Effect of predictive-distribution flexibility on epidemic forecasting

Determine whether replacing GENIE’s constrained parametric predictive distributions with more flexible non-parametric distributions, such as normalising flows, improves predictive performance for epidemic forecasting.

Background

GENIE models hospitalisations, deaths, infections, and the effective reproduction number using a specified parametric family: negative-binomial distributions for count outcomes and a log-normal distribution for the reproduction number. The factorisation also assumes conditional independence among the burdens and MSOAs, restricting the joint predictive distribution.

The authors identify an unresolved question about whether forecasting performance is limited primarily by this parametric choice or by the representations learned by the neural network. They point to normalising flows as a more flexible alternative capable of representing multimodal, skewed, and heavy-tailed distributions, but do not determine whether that flexibility benefits epidemic forecasting.

References

It remains to answer: to what extent is predictive performance limited by the choice of parametric family rather than by the representation learned by the neural network? One specific limitation here is that in equation~\ref{eq:distribution_q}, we treat each burden's predictive distributions as independent conditional on the context which is a restriction on the joint predictive distribution. The limited flexibility of the parametric family may constrain the ability of the model to represent the conditional marginal distributions. Non-parametric models, such as normalising flows, learn transformations of a simple base distribution and can represent substantially more complex distributions, including arbitrarily multimodal, skewed and heavy-tailed densities. Whether this additional flexibility translates into improved predictive performance for epidemic forecasting remains an unanswered empirical question.

— GENIE: Generative Neural Inference for Epidemics  (2608.20253 - Guzmán-Rincón et al., 20 Aug 2026) in Discussion, final paragraph under “limitations of this study”