GLEAM-AI: model and retrospective evaluation

The GLEAM-AI surrogate

GLEAM-AI is a neural surrogate of the Global Epidemic and Mobility model (GLEAM) — a stochastic, age-structured, metapopulation epidemic model in which local transmission within subpopulations is coupled to the movement of individuals between them. Running GLEAM at the scale required for calibration is computationally expensive, because each scenario is a large stochastic ensemble of simulations. GLEAM-AI eliminates this cost: it is a recurrent neural network that reproduces GLEAM's output for any scenario in the model's parameter space. Given a compact description of an epidemic scenario, it returns — from a single forward pass — the full distribution of state-level trajectories the simulator would have produced. For each of the 50 states and the District of Columbia it emits a Negative-Binomial predictive distribution for daily hospital admissions, so forecast uncertainty is represented explicitly rather than by re-running the simulator thousands of times; the weekly totals that FluSight scores are obtained by summing this output. GLEAM-AI is thousands of times faster than GLEAM on CPU, and more than 40,000 times faster on GPU.

The surrogate is trained once, offline, to match GLEAM configured to reproduce seasonal influenza across the United States. Its training set comprises 10,000 influenza scenarios drawn by a Sobol space-filling design over five epidemic parameters — transmissibility, seasonal forcing, epidemic timing, initial immunity, and the size of the initial seed — matched to a large precomputed library of GLEAM simulations. Because the surrogate maps scenario parameters to trajectories, observed surveillance plays no part in the network itself; it enters only at forecast time, through calibration.

Generating a forecast

At each weekly reference date, a GLEAM-AI forecast is produced in five steps:

1. Read the data as of the deadline. NHSN influenza-hospitalization admissions are taken as they stood at that week's Wednesday submission deadline, and the most recent, still-incomplete week is completed with a nowcast that uses only data available by then; later revisions are not used.
2. Propose a pool of epidemic scenarios, restricted to a realistic influenza effective-reproduction-number range so that calibration cannot draw on implausible, explosively-growing trajectories.
3. Evaluate every scenario through the surrogate and calibrate independently for each location by Approximate Bayesian Computation (ABC), keeping the scenarios whose surrogate trajectories are consistent with the observed data.
4. Form the predictive distribution from the retained scenarios and recalibrate its quantiles by conformal methods, bringing interval coverage toward its nominal level.
5. Emit the 23 FluSight quantiles for horizons 0–3 weeks ahead, subject to an as-of tail bound that caps any quantile at 1.5× the largest admissions ever observed for that location up to the forecast date.

The US national forecast is formed by aggregating the state-level forecasts.

Two-season design and out-of-sample controls

The retrospective uses two seasons with a strict separation of roles. The 2024-25 season is used for development: the model's settings — scenario priors, calibration tolerances, conformal parameters, the nowcast window, the tail-bound multiplier — are fixed on 2024-25 and then frozen. The 2025-26 season is held out, and the frozen configuration is applied to it once. Two caveats belong with this design, both stated plainly:

• At run time the evaluation is pseudo-prospective (next section): a forecast never uses any 2025-26 information that a real-time forecaster would not have had.

Pseudo-prospective (real-time) evaluation

The retrospective is pseudo-prospective: it reconstructs, week by week, only what a forecaster could have known in real time.

• As-of (vintage) data. Each forecast reads surveillance from the data vintage available at its Wednesday submission deadline — the values as known at that date, before later revision — and treats a week as observed only once it has ended. The backfill that later completed each recent week is not available to the forecast that had to anticipate it.
• Isolation of the scoring truth. The finalized, revised admissions used to score the forecasts are held in a separate data path that the forecasting code cannot access; the separation is enforced by an automated test. A forecast is therefore prevented from seeing the finalized value it will later be scored against.

The two GLEAM-AI forecasts shown

• GLEAM-AI — the calibrated surrogate described above.
• GLEAM-AI + TimesFM (hybrid) — an equal-weight (50/50) combination, at each quantile, of GLEAM-AI with TimesFM 2.0, a general-purpose time-series foundation model applied zero-shot (no influenza-specific fitting; the 500M-parameter checkpoint, pinned before the season, fed only the same as-of admissions history).

Evaluation

Forecasts are scored with the Weighted Interval Score (WIS) — a proper score for quantile forecasts — on a log(1 + count) scale, across horizons 0–3, the 52 scored locations (US national, the 50 states, and the District of Columbia), and every scored reference date of the season. The headline metric is the relative WIS (rWIS): a pairwise-tournament score, following CDC FluSight convention, normalized so that the FluSight baseline model equals 1.00 — below 1 is better than the baseline, and lower is better. A model is ranked only if it has forecasts for at least 75% of the season's cells, and 50% / 95% interval coverage is reported alongside skill. All scores use a single pinned, finalized surveillance dataset, so every model is judged against identical ground truth; applied to the official FluSight field for 2025-26, the procedure reproduces the CDC's published season-summary results.

Interpretation

These forecasts were not submitted to FluSight during 2025-26; they were produced afterwards under the controls above. The results are a single held-out season and should be read as such, together with the design-hindsight caveat noted earlier. On the charts, the Observed line is the finalized data; the optional as-of line shows what was known at each forecast date — the difference between them is the later backfill that the real-time forecasts had to anticipate.