This project explores short-term pattern searching in financial time series. Inspired by Gupta et al. (2025), we extract patterns using a geometrical criterion: patterns that precede one-sided market moves.
To evaluate the relevance of a pattern, we compute an information criterion based on Shannon entropy, which captures how strongly the pattern supports a directional view of the market.
By combining:
Geometric meaning (patterns that historically generated profits) and
Information-theoretic strength (confidence in directional bias),
we aim to isolate more “pure” patterns. This is particularly useful in noisy, high-frequency (1 min bar) data, where raw pattern frequency can be misleading.