Matrix Profile for Interpretable SVM-based Financial Time Series Classification

Abstract

Financial time series prediction remains challenging, and many existing approaches offer limited interpretability. This thesis investigates whether Matrix Profile-based historical pattern similarity can provide interpretable features with detectable predictive signals for financial time series trend classification. Using daily closing values of the S&P 500 index from 1990 to 2025, this thesis derives features from Matrix Profile similarity search, evaluates them with a linear Support Vector Machine, and uses a Random Forest classifier for comparison. The results show that certain Matrix Profile-derived features exhibit relatively consistent directional behavior in the classification decision. However, these features do not consistently outperform baseline statistical features, and overall classification performance remains modest. The main contribution of this thesis is to show that Matrix Profile similarity information can be transformed into interpretable features for financial trend classification, rather than serving as a high-accuracy prediction method.