LSTM-Based Stock Index Forecasting : Prediction Accuracy, Trading Performance, and Interpretability
Year: 2026
Type: dissertation
Abstract: This study examines the effectiveness of Long Short-Term Memory (LSTM) networks in forecasting stock index prices and evaluates whether such forecasts can be translated into an economically meaningful trading strategy. Using daily data from the OMXH25 index from 2010–2024 and a set of macroeconomic variables, a LSTM model is trained to predict next-day closing prices. The performance of the LSTM model is compared against a traditional autoregressive integrated moving average (ARIMA) benchmark, both in terms of statistical forecasting accuracy and economic performance in a rule-based trading framework. The empirical results indicate that the LSTM model does not outperform the ARIMA benchmark in forecasting accuracy. Furthermore, while the LSTM-based trading strategy generates higher returns than a buy-and-hold benchmark in a transaction cost free setting, the advantage disappears once transaction costs are introduced. Under such conditions, the strategy fails to produce economically meaningful excess returns in comparison to the buy-and-hold benchmark. The findings suggest that applying machine learning models into financial modelling does not necessarily translate into profitable trading opportunities, particularly when transaction costs are considered. To enhance interpretability, SHapley Additive exPlanations (SHAP) are applied to the trained LSTM model. The analysis reveals that the model relies primarily on recent lagged values of the index, while the additional macroeconomic variables provide limited incremental predictive value. This indicates that the model predominantly captures short-term autoregressive patterns rather than complex relationships across explanatory variables. Overall, the results suggest that, in the context of the OMXH25 index, the increased complexity of LSTM models does not yield superior economic performance compared to simpler benchmark models. The study highlights the importance of evaluating forecasting models within realistic trading frameworks and contributes to the literature by combining forecasting performance, trading evaluation, and model interpretability within a unified empirical setting.
Source: UTUPub (University of Turku)
Author: Salonen, Anton
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7220827696