Using Path Signatures to Classify Stocks by Country

Summary

This article applies rough path signatures to a supervised classification task: infer whether a company is based in the United States, United Kingdom, or Germany from its daily closing prices and trading volumes over a year. Stocks are represented as time-ordered price and volume data, while country labels are encoded as target points. The data is randomly split into training and test subsets, and signature-based models are evaluated at orders one through four.

The reported best result is 97% accuracy on the held-out test set at signature order four. This suggests that joint patterns in price and volume paths may contain information associated with country-level market characteristics. However, the account gives limited information about sample size, company selection, class balance, preprocessing, repeated splits, or robustness across periods. The result is a classification demonstration, not evidence of a trading strategy or proof that the learned relationships generalize to other markets or time periods.

Key ideas

• The model uses daily stock price and volume paths as inputs to country classification.
• Rough path signatures at several orders are compared in a supervised learning setup.
• The dataset is divided into training and test subsets before evaluating predictions.
• The article reports its best test accuracy at signature order four.
• Limited details on sampling and robustness constrain conclusions about generalization or trading value.

Tags

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