Forecastability of Infectious Disease Time Series
INFECTIOUS disease forecastability varied by season, pathogen, and population size, shaping how well hospital admissions were predicted.
A new analysis suggests that some infectious disease time series may be intrinsically harder to forecast than others, offering one possible explanation for why model performance shifts across seasons, pathogens, and locations. Using a forecastability metric based on spectral entropy, the researchers evaluated how much structure or disorder was present in respiratory virus surveillance signals, then examined whether that metric aligned with forecast performance.
Infectious Disease Forecastability Varied by Season
Across California syndromic influenza hospital admissions, forecastability differed substantially from one respiratory virus season to another. At the state level, scores ranged from 19.5% in 2021–2022 to 41.6% in 2017–2018. Higher forecastability was positively associated with both cumulative burden and peak burden, with the strongest relationship seen for peak burden. The 2017–2018 season, which had the highest peak weekly admissions and cumulative burden, was also the most forecastable.
Some disrupted seasons stood apart. The periods spanning the 2009 H1N1 pandemic showed lower than expected forecastability for their burden, suggesting that atypical epidemic behavior may reduce the regularity that forecasting models can exploit.
Population Size and Forecast Performance
For U.S. state and national laboratory-confirmed COVID-19 and influenza hospital admissions, forecastability increased with population size. Forecast performance also generally improved as forecastability rose, particularly when scores were assessed on the natural log scale to reduce the influence of target size.
For ensemble models, each 1% increase in forecastability was associated with an approximate 0.46–0.86% decrease in mean absolute error relative error, alongside smaller weighted interval scores and improved scaled relative skill. By contrast, baseline model performance did not show the same pattern, suggesting that higher forecastability benefited ensemble forecasting rather than simply reflecting easier baseline predictions.
The 2022–2023 influenza season was a notable exception, with mean absolute error and weighted interval score failing to track significantly with forecastability. The analysis suggested that unusual influenza dynamics after the COVID-19 pandemic and limited historical continuity for newer targets may have contributed.
Why Forecastability Matters for Clinicians
These findings suggest infectious disease forecastability could help contextualize why forecast performance changes between respiratory virus seasons and across geographies. The analysis also raised practical questions for public health forecasting, especially because forecastability was highly sensitive to reporting frequency and temporal smoothing. Daily rolling data produced different scores than weekly sampled time series, indicating that data structure itself may affect how predictable a signal appears.
For clinicians and health systems, the work points to a more nuanced interpretation of outbreak forecasts. Some targets may not be poorly modeled so much as intrinsically less predictable.
Reference
White LA et al. Forecastability of infectious disease time series: are some seasons and pathogens intrinsically more difficult to forecast? PLoS Comput Biol. 2026;22(4):e1014175.
Featured Image: MarkoVS87 on Adobe Stock.
