Interpretable Machine Learning for Financial Time Series: A Group-Based Analysis of Technical Indicators
Year: 2026
Type: conference-paper
Abstract: Financial time series prediction remains a challenging problem due to the non-stationary and noisy nature of market data. While recent studies often focus on maximizing predictive performance using complex models, there is a growing need for methodologies that emphasize interpretability, temporal robustness, and transparent evaluation. This paper proposes a structured experimental framework for cross-asset stock movement prediction using normalized technical indicators. The study employs an expanding-window evaluation methodology to respect temporal ordering and compares controlled balanced-test evaluation with an additional robustness setting in which test data are kept in their natural class distribution. A set of interpretable machine learning models, including Logistic Regression, Decision Trees, and Gaussian Naive Bayes, is used to analyze predictive performance under these conditions. In addition to full-feature models, the study introduces a group-wise analysis of technical indicators, separating features into functional categories such as momentum, volatility, volume, trend, price comparison, and statistical bands. This enables an examination of how different indicator groups contribute to model behavior across multiple market regimes between 2020 and 2025. The results show that predictive performance remains very close to the random baseline, reflecting the inherent difficulty of short-horizon financial forecasting. Small and temporally varying deviations from the baseline are observed, but these should be interpreted cautiously as weak statistical regularities rather than as evidence of robust or economically exploitable predictability. The interpretability analysis suggests that price-comparison and statistical-band indicators are used most consistently by the models.
Source: Procedia Computer Science
Authors: Lívia Kelebercová, Maroš Čavojský
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: diamond
OpenAlex ID: W7225083505