Interpretable AI for finance: LSTM-TSK fuzzy networks in stock prediction

Abstract

With the growing adoption of artificial intelligence in financial markets, deep learning models have become increasingly prevalent in predictive analytics and decision-making. Deep neural networks excel at capturing complex, nonlinear patterns in data, making them powerful tools for stock price forecasting and portfolio optimization. However, their black-box nature limits interpretability, which is a critical concern in high-stakes financial applications where transparency is essential for trust and decision-making. This project introduces an interpretable neuro-fuzzy system, the Interpretable LSTM-TSK Fuzzy Deep Neural Network (ILTFDNN), designed to enhance both predictive accuracy and interpretability in financial modeling. Unlike conventional deep learning models, ILTFDNN incorporates fuzzy rule mappings based on the TSK inference system, enabling its decision-making process to be aligned with human reasoning. The model dynamically learns fuzzy membership functions and rule parameters, making it adaptable to the evolving nature of financial data. Beyond improving transparency, this work explores the integration of ILTFDNN with momentum-based trading strategies, specifically through the forecasted MACD (fMACD) indicator. By using ILTFDNN predictions to compute MACD values ahead of time, the lag inherent in traditional technical indicators can be mitigated, potentially enhancing trading performance. Additionally, the project investigates the "infinite sliding plane" approach for function approximation, demonstrating how the model effectively captures nonlinear trends over time. The findings highlight the potential of interpretable fuzzy deep learning systems to bridge the gap between performance and explainability in algorithmic finance.