Why AAPL Option Risk Models Need Better Scenario Methods

A new paper on arXiv compares two families of scenario-generation methods used across derivatives pricing and risk management. The study tests nonparametric and semiparametric bootstrap techniques on time series and curve data. For anyone running an options book or a VaR model for a stock like Apple, the results have direct implications.

Scenario generation is the part of the risk engine that produces realistic possible paths for the underlying price. Most firms default to historical simulation, which resamples actual past returns directly. That is the nonparametric end of the spectrum. The semiparametric alternative fits a parametric model to the return distribution and then resamples from it. The paper finds that the choice matters most when tail events are in play.

A 5% down day for AAPL is rare. Historical simulation will reproduce whatever the last 500 days looked like. If the window didn't catch a crash, the scenario set will ignore it. The semiparametric approach can stretch the tails beyond what the data shows, depending on the model. The paper argues that the semiparametric bootstrap produces more stable risk estimates across different sample windows, while the nonparametric version tends to be more faithful to the observed co-movements.

That trade-off is central to any trading desk. A risk model that understates tail probabilities leaves margin too thin. A model that overstates them ties up capital unnecessarily. The paper's key contribution is a side-by-side comparison of how each method handles the balance, across both time series and yield curve data.

For a liquid option like AAPL weekly expiries, the scenario generation method feeds directly into the pricing of out-of-the-money puts. If the model uses pure historical simulation and the past year included a VIX spike, the scenario set might overweight a repeat of that move. If the model uses semiparametric bootstrap with a normal or t-distribution assumption, it might underweight fat tails. The paper's dataset covers equities and fixed income, so the findings apply across asset classes.

The authors stop short of prescribing one method over the other. They present the performance differences and let the user decide based on the use case. That fits the reality of quant teams, who need to justify methodology changes to risk committees.

For now, the paper gives desks a reason to audit their current scenario engine. The next step is a backtest against actual out-of-sample tail events for AAPL and other single stocks. A desk that finds its VaR exceeded more often than the confidence level claims may benefit from switching to the semiparametric bootstrap. A desk that already matches its backtest exceptions might prefer the simplicity of the nonparametric alternative.