Exogenous-Shock Position Sizing

Realized-volatility regime scaling, the correlated-book contamination haircut, and separating an endogenous reversal from a one-driver risk-off.

Last week we sized against a clock. This week the shock arrives with no clock at all. You cannot time what you cannot see coming, so you size for the regime it lands in, and you refuse to score a contaminated drawdown as a signal that failed.

Last Friday's lesson built the event-gated size: a rule that shrinks a position because a scheduled print sits inside its horizon and the leg is crowded enough to snap. That rule has one precondition it cannot supply for itself. It needs the event to be on a calendar. A jobs print is. An FOMC decision is. A token unlock is. But the drawdown that hit the book this week had no timestamp. A geopolitical shock repriced oil, equities went risk-off, and crypto re-coupled to the move inside a session. The clock-based defense had nothing to read.

This lesson supplies the defense that works when the calendar is empty. It does two things last week's rule could not. It scales size to the volatility regime the position is held through rather than to a scheduled event. And it teaches the book to grade a loss by its composition, so that a drawdown riding a market-wide driver is held on probation instead of tripping the exit gate a real signal failure would trip.

§ IFrame

Twenty-two days ago the desk opened a directional leg on a flow signal. For twenty-one of them the flow confirmed. Then this week's settle printed a net outflow that broke a three-day inflow streak, and on the surface that is exactly the shape of a thesis breaking. A signal-driven book that only reads its own P&L flattens here. It sees the streak end, calls the reversal, and exits.

The desk did not exit. It read one more variable before it graded the loss: the composition of the tape the loss arrived in. The outflow was coincident with a geopolitical shock that sent oil up and equities down together, and crypto moved with the equity risk-off rather than against it. Every book on the desk was losing the same day for the same reason. That is not a flow signal failing. That is one exogenous driver pushing correlated books off a cliff at the same time.

Name the missing layer the contamination read: a rule that asks, before grading a drawdown, whether the loss is idiosyncratic to the signal or shared with a market-wide driver, and holds the thesis on probation when the loss is shared. Three inputs feed the sizing and grading discipline that surrounds it — the volatility regime, the cross-book correlation, and the composition of the drawdown.

§ IIFoundations

Size is inverse to the volatility you are holding through

Carver's treatment of position sizing (Systematic Trading, Ch 10, pp. 173–176) puts one idea at the center: a position should be scaled so that its expected risk contribution is roughly constant across time, which means the dollar size moves inversely to the volatility of the instrument. When realized volatility doubles, the position that carries the same risk is half the size. This is volatility targeting, and its whole purpose is to stop the book from carrying more risk in a storm than it carried in calm simply because it never resized.

The exogenous shock is a volatility event before it is anything else. Whatever else the geopolitical print did, it widened the realized distribution of every risk asset at once. A book still sized for last week's calm is now carrying twice the risk it signed up for, on a signal that has told it nothing new. Volatility targeting cuts that automatically, before any judgment about the signal is required.

A crowded regime converts a shock into a stampede

Chan's account of the mechanics (Machine Trading, Ch 1, p. 38) frames sizing as the join between a signal's edge and the capital at risk behind it, and his microstructure treatment (Ch 6, pp. 191–194) supplies what happens under stress: liquidity thins exactly when everyone wants the same exit, so the realized cost of a move in a stressed regime is far larger than the same move in a calm one. Aldridge's statistical-arbitrage frame (High-Frequency Trading, Ch 8, pp. 217–218) adds the correlation piece: the assumption that a pair or a book is market-neutral holds only while the residual stays uncorrelated with the market factor, and a shock is precisely the event that collapses those correlations toward one.

Put together, the lesson is that risk in a shock is not the sum of the individual book risks. It is larger, because the books that were independent in calm become one book in the storm. The size discipline has to see the correlation, not just the volatilities.

Grade the drawdown by its composition, not its sign

The hardest discipline is the grading rule. A book that exits on every streak-ending loss will be shaken out of every true signal by the first market-wide risk-off day. A book that never exits will ride a genuine reversal into the ground. The line between them is not the size of the loss. It is the composition. A loss that is idiosyncratic to your signal — your instrument down while the market factor is flat, no shared catalyst — is evidence the signal is failing. A loss that is shared with the market factor — your instrument down on a day the whole risk complex is down on one exogenous driver — is evidence about the world, not about your signal. Same P&L number, opposite meaning.

§ IIIMechanism

The discipline runs in two loops that share the same inputs. A sizing loop that sets tomorrow's weight, and a grading loop that decides what a drawdown means.

The realized-volatility regime scale. Estimate realized volatility over a short trailing window. An exponentially-weighted estimate reacts faster to a regime break than an equal-weighted one, which matters when the whole point is to catch a shock early. Divide a baseline target volatility by the current realized estimate to get a size multiplier, and clamp it between a floor and one so the book never levers up into calm or vanishes in a storm. When realized vol sits at baseline, the multiplier is one and the signal's target weight passes through. When realized vol spikes to twice baseline, the multiplier is one-half, and the position is cut before any view about the signal is formed.

The correlated-book contamination haircut. Estimate the realized correlation of each book's returns to the market risk factor over the same short window. When that correlation is low, the books diversify and their risks add in quadrature. When it is high, the shock regime, they add closer to linearly, and the aggregate risk the desk is carrying is far larger than the per-book volatilities suggest. Apply a haircut to gross exposure proportional to how far the average cross-book correlation has risen above its calm baseline. The haircut is the correction for the fact that a shock deletes the diversification the position limits assumed.

