Diffusion Models for Synthetic Market Data

A diffusion model learns to generate realistic data by training on the reverse of a simple process: take real market data (say, a sequence of returns), gradually add small amounts of random noise over many steps until it's pure noise, then train a neural network to undo that noising one step at a time. Once trained, starting from pure random noise and running the learned reverse process produces a brand-new synthetic sample that looks statistically like the real data it learned from.

For quant use, this matters because historical market data is limited, you only get one realization of history, and rare events like flash crashes or specific liquidity crunches may appear only a handful of times, if at all. A diffusion model trained on real price paths or order-book sequences can generate large numbers of additional synthetic-but-realistic paths, useful for stress-testing a strategy against scenarios beyond the exact history observed, or for training downstream models that need more data than history provides.

A firm might train a diffusion model on ten years of five-minute equity returns, then generate ten thousand synthetic one-day paths sharing the same volatility clustering and fat-tailed return distribution as the real data, feeding those into a backtest to see how a strategy behaves across a much wider range of plausible-but-never-observed sequences than the single historical path allows.