A short-gamma term made a calm market replay historical crashes 20 percent harder

The two multipliers sit on the noise around each replay, and the catalogue underneath is untouched. Six historical drawdowns still go in, 2008 at -56.8 percent and 2000 at -78 percent among them ^[1], and the state now sets how hard each one comes out.

Back out the September 7 run and you can recover the size of the estimate. The short-gamma factor is 1 + max(0, -net_gamma) x 0.30, capped at 1.5 ^[5]. A 20 percent lift with the regime multiplier at 1.0 puts that factor near 1.20, which implies an estimated net-gamma term of about 0.67 ^[6][1]. The cap does not bind until 1.67 ^[2]. A day the regime module scored as calm was therefore already about 40 percent of the way to the amplifier's ceiling ^[3]. Panic regime on top of a saturated gamma term multiplies replay intensity by 3.75 ^[4].

Whether the 20 percent means anything outside this codebase rests on two things the author states himself. Net dealer gamma has to be estimated: full options-chain positioning from OCC is not freely available, so the term is approximated from SKEW and VIX percentiles, and he writes that the sign and direction are structurally right while the point estimate is not a calibrated GEX number ^[9]. The 0.30 weight is a parameter he chose, documented, capped and configurable, not fitted to market data ^[10]. He gives the +20 percent as a within-model result and says explicitly it does not claim the next crash will be 20 percent worse ^[11].

The structural half of the argument survives all of that. In the original design the market state fed the loss calculation while the danger calculation ran without it ^[15], so a calm market and a panic market replayed history the same way ^[2]. Dean, the derivatives-savvy reader whose critique the series is answering, wrote that "in capital markets, the hazard layer is created by the financial contracts themselves" ^[8]. A replay-only hazard layer has no input for positioning at all, so it cannot express that at any calibration.

One porting detail matters for anyone copying the design. The formula yields a single intensity scalar ^[3], so before the multipliers mean anything you have to decide whether your own hazard layer's intensity means severity per event or events per year.

Dean's later comment pushed at the liquidation side instead. Brokers mark at discrete intervals and thresholds sit on round numbers that many accounts share, so liquidation arrives in steps; his example is a 10 percent print skipping 12 and landing at 18 because everyone triggers at 15 together ^[12]. He also raised the liquidation window: a VaR that treats the Fed as a residual prices the news shock but not how long the forced-selling flow stays open ^[13].