Diffusion Model Outperforms Monte Carlo on European and Asian Options

When a pricing model misses fat tails, it can misjudge how risky a trade really is. That matters for exotic options and structured products, where payoffs depend on the whole path of prices, not just the final number. This paper adds a diffusion model to the task: a Diffusion-Conditional Probability Model that generates price paths with a composite loss built around volatility clustering, tail risk, and drift constraints. The model’s static checks show it matches market mean and volatility closely, although it still falls short on kurtosis, a measure tied to extreme outcomes. In adversarial backtesting, the P-model, which uses the diffusion-based trader, beats a GBM Monte Carlo market maker on European and Asian options. It earns higher profitability there because it tracks terminal values and path means better. But the picture changes for lookback options and structured products such as snowballs and accumulators. Those contracts are highly sensitive to extreme events, and the model tends to underestimate tail risks, which can lead to losses. The paper’s main takeaway is balanced: diffusion models with finance-specific losses can improve pricing for some derivatives, but extreme market risk remains a hard problem.

A price can end in the right place and still ruin a trade. That is why exotic options worry traders. The new DDPM model tries to follow the whole path, not just the finish line. In static checks, it tracked market mean and volatility well. In live-style games, it beat a Monte Carlo market maker, a system that prices by running many random paths, on European and Asian options. The catch is sharper. When a contract cared about rare spikes and drops, the same model could lose.

Why the path matters more than the finish line

The P-model did well on the basic shape of prices. It matched market mean and volatility in static tests. Mean is the average level. Volatility is how much prices jump around. The model also lined up closely with real market distributions. It still missed kurtosis, the part that tracks rare, extreme moves. That gap showed up in the trading game. The diffusion-based trader made much more profit on European and Asian options. Those contracts care most about the final price, or the average path. The same edge did not carry over to lookback options. Those payoffs depend on the best or worst price seen along the way. Snowballs and accumulators also exposed the weak spot. They are highly sensitive to extreme events. The model underpriced that danger.

How DDPM turns noise into a price path

DDPM starts from noise and learns to turn it into price paths. In this paper, the model adds finance-focused penalties to its loss, the score it tries to lower during training. One penalty pushes the paths to keep realistic volatility clustering, where choppy days bunch together. Another pushes the tails, the rare big jumps, to look more like markets. A drift rule keeps the path's long-run tilt in line with prices. The P-Q game then tests the result. The P-model acts like a diffusion-based trader. The Q-model acts like a GBM Monte Carlo market maker. GBM means geometric Brownian motion, a classic model with smooth random moves. The two sides face off in adversarial backtesting, a trade test that pits one model against another.

Where it helps, and where it still slips

This matters because pricing is not just about the finish line. It is about the whole shape of a price path. The P-Q game gives a new way to ask a blunt question. Does a model make money against a simple rival? That test exposed a split result. The diffusion model looks strong when payoffs depend on average movement or terminal value. It looks weak when payoffs hinge on rare shocks. That split helps traders and risk teams avoid false comfort. A model can pass neat fit checks and still miss the contracts that punish tail risk.

What the next stress test must prove

The surprise is still the same. A diffusion model can win on European and Asian options, then stumble on snowballs and accumulators. That split says path realism is not enough. Tail risk still rules the hardest products. The clearest next test is whether DDPM can stop underweighting rare shocks on snowballs and accumulators. If it can, the same kind of model could move from a nice fit to a useful price tool. If it cannot, the safest role stays narrower. It will help on contracts that care about average movement. It will not be the last word on extreme events.