[1K4-GS-3b] Knowledge utilization and sharing

4:10 PM - 4:25 PM JST(7:10 AM - 7:25 AM UTC)

[1K4-GS-3b-03] Time-Series ItemSB: Discovery of Interpretable Time-Series Rules in Financial Markets

〇Yasuhiro Suzuki¹, Kaoru Shimada¹ (1. Gunma University)

Keywords:

time-series data mining,genetic algorithm,knowledge discovery

Ensuring interpretability in financial market analysis is crucial for transparent investment decision-making. However, conventional inter-transaction association rules are limited to discrete classes and cannot predict continuous value distributions. This paper proposes "Time-Series ItemSB," a method extending the ItemSB concept to time-series data using GNMiner (GNP-based Rule Mining Method) with time-delay attributes. This approach enables the discovery of rules predicting future continuous distributions based on past multi-point conditions. We introduce a quadrant concentration rate to quantitatively evaluate rule quality. In experiments using 15 years of daily data for 20 major currency pairs, 31,526 rules were extracted. While the overall out-of-sample accuracy was impacted by market changes, specific pairs like GBPJPY achieved a quadrant match rate of 31.4%. Notably, the top 30 selected high-precision rules achieved an average match rate of 49.2%, validating the proposed method's effectiveness in discovering actionable financial patterns.