Training Is Not the Finish Line
After an AI strategy goes live, the market keeps changing.
Rates, trading structure, participant behavior, regulation, and data quality can all change the relationship the model learned.
That is model drift.
A Simple Example
A model learns that a group of tech stocks tends to rise for 3 days after earnings beats.
Then rates rise, and the market stops rewarding growth narratives. It focuses more on cash flow.
The same earnings beat now produces a weaker or opposite reaction.
The model is not broken. The world it learned has changed.
Where Drift Shows Up
- Win rate slowly declines
- Average gain shrinks
- Drawdowns deepen
- Signals cluster in fewer assets
- Live fills differ from backtest assumptions
The danger is that it may not be obvious in one day. Three months later, the model may already be stale.
How to Monitor
| Metric | What to Watch |
|---|---|
| Win rate | Outside historical range |
| Payoff ratio | Winners and losers worsen |
| Drawdown | Above threshold |
| Signal distribution | Sudden concentration |
| Data quality | Missing, delayed, abnormal |
Stopping Is Not Failure
Pausing a strategy is risk control.
If performance leaves the expected range, reduce size, stop, and review. Do not simply hope it comes back.
Quiz
Q1. Model drift means:
A. Market relationships change and the model weakens B. Chart colors change
Q2. Why is drift dangerous?
A. It can happen slowly B. It always explodes in one day
Q3. If drift appears, first:
A. Reduce size or pause and review B. Double down
Answer Key
Q1: A Q2: A Q3: A
Further reading: Wikipedia — Concept Drift · NIST — AI Risk Management Framework
For education only. Model performance changes with market regimes.
Get the pre-trade checklist.
We are turning these guides into a searchable checklist for checking terms, rules and risk before you trade.