Why I'm Betting on Diffusion Models for Finance

Everyone knows diffusion models for what they did to images. Stable Diffusion, DALL-E, Midjourney they made generative AI a household concept. But here's what most people outside quantitative finance haven't noticed: diffusion models are quietly becoming the most promising architecture for financial time series generation. And I'm building one. Why Finance Needs Generative Models? Traditional financial forecasting relies on parametric models GARCH for volatility, Black-Scholes for options pricing, VAR models for risk. These work well in "normal" markets. They fall apart during exactly the moments you need them most: tail events, regime changes, flash crashes. The fundamental limitation is that parametric models assume you know the shape of the distribution. Financial markets don't care about your assumptions. Generative models flip this. Instead of specifying a distribution and fitting parameters, you learn the distribution directly from data. The model captures whatever structure exists fat tails, volatility clustering, cross-asset correlations without you having to hard-code it. Why Diffusion Models Specifically? GANs were the first generative architecture applied to financial data. But GANs have well-documented problems: mode collapse, training instability, and difficulty capturing the full diversity of a distribution. Diffusion models solve these architecturally. The forward process gradually adds noise to real data. The reverse process learns to denoise step by step recovering structure from chaos. This iterative refinement is fundamentally more stable than the adversarial game of GANs. For financial data, there's an elegant connection: geometric Brownian motion (the foundation of modern options pricing) IS a diffusion process. When researchers align the forward noising schedule with GBM dynamics, the model naturally respects the heteroskedastic nature of financial returns. The Research Wave Is Here? This isn't speculative. Diffolio (2026) combines diffusion with hierarchical attention for portfolio construction. LFTD uses transformer-enhanced diffusion for firm-level data augmentation. Multiple groups are exploring GBM-aligned SDEs. I believe we're at the same inflection point that NLP hit when transformers arrived the right architecture meeting the right problem. I'll be sharing technical deep-dives on my implementation over the coming weeks. --- *Next up: what 90% of ML engineering actually looks like (spoiler: it's not the model).*