Introductory Context
"The history of options markets is punctuated by model risk failures: the 1987 Black Monday crash (when portfolio insurance strategies based on continuous delta hedging failed catastrophically because markets gapped down without allowing rebalancing), the 1998 LTCM crisis (when fixed-income convergence trades premised on normal correlation relationships broke down when Russia defaulted), and numerous smaller losses by firms that relied on model-generated hedge ratios that failed during market dislocations. Understanding model risk -- what causes it, how to identify it, and how to manage it -- is as important as understanding the models themselves. "
Sources of Model Risk in Options Trading
Parameter estimation risk: every model requires estimated parameters (Black-Scholes' σ, Heston's 5 parameters, Merton's jump parameters). These parameters are estimated from historical data or calibrated to current market prices -- both methods produce estimates with uncertainty. When the true parameter is outside the estimated confidence interval, the model produces systematically wrong prices and hedge ratios. For Nifty: the post-COVID period showed that historical volatility estimates from the calm 2017-2019 period drastically underestimated the March 2020 volatility -- a direct parameter estimation failure.
Model misspecification risk: every model assumes a specific form for the underlying's dynamics (log-normal diffusion, mean-reverting variance, Poisson jumps). If the true dynamics fall outside the model's assumed family of possibilities, the model is fundamentally misspecified. Black-Scholes misspecifies by assuming constant volatility. A jump-diffusion model misspecifies if the 'jumps' are actually regime changes that persist rather than instantaneous discontinuities. No model captures all aspects of reality -- every model is misspecified in some dimensions.
Calibration instability: a model calibrated to today's options market may have very different parameter values tomorrow, even if the underlying's price has barely changed. Heston model calibration (fitting 5 parameters to a vol surface) is notoriously unstable: small changes in market prices can produce large changes in calibrated parameters, creating large day-to-day changes in the model's risk calculations even when 'nothing important' has happened. This calibration instability propagates into hedging instability -- the hedge ratios change dramatically not because the position's true risk changed but because the calibration shifted.
Model Risk in Indian Options Markets - Specific Examples
Pre-event IV spike mismodelling: Black-Scholes models calibrated to normal-period IVs dramatically underestimate pre-Budget and pre-election option prices. A model based on historical 14% vol would price the RBI-meeting week 16.5% vol options as overpriced -- when in fact they are fairly priced for the event risk. Trading based on this model's 'overpriced' signal would create systematic losses.
Jump risk underestimation: diffusion-only models (Black-Scholes, Heston with no jump component) underestimate deep OTM put prices because they assign near-zero probability to the crash scenarios that justify those prices. A risk system based purely on diffusion models would show seemingly manageable VaR for a large short deep-OTM-put book -- until the actual crash occurs and the true tail risk materialises.
Model Risk Management Framework
- Model identification: document every model used, its assumptions, and its known failure modes. 2. Parameter sensitivity: test how P&L and risk metrics change when model parameters are varied within plausible ranges. 3. Model reserve: hold additional capital (a 'model reserve') proportional to the uncertainty in model valuations, particularly for illiquid or complex positions. 4. Multiple model comparison: price significant positions with two or more models; the difference is an estimate of model uncertainty. 5. Stress testing: run the portfolio through scenarios that violate model assumptions (sudden volatility spike, large jump, correlation breakdown). 6. Regular model validation: periodically compare model predictions to actual market outcomes. Models that systematically misprice should be re-calibrated or replaced.
The Model Reserve Concept
The model reserve is a buffer of capital held against the uncertainty of model valuations. For a position priced at Rs 100 lakh by Model A but at Rs 85 lakh by Model B (alternative model), a prudent firm might apply a model reserve of Rs 7.5 lakh (50% of the Rs 15 lakh model difference) -- reporting the position's value as Rs 92.5 lakh (the Model A value less the reserve). This conservative valuation accounts for the possibility that the true value is closer to Model B's estimate. SEBI's risk framework for registered investment advisers and portfolio managers requires some form of model validation, but does not prescribe specific model reserve methodologies -- individual firms develop these based on their risk management philosophy.
The Practical Retail Options Trader's Model Risk
For retail options traders who do not use sophisticated pricing models: model risk manifests primarily as reliance on Black-Scholes Greeks (delta, gamma, vega) from broker platforms that may not account for the volatility surface's slope and curvature. The 'sticky strike' delta (Black-Scholes delta that assumes IV is constant as the underlying moves) is wrong in a world where IV changes with the underlying's level. The 'sticky delta' adjustment (recognising that as the underlying falls, the IV at the new ATM increases -- moving up the skew) provides a better hedge ratio. The difference between sticky-strike and sticky-delta hedging is the retail trader's version of model risk: using the wrong delta because of an incorrect model assumption.
Model risk is the hidden variable in options trading: it is present in every trade, every hedge, every risk report, and every position value -- but invisible until the market violates the model's assumptions. The sophisticated practitioner maintains a healthy skepticism toward every model output, asking always: 'Under what conditions would this model give the wrong answer, and is the current market condition one of those conditions?' This skepticism is not paralyzing -- models are still the best available tools. But the practitioner who treats model outputs as facts rather than estimates is setting up for the moment when the model fails in exactly the way it was always known to be capable of failing.
Never Treat Model Output as Ground Truth for Illiquid Positions
The danger of model risk is greatest for illiquid positions where no market quote exists to challenge the model's valuation. A deep OTM Nifty call with 6 months to expiry may have a model value that cannot be independently verified because no one is actively trading it. The model could be significantly wrong without any market feedback. For illiquid positions: (1) apply conservative valuation adjustments (haircuts) to the model price, (2) use two independent models and take the more conservative value, and (3) assume the position cannot be exited at the model value -- always verify liquidity before relying on model values for P&L calculation.