Local Influenza Forecasts Outperform State-Level Forecasts in the United States

Most influenza forecasts in the United States are produced at the state level. This can miss what is actually happening in specific cities. This lack of granularity means a state might look stable while a metropolitan area faces a sudden surge in emergency room visits. A new study from researchers at the University of Texas at Austin and several other institutions demonstrates that moving from state-level to local-level forecasting improves accuracy.

The researchers report that local forecasts, generated for specific Health Service Areas (HSAs), outperformed state-based projections in 98.8% of tested regions at a one-week horizon. This is a significant gain. The authors find that local models achieved a 39.2% average reduction in the Mean Weighted Interval Score (MWIS). The MWIS is a metric used to evaluate how well a probabilistic forecast captures both the central tendency and the uncertainty of an event.

The masking effect of state-level aggregation

Public health responses, such as school closures or vaccine distribution, are typically implemented at sub-state levels like counties or cities. However, the current forecasting status quo relies heavily on state-level data. State-level averages act as a smoothing filter. This obscures "spatial heterogeneity," which refers to the significant differences in disease dynamics occurring between different regions.

As shown in [Figure 1A], local trends in influenza-related emergency department (ED) visits can deviate sharply from the statewide trend. These differences involve both timing and intensity. For instance, a local peak might occur weeks earlier or reach nearly triple the intensity of the state average. The authors note these discrepancies are particularly problematic during epidemic peaks. This is exactly when healthcare planners need reliable information to manage hospital surges. The study also finds an inverse relationship between population size and local-state agreement [Figure 1B]. Smaller HSAs tend to deviate more from their state's aggregate trend.

Scaling precision with gradient boosting

To address this gap, the authors developed a forecasting framework centered on Gradient Boosting Quantile Regression (GBQR). Unlike standard regression, which predicts a single value, GBQR estimates conditional quantiles. This allows the model to produce a full predictive distribution. This is essentially a range of possible outcomes with associated probabilities.

The methodology follows a structured pipeline: 1. Data Integration: The researchers used near-real-time data from the CDC’s National Syndromic Surveillance Program (NSSP). They focused on the percentage of ED visits attributable to influenza. Using percentages rather than raw counts provides an "intrinsically normalized" metric. This allows for fair comparisons between a small town and a massive city without complex population adjustments. 2. Model Training: The team employed the LightGBM framework to implement the GBQR. They trained the models retrospectively on data from three influenza seasons (2022–2025). 3. Ensemble Approach: To ensure stability, the models were ensembled across 100 random training subsamples. 4. Comparative Benchmarking: The local HSA-level forecasts were compared against two baselines. These were a "quantile baseline" (which produces flat, uninformative forecasts) and an automated ARIMA model (a classic statistical method for analyzing time-series data).

Superiority in local calibration and accuracy

The results favor local modeling across almost every dimension of accuracy. The authors report that HSA-level forecasts achieved better "coverage rates" than state forecasts. In forecasting, a coverage rate refers to how often the actual observed value falls within the model's predicted uncertainty interval. At a one-week horizon, the HSA-level models achieved a coverage rate of 0.91. State-level models lagged at 0.59. This indicates local models are better at "calibrating" their uncertainty. They provide a more realistic window of what might happen.

The performance advantage is most visible in the short term. As the forecast horizon increases, the relative advantage of local modeling begins to diminish. For example, the average reduction in MWIS drops from 39.2% at one week to 6.5% at four weeks. Despite this decay, local forecasts remain more accurate than state-level ones even at the four-week mark. Visualizing this in, the orange curves (HSA-level forecasts) track the red observed data points much more closely than the gray curves (state-level forecasts).

This is especially true during the volatile periods surrounding epidemic peaks.

Identifying the high-value targets

The study does not suggest that local forecasting is equally useful everywhere. Through a Generalized Linear Model (GLM), the authors identified associations between population structure and forecast improvement.

The data suggests that the added value of local forecasting is most pronounced in three specific scenarios: * Low Population Share: HSAs that represent a small fraction of their state's total population show larger improvements. * High Urbanization: The association between population ratio and forecast improvement is stronger in more urbanized HSAs [Figure 3A]. * State Fragmentation: States containing a higher number of Metropolitan Statistical Areas (MSAs) exhibit greater gains [Figure 3B]. MSAs are large, socially and economically integrated functional areas.

By projecting these findings onto a national map [Figure 3C], the authors suggest that resources for local forecasting capacity might be most effectively directed toward highly heterogeneous, multi-metro states like California and Texas.

Limitations and practical deployment

While the evidence for local forecasting is strong, the paper identifies several constraints. First, the analysis focused exclusively on HSAs with populations greater than 250,000. The authors explain that smaller jurisdictions often produce "noisy" surveillance signals due to low patient volumes. This can make them harder to model reliably. Consequently, the findings may not yet generalize to rural or very small communities.

Second, the researchers acknowledge that their comparative framework assumes state-level forecasts are used as proxies for local decision-making. In reality, local officials might attempt to manually adjust state forecasts using local data. This nuance is not captured in the study. Finally, maintaining hundreds of individual local models presents a higher computational and data-infrastructure cost than a single state-level model.

Is local forecasting ready for prime time? For large, metropolitan health systems, the answer appears to be yes. The ability to capture local peaks that state models miss is critical for surge management. Code for the implementation is reportedly available; see the paper for the canonical link. For practitioners in large, fragmented states, investing in HSA-level granularity offers a way to improve outbreak preparedness.

Figures from the paper