I keep seeing hedge engines that look careful in a simulator and then reveal they were partly just betting the drift.
They named worlds. They turned costs on. They budgeted path failures. Charts look like risk management. Then you ask which probability world the policy was trained under, and the room goes quiet. That silence matters. The measure you train under decides whether the policy is hedging or hunting drift.
Black-Scholes is the famous clean version of that story: in a complete market you can change measure, get a unique price, and get a unique replicating hedge (Harrison-Pliska completeness). That uniqueness is the textbook promise, not a universal property of every desk book. Live structured books around something like CEH-001 are usually incomplete. Frictions exist. Not every risk is tradable. Uniqueness dies. An AI that learns to hedge under the physical world can smuggle statistical arbitrage into a process labeled risk management.
This is the ruler problem wearing probability clothes. Anonymous measure choice is anonymous geometry on paths.
Two worlds, two jobs
Finance routinely uses two probability worlds on the same outcomes. The physical measure P weights paths the way history or a research simulator does: drifts, risk premia, fat crises. The pricing or hedge measure Q is a world in which discounted tradable gains behave fairly after costs, the classical risk-neutral setting where you price by expectation of discounted payoff, not by forecasting P-mean returns.
Girsanov-style reweighting is the continuous-time name for building Q from P: you do not invent new noise from nowhere; you reweight path probabilities so the same driving noise looks like Brownian motion with a different drift. (Brownian motion is the continuous-time random walk underneath these models: a path whose increments are independent, mean-zero, and grow in variance proportionally to elapsed time, the same object a martingale describes discretely, a process with no built-in drift, so that today's value is the best forecast of tomorrow's.) The market price of risk is the drift you remove when moving between "how stocks tend to grow under P" and "how discounted stocks behave under Q."
Quick desk test. A ticket that reports a hedge Sharpe under P while claiming "risk-neutral validation" has mixed geometries. Silence about which world produced the number is anonymous geometry. SILENT_MEASURE is the refuse.
Incomplete markets: disagreement returns inside pricing
Completeness is what purchases uniqueness of the equivalent martingale measure and uniqueness of price. On incomplete books, many measures are consistent with no-arbitrage. Honest output is a band, not a stolen point. Pasting a binomial or Black-Scholes float into a live incomplete gap is completeness theater: POINT_ON_INCOMPLETE. Pasting Black-Scholes as equity alpha is DENY_BS_AS_ALPHA, a famous closed form used as a spend passport.
CEH-001's dual clouds near 1.50 [1.42, 1.58] and 2.10 [2.01, 2.22] are incomplete honesty in mark space. Averaging them into 1.80 is midpoint laundering. A P-trained hedge that makes 1.80 look "hedgeable" in a friendly sim is laundering a peace number that was never honest under Q.
How this shows up in AI hedge stacks
- Model trench. Deep hedging (Buehler et al., arXiv:1802.03042) trained on raw simulator PnL under physical drift farms edge. Remove-the-drift stages exist in that literature precisely because the optimizer will harvest premia and call it skill. Residual honesty is not "the hedge made money." Money under P can be drift. Honesty asks: after remove-drift, is leftover edge still predictable by tradables?
- Agent trench. Agents that report a single float as the hedge value on an incomplete book are performing uniqueness theater. Agents that cite a P-Sharpe as hedge quality are citing the wrong world.
- Harness trench. Promotion packs that only score P-PnL will crown drift overlays. Residual-honesty tests with named thresholds must bind to the ticket. A correlation of residual with lagged tradable return near 1.0 against a tolerance of 0.25 is not a vibe disagreement. It is un-neutralised exposure with a number.
- Pipeline trench. Costs charged under P then reweighted can flatter policies when costs bite hardest in states Q upweights. Density estimates that do not integrate to one bias every Q-expectation smoothly. Both are production bugs that look like skill.
Walk the failure without pretending it is exotic
One-period toy, numbers small enough to check by hand. Stock 100, up 110, down 90, rate 5%. Risk-neutral up-weight 0.75. Physical simulator uses up-weight 0.80. A "hedge" that is mostly long the share funded by borrowing shows positive expected PnL under P and zero under Q before costs, negative after costs. Same policy, same payoffs. Only the measure and the cost changed. Verdict: DRIFT_OVERLAY.
| Object | Value | Reading |
|---|---|---|
| P up-weight / Q up-weight | 0.80 / 0.75 | physical vs pricing geometry |
| E^P[Π] long-share-funded-by-borrow | +1.0 | looks like hedge skill |
| Annualised P-ratio (toy) | ~+2.0 | career-making under P |
| E^Q[Π] before costs | 0 | no edge under pricing measure |
| After cost 0.40, Q-mean | negative | honest Q-after-cost ratio dies |
| corr(ε, R_lag) residual fixture | ~0.996 vs tol 0.25 | un-neutralised exposure |
| CEH-001 dual marks | ~1.50 [1.42, 1.58] / ~2.10 [2.01, 2.22] | incomplete honesty |
| P-hedge blessing of 1.80 | deny | midpoint laundering, not Q-honest |
| Verdict | DRIFT_OVERLAY (hedge ≠ alpha) | |
I have sat in reviews where hedge quality meant simulator PnL under physical drift, residual honesty meant the same chart with a smoother line, and uniqueness meant the optimizer converged. None of those sentences is an economic claim under a declared Q.
Frame conflict versus measure conflict
When CEH-001 shows 1.50 versus 2.10, ask two questions, not one. Is this frame conflict (spread cloud versus curve cloud, both legitimate, human chooses a ruler)? Is this measure conflict (a P-trained overlay sold as a Q-honest hedge, arithmetic has a right answer)? Either way, a midpoint blessed by a P-trained hedge is not an answer. It is a way to hide which question you refused to ask.
What to lock before you call it a hedge
- Name P, name Q, name the change-of-measure recipe (Girsanov / discrete reweight id).
- Require remove-drift before capital hedge training.
- Name residual-honesty test, threshold, predictor set, and re-test cadence.
- Default completeness to incomplete for production equity unless a written, scoped, expiring completeness argument exists.
- Charge costs under Q; report every ratio with its measure in the same field.
- Refuse silent measure, drift overlay, point on incomplete, and BS-as-alpha passports.
- Clearance for hedge autonomy carries measure_card_hash; a token without a named world is not a hedge right.
- CEH-001 dual marks stay dual; 1.80 cannot be sold as hedgeable peace under a P-trained overlay.
Curious how others separate make-money-from-the-simulator policies from hedge-the-claim policies without pretending the market is complete. Next post: the stack that makes remove-drift a hard stage.
Research fact that belongs in the trench
Deep-hedging research (Buehler et al., arXiv:1802.03042) that separates remove-the-drift from learn-the-hedge is not academic etiquette. It is the only way gradient search stops farming premia. Girsanov and discrete change-of-measure are the arithmetic: same payoffs, different path weights, different story about skill. Incomplete-market theory then forbids uniqueness theater: many Q's, bands not points. Minimum-entropy martingale measure (MEMM) ideas are one way to pick a Q when many exist; they are still a choice, not free uniqueness.
Book: Chapter 7 in Rulers Before Models.