---
pretty_name: "Predictions \u2014 COVID Tracking Project \u2014 US states daily (archived)"
license: other
tags:
- epi-eval
- predictions
- forecast-evaluation
- companion-of-covid-tracking-project
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*.parquet
---

# Predictions for COVID Tracking Project — US states daily (archived)

Community-submitted forecasts targeting [`EPI-Eval/covid-tracking-project`](https://huggingface.co/datasets/EPI-Eval/covid-tracking-project).
Each row is one quantile (or point) forecast for one target date — see the
schema below.

This repo accumulates accepted submissions from many forecasters. New
predictions arrive as community pull requests opened from the [EPI-Eval
dashboard](https://github.com/ChrisHarig/apart-forecasting-tool); a
maintainer reviews each PR before merging.

## Schema (v1)

| column | type | notes |
| --- | --- | --- |
| `target_date` | string (`YYYY-MM-DD`) | The date being forecast |
| `target_dataset` | string | Always `covid-tracking-project` |
| `target_column` | string | Truth column being forecast (see below) |
| `submitter` | string | Forecaster name or HF username |
| `model_name` | string | Identifier for the model run |
| `description` | string | Free-form notes on the model |
| `quantile` | float (nullable) | In `[0, 1]`. `null` = point estimate |
| `value` | float | Forecast value (in the truth column's units) |
| `submitted_at` | string (ISO 8601) | UTC submission timestamp |
| _(pass-through dims)_ | string | Categorical dims from the source CSV |

Long format: one row per `(target_date, [dim values…], quantile)`. A
forecaster providing the median plus 50%/80%/95% intervals emits 7 rows per
date (one point + 6 quantiles). Multiple submissions from the same forecaster
land as separate parquet files under `data/`.

## Forecast targets

Truth columns from `EPI-Eval/covid-tracking-project` you can forecast:

- `positive` (cases (cumulative))
- `death` (deaths (cumulative))
- `hospitalizedCurrently` (patients)
- `inIcuCurrently` (patients)
- `onVentilatorCurrently` (patients)
- `totalTestResults` (tests (cumulative))
- `positiveIncrease` (cases/day)
- `deathIncrease` (deaths/day)
- `hospitalizedIncrease` (admissions/day)
- `totalTestResultsIncrease` (tests/day)

## Submitting

The dashboard at [apart-forecasting-tool](https://github.com/ChrisHarig/apart-forecasting-tool)
handles the full submission flow: drag-drop a CSV, pick this dataset as the
"Compare to" target, review your scores against the truth, and click "Submit
to HuggingFace." The dashboard serializes your CSV into the schema above and
opens a community PR here.

## Notes

- Predictions whose `target_date` falls outside the truth dataset's coverage
  (forecast horizon) are still accepted. Comparison metrics on the dashboard
  compute only on the dates where truth is available.
- A submission's `submitter` value is its only identity claim — there's no
  signed authentication. Reviewers should sanity-check unfamiliar submitters.

_Initialized by `upload_pipeline.core.bootstrap_predictions_repos`._