. A fully documented UML-based workflow and publicly available implementation code ensure transparency and reproducibility. The study contributes to the literature on deep learning-based financial forecasting by combining predictive performance, transfer learning validation, explainability analysis, and robustness testing, and offers practical guidance for integrating interpretable artificial intelligence techniques into quantitative investment decision-making.