National and Regional Influenza-Like-Illness Forecasts for the USA

Michal Ben-Nun; Pete Riley; James Turtle; David P. Bacon; Steven Riley

Broadly, for each different forecast target and each forecast lead-time, there has been a 513 gradual progression over time such that objective forecasts become more accurate than 514 subjective forecasts. We note also that although we describe the subjective process as it 515 was conducted, we also provide a thorough retrospective assessment of the predictive 516 performance of each model variant. 517 We may refine our ensemble approach for future iterations of the competition. It seems 518 clear that the coupled models produce more accurate forecasts than the uncoupled 519 models for most targets, so we would consider an ensemble only of the coupled model 520 variants. We will also consider weighted ensembles of models and attempt to find 521 optimal weights by studying all prior years. Also, data were often updated after being 522 reported and we did not include an explicit reporting model in our inferential framework 523 (also sometimes referred to as a backfill model). Rather, we used knowledge of past The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/309021 doi: bioRxiv preprint adjustments to data during our discussions and eventual subjective choice of models. 525 We aim to include a formal reporting model in future versions of our framework. when evaluated using the historical data. This prospective study supports recent 530 retrospective results suggesting that influenza forecasts can be more accurate if they 531 explicitly represent spatial structure [33, 34] . Given that the model structure we used to 532 represent space was relatively coarse [24] , further work is warranted to test how forecast 533 accuracy at finer spatial scales can be improved by models that include iteratively finer 534 spatial resolution.