Realism VS Accuracy: Event Sequence Forecasting from a Generative Modeling Perspective

Realism VS Accuracy. Event Sequence Forecasting from a Generative Modeling Perspective. By Dmitry Osin and seven others, from the Applied AI Institute, ITMO University, and two other institutions. Published on the eighth of October. Its code repository is worked out: about one hundred and thirty source files, with a readme and a licence.

An event sequence is a chain of random events, each one described by a mix of numerical and categorical features. Forecasting them comes up in finance, healthcare, business process intelligence and elsewhere. The authors treat forecasting several steps ahead as a conditional generative modeling problem, and that framing changes what a good forecast is. It is not only about being close to what happens. It is also about whether the forecasts look like real sequences.

Forecasts are therefore scored on two separate quality aspects. The first is prediction accuracy, the familiar question of how close a forecast lands to what actually happened. The second is distributional realism: whether the forecasts, taken together, have the character of genuine event sequences. Treating forecasting as generation is what makes the second aspect a natural thing to ask for.

To measure both, the authors assemble a benchmark with four parts. There are metrics covering each quality aspect. There are strong statistical baselines. There are recent models adapted from adjacent fields. And there are curated real-world datasets to run them all on.

Careful evaluation shows that no method excels at both accuracy and realism. The authors read this as a fundamental trade-off between the two. They then introduce Wasserstein BaryBooster, an algorithm that aggregates samples from a diffusion model strong on realism into a state-of-the-art predictor. The result demonstrates that distributional realism can be converted into prediction accuracy.

The paper ends up arguing both sides of its own title. No method wins on realism and accuracy at once, yet realism is not wasted: a model that produces realistic samples can be turned into a source of better predictions. The code is publicly available.