Discretionary Overlays On A Systematic Book
When and how a human trader is allowed to override or adjust a systematic strategy's output — and why an overlay policy needs to be as explicit as the strategy itself, or it quietly becomes the whole strategy.
Prerequisites: Systematic vs Discretionary Trading
A systematic strategy generates its signals and sizing from a fixed, backtested process, and its whole value proposition is consistency — it does the same thing in the same situation every time, without a human's mood or memory of yesterday's loss creeping in. A discretionary overlay is a human's ability to step in on top of that and adjust, delay, or veto what the system wants to do. Overlays exist for real reasons: a system trained on historical data has no way to know about a specific piece of breaking news, an unusual market structure event, or a data feed error that's about to feed it a bad input. But an overlay that isn't tightly scoped in advance tends to expand until the "systematic" strategy is really being run at a human's discretion most of the time, which quietly throws away the reason the strategy was built to be systematic in the first place.
The fix is treating the overlay itself as a rule, not a general license. A well-run desk writes down, in advance, the specific and narrow situations where a human is allowed to intervene — a known data outage, a trading halt, a regulatory event that the model has no way to anticipate — and requires anything outside that list to go through an explicit, logged decision, ideally reviewed after the fact, rather than a quiet keystroke that overrides the system's order.
Worked example
A systematic equity strategy generates a large buy order in a stock right as the company announces an unscheduled halt pending material news. The strategy has no concept of a trading halt; left alone it would keep trying to send an order that the exchange will reject, and once trading resumes it might chase a gapped price with no awareness that anything unusual happened. This is exactly the kind of situation an overlay policy should explicitly permit: a human pauses the order, waits for the halt to resolve and the news to be digested, and only then decides whether to let the system re-engage or skip the trade for the day. That's a scoped, defensible override.
Contrast that with a trader who, on a normal day with no halt, no news, and no data issue, decides the system's signal "doesn't feel right" and skips the trade anyway. That's not a scoped exception — it's discretion creeping into every trade, undermining the entire premise that the strategy's historical track record predicts its future behavior, since the live version is no longer actually the backtested system.
An overlay policy should list the specific situations that justify human intervention — not leave it to case-by-case judgment — and every override taken should be logged and reviewed, so the desk can tell, after the fact, whether overlays are improving results or quietly turning a systematic strategy into a discretionary one.
An overlay that's used often enough stops being an overlay and becomes the real strategy — at that point the backtest and track record built for the systematic version no longer describe what's actually being traded, and the strategy needs to be re-evaluated as the discretionary process it has become.
Practice in interviews
Further reading
- Chan, Quantitative Trading, ch. on discretionary risk overrides