Tapestry

This is the project with all my wishlist for infectious disease modeling: Flusion's idea of learning across surveillance sources, InfluPaint's modeling ideas and infrastructure, CRPS fitting like that very smart DeepMind paper, and a way to bring it all together. The project builds on InfluPaint's Slurm job manager and simulation storage, and EpiBenchmark's scoring.

The idea is simple: give a model the recent history of several pathogens and ask it to generate possible futures. Learn across signals, track what is observed, and fit the predictive distribution. The current models use MLP encoders, convolutions, and transformer-style attention to exchange information across locations and targets. The architecture notes explain the contributions from Flusion, InfluPaint, and DeepMind's Functional Generative Networks paper, and their adaptations here.

Unicorns exist

A unicorn is a model formulation that beats the ensemble across three pathogens, on both admissions and ED visits — six targets — across the past three seasons. It was unclear whether such a model existed, given how different the seasons are. These retrospective experiments found candidates. Evaluation uses season cross-validation: train on two seasons and forecast the third, rotating through 2023–24, 2024–25, and 2025–26.

These models see finalized data with revisions unavailable to the hub ensembles at forecast time. Whether the gains carry over to vintaged data remains to be established. Ensemble comparisons cover flu admissions in 2023–24, flu and COVID admissions in 2024–25, and all six targets in 2025–26. The unicorn finding applies to this available support, not a complete six-by-three grid. These same seasons guided architecture selection, so this is exploratory cross-validation, not an untouched final evaluation.

Results — Fan plots. Green shows the new models; grey shows the ensemble. Measured interval coverage remains below nominal levels.

Across B0.1, 36 of 172 configurations beat the ensemble on the combined score, with a best score of 0.883 (1 is ensemble parity; lower is better). That count is not a count of unicorns winning every target/season comparison. The top five configurations fit separate models by target or pathogen while receiving all six input histories.

The B1 overnight results compare direct forecasts, supplied-final flags, auxiliary nowcasts, and two-stage forecasts under vintage inputs and artificial masking, with rankings and matched forecast fan plots. The page also reports the completed 300-epoch follow-up, with matched-seed comparisons and 100-vs-300-epoch fan plots.

B0: six channels, no revision nowcasting

B0 is the finalized-data experiment. The channels are:

	Influenza	COVID-19	RSV
NHSN hospital admissions	Channel 1 · counts	Channel 2 · counts	Channel 3 · counts
NSSP ED visits	Channel 4 · proportion	Channel 5 · proportion	Channel 6 · proportion

Every channel has an observation mask. A missing value is different from an observed zero. The panel contains 50 states, DC, and the native US series; US is forecast directly rather than constructed by adding state predictions.

flowchart TD
    A["NHSN finalized admissions<br/>Flu · COVID-19 · RSV"] --> C["Shared source selection<br/>Weekly dates · native geography · units"]
    B["NSSP latest frozen ED snapshot<br/>Flu · COVID-19 · RSV"] --> C
    C --> D["Saved B0 panel<br/>week × 6 channels × value/mask × 52 locations"]
    D --> E["History window<br/>8, 12, or 26 weeks · all six channels"]
    D --> F["Future labels + separate supervision masks<br/>Four weeks after the context end"]
    E --> G["Fold-fitted transforms and scales<br/>Fit on training seasons only"]
    G --> H["Conditional sample generator"]
    F --> I["Fair CRPS in native units<br/>Training-side scaling and target/location weights"]
    H --> I
    H --> J["Sampled futures → 23 quantiles<br/>EpiBenchmark WIS vs hub ensembles"]

Admissions are stored as counts; ED percentages are converted to proportions. Admission-rate transforms and normalization are fitted inside each training partition. B0 treats the frozen latest NSSP values as retrospective truth; they are not guaranteed immutable. The input mask handles existing gaps, but this experiment does not establish performance with arbitrary missing sources or historical reporting delays. See the data contract.

