Key takeaways
  • Pathwise labels can mislead on jumps
  • Likelihood-ratio labels suit barrier options
  • Gamma adds extra regularization
  • Price and sensitivity errors can fall

When a pricing model has sharp jumps, a neural network (a software system inspired by how brain cells connect) can learn the wrong lesson from the wrong kind of sensitivity. This paper looks at differential machine learning, a way to train networks on both prices and price sensitivities, and shows where the usual pathwise adjoint differentiation breaks down. For discontinuous payoffs like digital or barrier options, those pathwise sensitivities are biased, and feeding them into the loss function can actually magnify errors. The authors test alternative sensitivity estimators and find that they can substantially reduce test errors in both prices and sensitivities. They also show that differential labels built with the likelihood ratio method broaden the approach to discontinuous payoffs, while a hybrid scheme that adds gamma estimates alongside delta estimates gives extra regularization. The result is a cleaner route to fast approximations for derivatives pricing and risk management when the payout is not smooth.

A barrier option can pay only if a market price crosses a line. Digital options can jump from nothing to something in one step. That sharp edge is where Differential ML gets tricky. Differential ML trains a neural network. A neural network is an artificial brain-inspired system. It learns from prices and price sensitivities together. Sensitivities answer a plain question. How much does the price move when the market shifts? Traders call those numbers Greeks. The catch is blunt. When the payoff jumps, the usual sensitivity label can mislead the network. It can even make test errors larger. This study's surprise is that a bad clue can do more harm than no clue.

When smooth clues fail

The main result is practical. The usual pathwise sensitivity follows each simulated market path. It works poorly when the payout has a jump. For digital and barrier options, that label is biased. When that biased label enters training, the network can lock onto the wrong lesson. Swapping in sensitivities from the likelihood ratio method cuts that risk. Tests show these alternative labels can cut test errors in both prices and sensitivities. It also pushes Differential ML beyond smooth contracts. A hybrid version goes one step further. It adds gamma, the second sensitivity, alongside delta, the first sensitivity. That extra signal gives more regularization. Regularization means a small nudge that keeps the model from memorizing noise.

How the labels change

Training starts with simulated market paths. Each path gives a price label. Differential ML adds sensitivity labels too. The older route uses pathwise adjoint differentiation. That is a way to trace small input changes through a payout. It works when the payout changes smoothly. It breaks when the contract jumps at a barrier. The alternative is the likelihood ratio method. It builds a sensitivity label from how likely a path is. It does not depend on a tiny bump in the input. The hybrid method then adds gamma labels. Gamma is the second sensitivity. Delta is the first sensitivity. The network sees both.

  • Pathwise adjoint differentiation tracks each simulated path and works best for smooth payoffs.
  • The likelihood ratio method scores paths by how likely they are and still works on jumps.
  • The hybrid scheme adds gamma, the second sensitivity, to regularize training further.

“the derivatives of a good approximation are not always a good approximation of the derivatives.”

— Huge and SavineFrom the introduction

Why this matters for jumpy payoffs

For fast pricing, speed only helps if the answer is trustworthy. Differential ML aims to replace slow simulation with a quick network. This result makes that aim wider. Likelihood-ratio labels let the same idea reach digital and barrier options. Those contracts change in a jump, so smooth tricks often stumble. The hybrid delta-and-gamma setup adds another check on training. In plain terms, the network gets better clues. Better clues can mean better prices and better hedges. When the clue is wrong, both can drift together. That matters because risk control depends on the sensitivities, not just the price.

What still has to hold up

The surprise from the start still holds. A sensitivity signal can hurt if it comes from the wrong estimator. The fix is not to give up on sensitivity labels. It is to choose a label that survives jumps. That makes Differential ML useful for barrier and digital options. They no longer sit outside sensitivity-trained neural nets. That is the concrete gain from the switch. The next check is simple. The likelihood-ratio version still has to hold up across more contracts and market settings.