SETAR-Based Autoregressive Drift Detection for Trading Models

Summary

This document explains ADDM, a method for detecting changes in a trading model’s prediction errors and adapting the model when market conditions shift. Its detector uses a Self-Exciting Threshold Autoregressive (SETAR) model to divide error behavior into regimes. A lagged error value is compared with a threshold; when a change is detected, recent labeled data trains a replacement model, which is blended with the original according to an estimated drift severity.

The article outlines a workflow that starts with a baseline model and validation errors, then monitors incoming data and updates the model after detected drift. It proposes using the third quartile of errors to estimate severity and mentions a comparison of daily retraining with and without drift detection, alongside buy-and-hold. However, the supplied text omits much of the implementation and presents no backtest results, so it does not establish that ADDM outperforms alternatives. Its claims about broad model compatibility should be treated as a proposal requiring validation on the target data and strategy.

Key ideas

• Concept drift can make historical relationships less useful for a predictive trading model.
• ADDM monitors prediction errors and uses SETAR regimes to detect changes in their behavior.
• The detector compares a lagged error with a threshold rather than relying on the latest error alone.
• After a detected shift, recent labeled observations train a new model that is blended with the original according to estimated drift severity.
• The article describes a backtest comparison but supplies no reported performance evidence in the provided text.

Tags

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