Architectures tried

The common model is a conditional sample generator. A history encoder describes what is happening; random noise modulates a decoder to produce a possible future. Repeating this gives a forecast distribution. Fair CRPS rewards samples that are both accurate and appropriately spread, in admission counts or ED proportions. Forecast quantiles are then evaluated with WIS.

flowchart TD
    A["Local six-channel history + masks<br/>Calendar · population · optional recent dynamics"] --> B["Temporal encoder<br/>MLP / convolution / multiscale convolution"]
    B --> C["Context exchange<br/>None / spatial / pathogen / target / joint attention"]
    C --> D["Residual sample decoder<br/>Shared, pathogen-specific, or target-specific heads"]
    Z["Fresh random draw per member<br/>Global noise; optional local noise"] --> D
    D --> E["Four future weeks × requested targets × locations<br/>Inverse transforms → native-unit samples"]
    E --> F["Predictive intervals and quantiles"]

Part	Tried in B0.1	Findings so far
History encoder	Flattened-history MLP, temporal convolution, multiscale convolution	MLP and ordinary convolution improved the tested matched references. More complexity did not consistently help.
Information exchange	Local only; shared spatial, pathogen-specific, target-specific, or joint location–target attention	These are transformer-style attention blocks over encoded histories, not a separate temporal-transformer encoder. Joint attention can see 52 × 6 = 312 tokens; its benefit depends on the recipe.
Sharing	Shared heads, three pathogen heads, six target heads; also fully independent fits by pathogen or target	The five leading configurations use independent fits. Each component still sees all six local input channels. Sharing inputs and sharing model weights are different choices.
Uncertainty	Noise-modulated residual decoders, global/local noise, stochastic trend decoder	The trend decoder worsened all four matched comparisons and has been removed from the active grid. Historical results retain it.

The completed experiment crossed 172 configurations and three random seeds, with three held-out-season folds per seed. The B0.1 specification has the exact recipes; the results have the comparisons. These findings are about the tested combinations, not a general ranking of MLPs, convolutions, and transformers.

From prediction to nowcasting

The next steps are to evaluate masking with arbitrary source availability, add covariates, and establish nowcasting performance. The B1 implementation has the Wednesday-vintage data and correction/forecast path; the unicorn results above are still B0 results. Nowcasting here means estimating the eventual values of recently completed weeks whose reports are missing or provisional.

flowchart TD
    A["Finalized older history<br/>Recent Wednesday reports or flagged finals"] --> B["Six-channel history<br/>Availability and known-final flags"]
    B --> C["B1 history encoder"]
    C --> D["Correct unknown recent values<br/>Pass through visible known finals"]
    D --> E["Condition on each sampled correction<br/>Generate the next four weeks"]
    Z["Random member draws"] --> D
    Z --> E
    T["Separately pinned reference truth<br/>Recent and future labels + supervision masks"] -. "training only" .-> D
    T -. "training only" .-> E
    E --> F["Forecast samples → hub quantiles → evaluation"]
    G["Later: additional sources and covariates"] -.-> B

The key distinction is event time versus release time: what happened last week may differ from what was reported by Wednesday. B1 uses finalized older history and keeps eligible Wednesday reports for the two recent weeks. Missing recent reports receive flagged reference finals, which bypass nowcasting and are excluded from its loss and scores. This is retrospective conditioning on later information, not evidence of operational Wednesday skill.

The goal is ensemble-matching performance across all six hub targets. Reporting and revision timing remain central to that evaluation. An ensemble of the three cross-validated models is one option for the FluSight submission; the mixture still needs to be evaluated.

Working with Tapestry

Start with Getting started, the training workflow, or the data explorer. The architecture notes explain the model and masks; the B1 page documents the current vintage-aware work.

Documentation note · September 16, 2026: the “unicorn” terminology comes from the project research update; numerical summaries come from the saved B0.1 report. No scores were recomputed for this update. “Across three seasons” means the available target/season evaluation support, not a complete six-by-three grid. Diagrams distinguish evaluated B0, implemented B1, and later source/covariate extensions.