Datasets:
File size: 2,240 Bytes
5f65667 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 | ---
license: other
task_categories:
- time-series-forecasting
tags:
- forecasting
- benchmark
- time-series
- llm-evaluation
pretty_name: Frontier Forecasting Arena (FFA) v0.1
size_categories:
- n<1K
---
# Frontier Forecasting Arena (FFA) — v0.1 (items only)
Anonymized, **answer-free** evaluation items for the Frontier Forecasting Arena, a
historical-replay benchmark that measures forecasting skill with proper scoring
rules and is designed so that **memorization is measurable rather than assumed
away**.
This dataset contains the **model-facing items only** — no ground truth. Skill is
verified by the maintainer against held-out truth. That split is the whole point:
no real date, absolute level, station, or year ships to a model, so answers can't
be looked up or memorized.
## Contents
| File | Arm | Items | Unit |
|---|---|---|---|
| `energy/items.jsonl` | Energy load | 50 | 168h history + 24h future drivers → 24×(q10,q50,q90) |
| `synthetic/items.jsonl` | Synthetic (DGP) | 50 | contamination-proof floor |
| `kalshi/items.jsonl` | Kalshi weather | 50 | settled-market snapshot ladder |
| `crypto/items.jsonl` | Crypto | 50 | hourly panel → forward window |
| `stock/items.jsonl` | Stock | 50 | daily panel → forward window |
Each line is one JSON item with a compact `columns` + packed-row encoding.
## How to run it
Use the FFA CLI (code: https://github.com/steves-brain/ffa-benchmark):
```bash
huggingface-cli download userr99/ffa-v0.1 --repo-type dataset --local-dir ffa-v0.1
ffa predict --dataset ffa-v0.1 --out submissions/<your-name> \
--provider openai-compatible --base-url http://localhost:11434/v1 \
--model <your-model> --harness Ollama --dataset-version v0.1
```
`ffa predict` never sees the answers. Send the resulting submission back (see
`CONTRIBUTING-SUBMISSIONS.md` in the GitHub repo) to be scored against private
truth and added to the leaderboard.
## Scoring spine
CRPS for the continuous arms, Brier/log-loss for binary. Never a single blended
skill number — skill is a per-arm vector; only calibration aggregates across arms.
## License
See the repository for terms. Built from free public data (EIA, NOAA/NWS, Kalshi,
public crypto/stock markets), anonymized.
|