Interpretable Tsetlin Machine Rules for Market Classification

Summary

The article explains how a Tsetlin Machine classifies market conditions using readable logical rules. Small integer-state automata decide whether each condition or its negation belongs in an AND clause. Groups of clauses vote for or against each class, and the class with the highest net vote is selected. Training adjusts the automata through reward and penalty steps rather than gradient-based weight updates.

It describes turning technical indicators into Boolean inputs, labeling historical observations by forward return, training the classifier, inspecting its learned rules, and viewing active rules on a chart. The article says the implementation is checked against a separate reference and tested on known Boolean problems, including XOR. In its EURUSD hourly example, the model did not show predictive advantage over a five-bar horizon using the chosen features. The author presents interpretability and runtime visibility as practical benefits, while making clear that readable rules do not guarantee profitable predictions. Results depend on feature choices, labels, and market data; readers are encouraged to evaluate the model against an appropriate baseline and their own trading horizon.

Key ideas

• The model represents learned decisions as clauses made from included Boolean conditions and their negations.
• Automata use bounded integer states, with reward strengthening a choice and penalty moving it toward a change.
• Clauses vote for or against each class, and prediction selects the class with the strongest net vote.
• Market inputs are formed by converting indicators into Boolean features and labeling observations by forward returns.
• The reported EURUSD hourly experiment found no predictive edge over its stated five-bar horizon and feature set.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.