Live results of general-purpose agents that forecast real-world time series with information they find themselves, next to time-series foundation models (TSFMs) given human-curated inputs and the task's official or expert forecast. Every forecast is submitted before its outcome exists and scored on the first published value. The website shows these results from its own copy of their summaries.

The tasks (ALE Task format) are in agentic-forecasting-tasks; the pipeline that runs them is agentic-forecasting.

Layout

runs/<task>/
├── truth.csv                          first published value per date
├── <run day>/
│   ├── live/<system>.json             live forecasts: agents, TSFMs, expert forecast
│   ├── mock/<system>.json             agents replaying the day offline
│   ├── x/                             X_d: the curated input the TSFMs received
│   ├── <arm>-task/, <arm>-first-task/ the Task a continual arm ran: the live Task
│   │                                  with that arm's paragraphs in the instruction
│   └── ale/<run>-<agent>/<episode>/   ALE episode: trajectory.json, transport and
│                                      execution traces, result.json, lock.json;
│                                      <run> is live, mock, continual or continual-tsfm
└── logs/<date>.log                    the pipeline's log of each day

A forecast file holds run_date, one entry per target date with nine quantiles (or a point for the expert forecast), and its provenance: harness and model, timing, model calls, tokens and ALE's list-price cost estimate where the model has a listed price.

Protocol

On each run day, agents receive the instruction, the run date and target dates, and the open internet; nothing else. TSFMs receive X_d, the target history and covariates the task's author chose. Each agent also runs in two continual arms, each a chain that resumes the agent's sandbox and session from its previous run day: continual is asked to learn from its earlier submissions, and continual-tsfm is also told to run the TSFMs (system ids end in -continual and -continual-tsfm). Each run day also becomes an offline replay built from the web materials the live agents saved; replays run on demand. Up to and including 2026-10-23, live agents also received X_d.

The agents' saved web pages are not included here; they stay with the Task assets. Agent trajectories record the commands agents ran and excerpts of the pages they read.

Sources

Target and covariate data come from their publishers, for example ERCOT public reports and Open-Meteo (ERA5 reanalysis and ECMWF IFS forecasts, CC BY 4.0). Each X_d carries a MANIFEST.json with its sources and file hashes.

Downloads last month: 3,210

Total file size:

300 MB