Neural Network Option Pricing

A closed-form option pricing formula is fast; a Monte Carlo or PDE solver for an exotic model can take seconds per price, and a real-time calibration might need that price evaluated tens of thousands of times as it searches for the parameters that fit the market. Waiting seconds, tens of thousands of times, is not something a trading desk can do between quotes. The fix sounds almost too simple: run the slow pricer once, offline, on a huge grid of inputs, and train a neural network to reproduce its outputs. Afterward, evaluating the network takes microseconds.

The world's most expensive lookup table

Think of a chef who spends a weekend cooking every possible combination of a dish's ingredients and writing down how it tasted, then hands a new cook a laminated card summarising the pattern — "more salt up to about a teaspoon per pound helps, then it stops helping." The new cook can now guess how any similar dish will taste almost instantly, without ever touching a stove, because the pattern was learned once and reused. A neural network pricer is that laminated card: the expensive cooking happens once, offline (training), and the fast guessing happens every time afterward (inference).

What the network is actually fitting

A feed-forward network is just a chain of simple functions. With one hidden layer, it computes

$$\hat{V}(x) \;=\; W_2 \, \sigma\!\big(W_1 x + b_1\big) + b_2,$$

where $x$ is the vector of inputs (strike, maturity, and model parameters like volatility or mean-reversion speed), $W_1$ and $W_2$ are matrices of weights, $b_1$ and $b_2$ are offset vectors, and $\sigma$ is a nonlinear "squashing" function applied entrywise. In plain English: take the inputs, mix them together with learned weights, bend the result through a nonlinearity so the network can represent curves rather than only straight lines, then mix again to produce one number, the price. Training means adjusting every weight in $W_1, W_2, b_1, b_2$ so that $\hat V(x)$ matches the true pricer's output $V(x)$ as closely as possible across the whole training grid, done by gradient descent via backpropagation.

Worked example 1: fitting three points with a one-node network

Suppose the true (slow) pricer, evaluated at three strikes for a fixed maturity, gives prices $V(90)=12.0$, $V(100)=5.0$, $V(110)=1.5$. A tiny network with one hidden unit and a linear activation is just $\hat V(K) = w_2(w_1 K + b_1) + b_2$, which collapses to a straight line $\hat V(K) = aK + b$. Fit it with two of the three points, say $K=90$ and $K=110$: slope $a = (1.5-12.0)/(110-90) = -10.5/20 = -0.525$, intercept $b = 12.0 - (-0.525)(90) = 12.0 + 47.25 = 59.25$. The line predicts $\hat V(100) = -0.525(100) + 59.25 = 6.75$, against a true value of $5.0$ — an error of $1.75$. A linear network cannot bend, and option prices are convex in strike, so it systematically misses the middle. This is exactly why real pricing networks use a nonlinear $\sigma$ and several hidden units: enough bends to trace a convex curve through all the training points, not just fit a straight line between two of them.

Worked example 2: training speed versus inference speed

Say the true pricer (a Monte Carlo engine) takes $0.4$ seconds per price, and a calibration routine needs $50{,}000$ evaluations to search for model parameters that match the market. Direct calibration: $50{,}000 \times 0.4\,\text{s} = 20{,}000$ seconds, about $5.6$ hours — unusable intraday. Training the network instead uses, say, $200{,}000$ training prices computed once offline: $200{,}000 \times 0.4\,\text{s} = 80{,}000$ seconds, about $22$ hours, but this happens overnight, not during trading. Then each network evaluation during the day takes about $0.00002$ seconds ($20\ \mu s$), so the same $50{,}000$-evaluation calibration costs $50{,}000 \times 0.00002\,\text{s} = 1$ second. A calibration that was impossible intraday becomes instantaneous, at the one-time cost of an overnight training run.

What this means in practice

Neural pricers are used to accelerate calibration of expensive stochastic-vol and local-stochastic-vol models, and to price American or path-dependent products where a closed form doesn't exist, letting a full model calibration that used to take hours run in seconds. The network never replaces the underlying model — it replaces the cost of evaluating it.