LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
A neural network can learn a pricing function the way it learns any other map from inputs to an output. Black-Scholes is the classroom case, because the correct price is already a formula. The useful case is a model with no usable formula. Simulate it many times offline, train the network on parameter sets and prices, and evaluate the network when a live calibration cannot afford the simulation. The lecture's Black-Scholes check is close on some rows and nonsense on others. Learning happened. It was not exact, and it was not free of error outside the training box.
Why it matters
Black-Scholes assumes geometric Brownian motion, constant volatility, no arbitrage, no transaction costs, and continuous hedging. Markets violate those. Richer models put volatility on its own mean-reverting path, as Heston does, let volatility depend on the spot, or add jumps. Some rough-volatility models use a fractional driver and long memory. The richer the model, the less often a closed price exists. You can still simulate paths, discount the payoff, and average. Doing that inside a calibration, where every trial parameter set needs a fresh grid of prices, is too slow for a book that must be remarked. The literature the lecture describes moves the simulation offline.
The idea in plain words
Five Black-Scholes inputs, spot, strike, time, rate, and volatility, return one price. A Monte Carlo estimate of the same contract is the discounted risk-neutral expectation of the positive part of spot-at-expiry minus strike. Same inputs, approximate price. A network is a third map: parameters in, price out, with weights chosen to shrink the gap to the simulated or analytic prices. If that gap can be driven to zero, the three maps agree. The lecture does not claim a theorem that the gap vanishes. It says the hope is reasonable and the check is empirical. The warmup is f(x) = x squared on the interval from minus 2 to 2. In sample, a lightly trained network fits the parabola closely. Universal approximation is the name for that ability. Outside the interval the fit breaks. Vertical lines on the lecture's plot mark the edge of the training data, and past them the curve leaves the parabola. Option networks have the same habit. A price request outside the simulated box is an extrapolation, and the warmup says not to trust it. For Black-Scholes the lecture builds 5,000 parameter rows and 5,000 formula prices, then trains a small network for a thousand epochs by minimizing error. Row 3,817 had a true price of 30.449 and a network price of 30.679. Row 425 had a true price of 0.142 and a negative network price.
How to use it
Simulate the model you actually want. For Heston the lecture lists an eight-dimensional input: initial spot, initial variance, time, rate, and the variance-process parameters, including the correlation between the two Brownian motions. There is no elementary price. A discretization produces paths. The average discounted payoff is the label. Train on many such pairs while the clock does not matter. The lecture's Heston demo used only 500 paths and 1,000 epochs and called the errors poor on purpose. Raise both if you need a stated accuracy. The error of this two-stage approximation, Monte Carlo noise inside the labels, plus network error on top, is described as well studied. A confidence interval is available. The lecture does not report one for the Heston demo.
Use the trained map only inside the parameter box you simulated. Recalibration against a live volatility surface is then a fast function call instead of a nested simulation. You pay in approximation error. The lecture's claim is that the payment can be small enough to be worth the dynamics you refused to assume away.
What usually goes wrong
Deploying the network past the training range, which the parabola already fails. Accepting a negative call price, as row 425 did. Training on 500 paths and expecting production accuracy. The lecture says that setup is awful if precision matters. A fourth error is forgetting that a network trained on model prices inherits the model's mistakes. It does not learn the market. It learns the map you simulated. Fast Fourier methods for Heston exist. The lecture sets them aside to keep the example about simulation.
Questions people ask
What were the two Black-Scholes spot checks?
Parameter row 3,817: formula 30.449, network 30.679. Parameter row 425: formula 0.142, network negative. Nearby rows such as 26.97 versus 27.1 were tighter. All of them are approximations.
Why train on Black-Scholes at all?
Because the formula can grade the homework. If the network cannot stay near a known price inside a box of 5,000 examples, it will not be a trustworthy stand-in for a model you cannot grade so easily.
What did the Heston demo refuse to claim?
A calibrated production error. It used 500 paths and 1,000 epochs, showed a chart of fit and error, and said the point was the offline-training idea, not an optimized architecture.