Own-Trade Impact Feedback
A backtest usually treats the price series as something that happens to you. If your own trading is big enough to move the tape, the price series partly happens because of you, and a strategy can end up "predicting" moves it caused itself.
Prerequisites: Market Impact
Almost every backtest treats historical prices as a fixed, exogenous record — something that already happened, independent of the strategy being tested. That assumption is exactly backwards once a strategy is big enough to move the tape it's trading on. If your own selling pushes a price down and it later recovers as your impact fades, a mean-reversion signal trained on that price series will "discover" a mean-reversion pattern that is, in truth, just the market healing from a wound the strategy itself inflicted.
Worked example: the reversal signal that's chasing its own footprint
A mid-cap mean-reversion strategy sells into weakness and buys the subsequent bounce. Backtested on historical prices with no impact model, it shows a strong edge: sell after a 2% intraday drop, the price recovers 1.1% over the next hour on average.
Now separate the sample by whether the fund's own trading was large relative to that name's volume at the time:
| Regime | Participation in that name | Recovery captured | Interpretation |
|---|---|---|---|
| Fund is a small fraction of volume | < 1% ADV | 0.3% | genuine market mean reversion |
| Fund is a meaningful fraction | 3-5% ADV | 1.1% | partly the fund's own impact decaying |
In the second bucket, roughly two-thirds of the "recovery" the backtest is crediting to the signal is arithmetic: the fund pushed the price down with its own selling, and price snapping back toward where it would have been anyway looks, in a backtest that assumes prices are exogenous, exactly like a profitable reversal call. Run the same signal on a simulator that removes the fund's own price impact from the reference series, and the edge in the high-participation bucket collapses toward the 0.3% seen in the low-participation bucket.
Once a strategy's own participation is large enough to move prices, the historical series used to test it is contaminated by the strategy's own footprint — a signal can look predictive while it's really just recognizing the decay of its own impact.
This feedback is invisible in a standard backtest because both the impact and the "prediction" live inside the same price series with no label distinguishing them. It only shows up when you compare edge across participation-rate buckets, or when you simulate with an impact model that separates the fund's footprint from the underlying market's.
The practical check is exactly the bucket comparison above: split historical trades by how large the fund's own participation was at the time, and treat any edge that only appears in the high-participation bucket as suspect until proven otherwise with an impact-adjusted price series.
Practice in interviews
Further reading
- Kyle & Obizhaeva, Market Microstructure Invariance