Conditional diffusion for uncertainty-aware dynamic microsimulation: multivariate trajectory inpainting, forecasting, and scenario generation

Dynamic microsimulation requires generating coherent multivariate micro-trajectories over time across multiple outcomes (e.g., employment, income, hours), while handling panel gaps and propagating uncertainty into downstream indicators. Common approaches—transition models, chained regressions, and hot-deck imputation—often yield a single deterministic completion and can struggle to preserve high-dimensional joint structure, especially under block nonresponse.

We propose a conditional diffusion approach for dynamic microsimulation in which each unit is represented as a multivariate monthly trajectory, conditioned on static covariates and an observation mask. The method borrows the same core mechanism that made diffusion models widely known through text-to-image systems such as Stable Diffusion: a sample is generated by starting from noise and repeatedly denoising until realistic structure emerges—images are simply a particularly dramatic, high-dimensional domain where this iterative reversal is easy to appreciate. Here we apply the same denoising principle to structured longitudinal microdata: a one-dimensional residual convolutional denoiser with timestep and month positional embeddings learns to reverse a gradual Gaussian corruption process on trajectories, so observed months are pinned while missing months (or, via one-sided masking, future months) are generated as multiple plausible, jointly coherent completions consistent with the observed history and covariates.

We demonstrate the method on a large household panel dataset as a case study (SIPP). We evaluate the model under both random missingness and contiguous block gaps, and assess not only point error but dynamic and joint realism, including unemployment spell-length distributions, employment transition matrices, and income distributions conditional on employment status. A key benefit is multiple imputation: repeated sampling yields a distribution over plausible completions and uncertainty bands for downstream statistics, allowing the model to correctly express uncertainty instead of returning a single deterministic trajectory. The width of these bands reflects how identifiable missing months are given observed history and covariates; poor empirical coverage provides a diagnostic for missing predictors or miscalibration. We further demonstrate one-sided masking as a forecasting/nowcasting use case, and scenario-style constrained sampling for stress-testing counterfactual assumptions (e.g., income floors or top-ups), while noting that causal policy inference requires additional identification assumptions.

Overall, conditional diffusion offers a flexible, uncertainty-aware generative layer for microsimulation that can preserve multivariate temporal structure and support robust uncertainty propagation.