Don't pay for caution you can't justify

Caution sells itself. “Don’t trade in the last 24h before FOMC” sounds prudent. “Don’t carry overnight risk through earnings” sounds prudent. “Don’t fire signals during the first 30 minutes of the session” sounds prudent. And every one of them, shipped without measurement, is a tax on the strategy whose only justification is that the trader thought it was prudent.

This is a note about which gates survive measurement, and which ones get smuggled in past it.

The two bars

A risk-reduction gate ships when it clears both of these:

Bar 1. The signal has no event-specific edge to defend. A cross-sectional momentum signal calibrated against panel-clock returns doesn’t “know” what an FOMC announcement is. Its lookback window absorbs FOMC-day shocks the same way it absorbs every other day’s shocks. There is no per-event positioning logic that the blackout is protecting; the signal would happily over-position into the event and the post-event mean-reversion would (on average) cost some of that back. The gate isn’t defending alpha — it’s reducing exposure.

That’s not automatically wrong. Exposure-reduction has a place. But it has to clear the second bar.

Bar 2. The cost is finite, measurable, and tractable on synthetic data first. “Cost” here means the Sharpe / drawdown / turnover delta of the gated strategy vs the ungated one, on the same price stream, over the same window. Not the cost in the real world. Not the cost over a decade of real FOMC days. The synthetic stand-in, with a calibrated event-day vol multiplier and a single panel signal, gives you a direction and an order of magnitude. If the synthetic A/B says the gate costs you 0.4 Sharpe, real data will probably cost more (selection bias on the event calendar is a real thing). If it says 0.05, real data might still cost more — but you’re in the noise band and the decision is defensible either way.

A gate that increases Sharpe in the synthetic A/B is also worth checking against bar 1: it’s not actually reducing risk, it’s adding edge. Sometimes that’s because the simulation has a structural quirk the gate exploits (turnover discount on event days), and sometimes it’s because you’ve discovered an actual event-specific signal. Either way, treat it as a different conversation than the original “caution” framing.

What the FOMC blackout looks like through this lens

The platform shipped the event-clock surface recently, and the first concrete strategy on it is a momentum signal wrapped in a 24h pre-FOMC blackout (XsMomentumWithFomcBlackoutStrategy). The wrapper exists for two reasons that have nothing to do with alpha:

It’s the smallest concrete consumer of the event-clock surface. Without a strategy that cites events, the cooldown rule and the citation graph have nothing to act on.
It demonstrates the gate shape so the next operator who wants to add a “no positions during X event” rule has a worked example.

But the wrapper does not claim to add alpha. The A/B harness (examples/fomc_blackout_compare.py) confirms: on a 200-day synthetic walk with a 3× FOMC vol multiplier, blackout-on costs ~0.1 Sharpe vs blackout-off. Different seeds move that around in the ±0.2 range. That number says one thing only — shipping this gate in production requires either (a) accepting the Sharpe cost as the price of an operational guarantee you can name, or (b) finding the event-specific edge that fills the gap.

The first option is fine. “I don’t want positions through FOMC because if my data feed lags during the announcement I have no recourse” is a real operational concern, the cost is bounded, and a 0.1 Sharpe insurance premium is cheap. Document the rationale; ship the gate.

The second option is the research question. “Is there a sentiment shift in the 4-hour window after FOMC that justifies adding event exposure rather than removing it?” is a real bar-1-crosser: the panel signal has no language model behind it, no event-specific edge to defend, and an LLM-scored event reduction would be a different signal class. If the answer is yes, the blackout was the wrong shape — what you wanted was a switch to a different signal during the event window, not a reduction of the panel signal’s exposure. Either way the A/B is the artifact that tells you which one.

Anti-patterns

The shapes that look like caution but fail one or both bars:

“Add the gate because it makes the operator feel better.” Bar 2 fails: there’s no measurement, just a vibe. The vibe gets fed by every quarter where the gate didn’t fire and the strategy did fine — survivor bias on the operator’s confidence, not on the strategy’s PnL.
“Add the gate because the literature says event days are volatile.” Bar 1 partial-credit (sometimes): event-day vol is documented. But the signal you’re gating wasn’t calibrated to exploit event-day volatility either way. The literature is describing a market phenomenon; the gate would be paying for it twice if the signal didn’t already see it.
“Add the gate as a default for every strategy.” Both bars fail: the gate runs against signals it was never measured against, including ones that genuinely want exposure during the event (e.g., a vol-strategy short straddle calibrated to capture the post-event vol crush). A default-on gate is a default-on tax.
“Add the gate because we got hurt once.” This one’s the hardest. A single bad event-day fill is sample size 1; the gate pays its cost every event day going forward. Bar 2 says: do the A/B over a multi-year window before letting one bad day calibrate the policy. If the gate would have prevented the loss but collectively costs more PnL than the loss it would have prevented, the right answer is better order-handling on event days, not no orders.

What the discipline buys

The same shape from walk-forward without leakage — measure first, ship second — but at the strategy-construction layer rather than the feature-evaluation layer. The leakage checklist catches features that fake alpha; the two-bar gate test catches risk-reduction logic that fakes safety.

Both fail the same way: the strategy looks fine because nobody questioned the wrong piece.