Beyond Black-Scholes: AI in Options Pricing

Black-Scholes is the most famous wrong model in finance. It assumes geometric Brownian motion, constant volatility, lognormal returns, and continuous hedging, every one of which is empirically false. Despite this, it remains the default options pricing benchmark for two reasons: it has a closed form and the standard deviations from its predictions are interpretable. The job of modern options pricing models, including AI-augmented variants, is to predict the deviations from Black-Scholes more accurately than the model itself.

This article explains where Black-Scholes fails, what classical fixes exist, and how machine learning approaches improve on the classical fixes. The audience is someone who knows what an option is and what Black-Scholes claims to do, but is curious about the modern toolkit.

Black-Scholes recap and its failures

Black-Scholes prices European options under five core assumptions: (1) the underlying follows geometric Brownian motion with constant drift and volatility, (2) no arbitrage, (3) continuous trading is possible, (4) there are no transaction costs, and (5) risk-free borrowing and lending are available at a constant rate. The famous closed-form solution depends on five inputs: spot price, strike, time to expiration, risk-free rate, and volatility.

Empirically, options do not trade at Black-Scholes prices. The market produces a volatility "smile" or "skew", implied volatilities that vary with strike and time to expiration. Equity index options show a strong negative skew (deep OTM puts trade at higher implied vol than ATM and especially OTM calls). Single-stock options show different shapes depending on the underlying. None of this is consistent with the constant-volatility GBM assumption.

The deviations from Black-Scholes are not random, they are systematic and exploitable. Knowing the shape of the volatility surface is essential for any realistic options pricing, risk management, or hedging.

Classical fixes: local and stochastic volatility

Two classical extensions of Black-Scholes address the volatility surface. Local volatility models (Dupire 1994) make volatility a deterministic function of spot price and time: σ(S, t). The local-volatility function is calibrated to match the observed market option prices exactly. This produces a model that, by construction, fits the volatility surface, but it has known limitations: it produces unrealistic forward smile dynamics and is unstable in regions with sparse option data.

Stochastic volatility models (Heston 1993 being the canonical example) introduce a separate random process for volatility: dσ² = κ(θ − σ²)dt + ξσ dW, correlated with the asset price process. This is closer to economic reality (volatility is itself a random variable) and produces more realistic forward dynamics. The cost is two extra parameters (κ, θ, ξ, and the correlation ρ), and the closed-form solution is much more complicated than Black-Scholes, typically requiring numerical Fourier inversion to compute.

Hybrid local-stochastic volatility models combine both, local volatility provides the fit to the static surface, stochastic volatility provides realistic dynamics. This is the workhorse of institutional derivatives desks.

Where AI improves on classical models

Machine learning approaches to options pricing solve specific subproblems within the classical framework, not the framework itself. The two main applications:

First, volatility surface fitting. Classical methods fit parametric models (SVI, SABR) to the observed option prices. ML approaches use neural networks to fit the surface non-parametrically, with arbitrage-free constraints enforced via the network architecture or loss function. The result is a smoother, more accurate surface fit, especially in regions where option data is sparse. Companies like Numerai and many bank derivatives desks use neural-network surface fits in production.

Second, calibration of complex models. Stochastic volatility models with closed-form solutions (Heston, SABR) require numerical calibration to fit market prices. ML approaches train neural networks to approximate the price-to-parameter map, making calibration thousands of times faster. Once trained, the network produces calibrated parameters in microseconds, vs. seconds for direct numerical calibration. Hernandez (2017) and follow-up papers showed that this approach is now standard in institutional derivatives.

End-to-end ML pricing

A more radical approach is to skip the classical model entirely and train an ML model to predict option prices directly from market state. The input is the spot price, strike, expiry, and a rich set of features (historical volatility, term structure, microstructure indicators). The output is the option price. This works well empirically in liquid markets, neural networks can learn the volatility-surface structure from data without being told about Black-Scholes.

The downside of end-to-end pricing is interpretability and risk management. Black-Scholes has explicit Greeks (delta, gamma, vega) that traders use to hedge. A neural network does not produce Greeks natively, though they can be extracted via automatic differentiation. For pricing-only applications, end-to-end ML is competitive; for hedging applications, the model-based approaches (local vol, stochastic vol, with ML for surface fitting and calibration) are more practical.

Implied volatility forecasting

A separate application of ML in options is forecasting implied volatility, predicting how the volatility surface will evolve. This matters for any strategy that involves holding options over time, where changes in implied vol affect P&L through vega exposure. Classical models (GARCH, HAR-RV) provide baselines; ML models (LSTM, gradient boosting) can outperform them by exploiting cross-sectional structure across strikes and maturities.

The signal-to-noise ratio in implied vol forecasting is low, implied vols are themselves forecasts and incorporate most predictable information. The marginal improvement from ML is usually a few percent in RMSE, which compounds meaningfully over many trades but does not produce dramatic alpha on its own.

Where Black-Scholes still wins

Black-Scholes remains the standard for one critical purpose: the implied volatility quote convention. Market participants quote and reason about options in implied volatility space because BS provides a stable mapping between dollar prices and a more comparable "volatility" number. The price quoted in IV space is comparable across strikes and maturities; the dollar price is not. Even sophisticated stochastic-volatility traders quote in BS implied vols and translate to dollar prices via BS inversion. The model is wrong, but the convention it provides is irreplaceable.

Black-Scholes is also useful as a sanity check. If your sophisticated ML pricing model produces a price that is wildly different from the BS price at the BS implied vol implied by the market, you have a bug. The model is wrong, but its errors are small enough that BS prices are useful approximations for sanity-checking other models.

ARIA Analyst and options pricing

ARIA Analyst does not provide options pricing as a service, it provides options-informed features for equity scoring. The volatility surface for a stock reveals information that does not appear in the cash equity price: the skew tells you about institutional hedging demand, the term structure tells you about expected events (earnings, regulatory decisions), and the put-call ratio is a sentiment indicator. ARIA Analyst extracts approximately 15 features from the options market per stock and feeds them into the deterministic scoring agents and the ML ensemble.

For users who want options pricing directly, the standard institutional vendors (Bloomberg, OptionMetrics) provide it. ARIA Analyst integrates with these via the Premium tier for users who need both stock scoring and options pricing in one pipeline.

Conclusion

AI does not replace Black-Scholes, it sits alongside classical options pricing models, improving the parts that the classical framework gets wrong (surface fitting, calibration speed, implied vol forecasting) while leaving the parts that work (no-arbitrage structure, Greeks for hedging, IV quote convention) intact. The state of the art in options pricing is a hybrid: classical model for structure, ML for parameter fitting and surface representation, end-to-end models for liquid pricing-only applications.

For equity-focused investors, the most actionable use of options data is as a feature for stock scoring, and ARIA Analyst does this by default. Create a free account to see options-informed scoring on any stock, or read our GARCH guide for the volatility-modeling foundation. See feature engineering for the features ARIA extracts from option chains.