Demystifying Deep Learning in Finance : A SHAP-Driven Study of Technical Indicators

This thesis replicates and extends the work of Mostafavi and Hooman [1] by evaluating the predictive utility of technical indicators in financial time series forecasting, using daily price data from Swedish equities listed on the NASDAQ Stockholm exchange. Bidirectional Long Short-Term Memory (BiLSTM) models were trained on a per-stock basis, using compact feature sets selected via wrapper-based forward selection from a 97-feature dataset that included both engineered indicators and raw market variables. Unlike the original study, which used pooled U.S. stock data and principal component analysis (PCA) to derive global feature importance, this study focuses on per-stock modeling and interpretability. SHAP (SHapley Additive exPlanations), a model-agnostic explainability framework, was used to attribute predictions to specific input features at both global and local levels. Results show that a small number of indicators (typically 1–5) account for most of the model’s predictive capacity, and that SHAP enables consistent, interpretable insights into how predictions are formed. Compared to full-feature baselines, wrapper-based models achieved similar or better accuracy while improving transparency. This work, conducted in collaboration with Softwerk AB, demonstrates that deep learning in finance can be made more interpretable through post-hoc analysis, supporting more trustworthy and auditable forecasting pipelines.