Event-sequence forecasts can't be both accurate and realistic: pooling a diffusion model's samples into one beats the dedicated predictor

Event-sequence forecasting is recast as a conditional generation problem, and forecasts are judged on both prediction accuracy and distributional realism. Earlier evaluations of event sequences focused almost exclusively on prediction accuracy and overlooked realism. Across three card-transaction datasets, methods with high prediction accuracy show low distributional realism and vice versa, and no method excels at both.

Aggregating several samples from a diffusion model converts realism back into accuracy. The diffusion model on its own is not the most accurate forecaster, though it is the most realistic. After Wasserstein BaryBooster aggregates LDM samples, TMS reaches 0.495, 0.436 and 0.522 on Alphabattle, Gender and Age, above DEF's 0.424, 0.413 and 0.508, while cDFS stays at 0.177, 0.256 and 0.224.

The procedure treats one forecast as the barycenter of many samples. Previously the many samples of a generative model were treated only as a distribution, never collapsed into a single forecast. It alternates alignment and aggregation under TMS, aligning with Hungarian matching and taking medians for time and numerical features and modes for categorical ones, with no retraining.

The benchmark targets long-horizon event-sequence forecasting, where events carry timestamps and a mix of numerical and categorical features, and the object of evaluation is the future continuation given a history. All three main datasets are card-transaction records, and each event includes at least a timestamp, a transaction amount and a merchant category code. The authors additionally stress-test on two non-transactional point-process corpora, Retweet and Amazon, which have far fewer and much shorter sequences and carry only a timestamp and a categorical mark per event.

The authors state that the empirical study uses a relatively small number of datasets and focuses narrowly on card transactions, and that practical constraints on data quality and schema compatibility limited the final pool. Wasserstein BaryBooster introduces substantial computational overhead, since it requires generating multiple candidate trajectories and solving an iterative barycentric aggregation procedure at inference time, which may be prohibitive in latency-sensitive or resource-constrained deployments; the overhead is an inference-time choice, however, and 5 samples already retain about 60 percent of the TMS gain at one sixth of the sampling cost, with samplings parallelizing across GPUs.