BOSS Symbolic Time-Series Classification for Market Regimes

Summary

This document explains a market-regime indicator built on Bag-of-SFA-Symbols (BOSS). It converts normalized price windows into symbolic words using a low-pass Fourier representation and data-learned bins, then compares stretches of market history as word-frequency bags. The indicator labels trailing segments as ranging, trending, or volatile, and exposes regime codes for use as filters in automated strategies.

The implementation uses an ensemble across candidate window lengths, retaining classifiers with leave-one-out accuracy near the best member and combining their votes. The document reports that this raised macro accuracy on real bars from 25.3% for a single classifier to 62.5% for the ensemble. It also describes a coefficient sweep in which shorter symbolic words were more robust to noise. These results are specific to the presented setup; regime training labels come from rules based on return autocorrelation and volatility, so the classifier reproduces those definitions rather than identifying an objective or predictive market state. The document does not establish that using its labels improves trading returns.

Key ideas

• BOSS represents normalized price windows as symbolic words and summarizes segments with word-frequency histograms.
• Fourier filtering and learned coefficient bins reduce sensitivity to noise and price scale.
• An ensemble across window sizes addresses the sensitivity of a single classifier to its chosen window.
• The reported ensemble accuracy exceeds the reported single-classifier accuracy on the document's real-bar evaluation.
• Training regimes are rule-labeled using autocorrelation and volatility, which limits what the classifier's labels mean.

Tags

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