The composition-over-net grading rule. When a drawdown lands, do not grade it on the net number. Decompose it. Measure the drawdown's correlation with the market risk factor over the shock window and check for a shared exogenous catalyst. If the correlation is high and a catalyst is present, mark the loss contaminated: hold the thesis on probation, keep the asterisk on the record, and do not trip the exit gate. If the correlation is low and no catalyst is present, mark the loss idiosyncratic: this is the endogenous reversal the exit gate exists to catch, and it should trip. The sizing loop has already cut the position for the volatility regime; the grading loop decides whether the position survives to be cut again tomorrow or is closed today.

The order is fixed. Realized-vol regime scales the size. The correlation haircut caps the aggregate. When a loss arrives, the composition read grades it before the exit gate is allowed to fire. Only an idiosyncratic loss trips the gate.

The DevOps layer: the regime is a stream, not a snapshot

A regime read taken once a day at settle is a regime read that is stale by the time it is used. The operational discipline is to run the volatility and correlation estimates as a continuous stream that the sizer subscribes to, updating the multiplier and the haircut as new prints land, with the last good estimate held under a freshness deadline so a stalled feed degrades to a known-conservative default rather than to a silent stale number. The sizer asks the stream one question — what is the current regime multiplier and the current contamination haircut — and the stream answers from live data with a timestamp, so a size decision made under stress is reproducible and its inputs are auditable after the fact.

§ IVWorked Example

Take the live book. The flow signal has held a directional leg for twenty-one sessions. This week's settle prints a net outflow that breaks a three-day inflow streak, on its face the streak-ending loss that grades a thesis.

Read the regime first. Realized volatility across the risk complex has jumped: oil spiked on the geopolitical print, equity realized vol widened, and the crypto leg's realized vol rose with it. The regime multiplier drops well below one, so tomorrow's target weight is already cut by the volatility scale alone, independent of any grading decision.

Read the correlation. Over the shock window the crypto leg's return correlation to the equity risk factor has risen sharply from its calm baseline. The books that were near-independent last week are moving as one book today. The contamination haircut binds, and aggregate gross is cut further so the desk is not carrying stacked risk it believes is diversified.

Grade the drawdown by composition. The outflow is coincident with the exogenous catalyst, and its correlation with the market risk-off is high. Mark it contaminated. The thesis is held on probation with the asterisk intact; the exit gate does not fire. The loss is real money, but it is evidence about the shock, not about the signal.

Now the discriminating test the composition rule sets up. The next settle grades the crack. If flow re-asserts as the exogenous day gives back — an inflow that reverses the contaminated print — the contamination read is confirmed and the thesis is clean, restored off probation. If instead a second consecutive outflow lands after the correlation has faded, with the market factor no longer supplying the shared driver, then the loss has turned idiosyncratic, the probation was the last grace the composition rule allows, and the exit gate trips on an endogenous reversal that has finally isolated itself from the shock. The rule does not refuse to exit. It refuses to exit on a contaminated print, and it names in advance the observation that would flip the verdict.

§ VConnection to Prior Lessons

This is the fourth beat of the sizing-and-exit sequence. June 30 built the sequenced exit gate, the machinery that confirms a withdrawal across ordered legs and filters a fake reversal. July 2 built the dispersion classifier that reads the regime and names it. July 3 built the event-gated size that shrinks a position against a scheduled shock with a public clock. Today supplies the piece those three assumed away: the shock with no clock, graded by composition so the exit gate the June 30 lesson built is not tripped by the market-wide day the July 3 lesson could not have seen coming.

The through-line is the one the week's synthesis kept surfacing: isolation. The dispersion classifier isolates the rotation axis from the cohort's shared direction. The event gate isolates the scheduled hazard from the signal's own horizon. The composition rule isolates the exogenous driver from the endogenous reversal. Each instrument does the same work on a different object, finding the one true driver and holding it apart from the correlated surface that hides it.

§ VIConnection to Today's Dev Lesson

The regime scale and the contamination haircut are only useful if they update fast enough to matter, from several noisy feeds, without blocking the sizer when one feed stalls. That is a streaming-concurrency problem: fan several volatility and correlation estimators into one regime read, bound each read by a deadline so a slow feed cannot hang the decision, and let the sizer read the current multiplier without racing the writers. Today's Go lesson builds exactly that — a streaming volatility-regime sizer using Go's pipeline pattern and context, fanning realized-vol estimators into one regime channel, bounding each estimate under a deadline, and guarding the size multiplier with an atomic so reads never tear. The Ops lesson defines what the regime multiplier means; the Go lesson makes it fresh, bounded, and safe to read under stress.

§ VIIClosing

A signal that has been right for twenty-one days does not stop being right on the first day the whole market falls together. Size for the volatility regime before you form any view, cut the aggregate for the correlation the shock deleted, and grade the drawdown by whether it is shared with the market or idiosyncratic to your signal. Hold the thesis on probation when the loss is contaminated, and name in advance the isolated loss that would finally trip the gate. The exogenous shock has no clock, so you defend against it with regime and composition instead of with time.

A loss shared by the whole market is evidence about the market. Only a loss your signal owns alone is evidence about your signal.