Wasserstein BaryBooster Turns Diffusion Realism into Event Forecast Accuracy

The authors of preprint arXiv:2610.12586 built a benchmark for event-sequence forecasting and evaluated models on two axes: prediction accuracy and the distributional realism of the generated trajectories. Their main conclusion is blunt: no method wins on both axes at once. The team then showed a workaround. The Wasserstein BaryBooster algorithm aggregates samples from a diffusion model that is strong on realism and turns them into a state-of-the-art predictor. The code is publicly available.

What is being forecast here

An event sequence is a chain of random events, each with heterogeneous features, both numerical and categorical. A typical example is bank transactions with an amount, a merchant category, and a timestamp. The same structures appear in medical histories and business-process logs. The authors work in a multi-horizon setting: the model must predict not just the next event but several steps ahead.

They frame the task as conditional generative modeling. The model receives a history and samples possible continuations. This changes how forecasts must be evaluated. It is not enough to know how close a forecast is to the actual trajectory. You also need to check whether the generated chains look statistically like real ones: event frequency, intervals between events, and feature combinations.

The benchmark and the trade-off

Practitioners know this gap well. A model that tries to minimize error tends to produce averaged, "smoothed" trajectories. They score decently on accuracy, but such chains almost never occur in real data. A generative model, by contrast, reproduces the real distribution, but any individual trajectory may stray from the ground truth. The authors tested this on a single suite of tests.

The benchmark includes:

• metrics for both aspects of quality: accuracy and distributional realism;
• strong statistical baselines;
• recent models from adjacent fields, adapted to event sequences;
• selected real-world datasets.

After careful evaluation, a fundamental trade-off emerged. The most accurate methods lose on realism, and the realistic ones lose on accuracy. For industry this is an uncomfortable signal: choosing a model by a single metric can quietly break the other one.

How BaryBooster gets around the trade-off

The second part of the work is constructive. In the experiments, the diffusion model proved strong specifically on realism. Wasserstein BaryBooster takes a set of its samples and aggregates them into a single forecast. The name points to a barycenter in the Wasserstein metric: the "mean" of a set of objects that accounts for the geometry of the space, rather than a simple arithmetic average. The abstract does not disclose the details, so the specifics are best checked in the paper and the repository.

The key takeaway is that realism can be converted into accuracy. A model that describes the distribution of future events well can serve as the source for a predictor that outperforms specialized methods.

What to do with this in practice

Teams in fintech, healthcare, and process mining that forecast event chains get a ready-made tool for testing their models. The benchmark and code are available in the xaosina/event-sequence-forecasting repository. It makes sense to run your own solutions there and see whether you have paid for accuracy with realism.

The choice of metric should depend on the task. For scoring and point forecasts, accuracy matters. For simulating customer behavior, stress tests, and synthetic data, realism is critical. If you need both, a "diffusion plus sample aggregation" scheme looks like the first candidate to test.

There are caveats. We are relying on the abstract, which contains no numerical results. By the method's logic, aggregating many samples costs more than a single direct forecast. Before moving the approach into production, measure the accuracy gain on your own data and weigh it against inference cost.