Diachronic Sample Integration: Robust Tail-Risk Estimation with Generative Models

This paper introduces Diachronic Sample Integration (DSI), a test-time inference method that aggregates generated samples from multiple checkpoints along a stochastic training trajectory. Instead of trusting a single endpoint, DSI constructs a checkpoint-mixture distribution intended to average localized tail fluctuations. The authors provide a finite-budget bias-variance analysis and evaluate the method on multivariate synthetic processes and high-frequency trading data. According to the abstract, DSI reduces tail-estimation error under fixed simulation budgets and outperforms single-checkpoint, standard diffusion, and tail-aware baselines without changing the generative training objective.