cascade-eval-pool / README.md
tensorlink-dev's picture
refresh card (as_of=2026-08-01)
225ec5d verified
|
Raw
History Blame
1.79 kB
metadata
license: cc-by-4.0
pretty_name: cascade held-out eval pool (24h-lagged reveal)
tags:
  - time-series
  - forecasting
  - benchmark

cascade eval pool — lagged public reveal

Daily snapshots of the held-out evaluation pool used by the cascade subnet, built from the tsbench-forge live catalog of real public time series across the 7 GIFT-Eval domains.

Each snapshot is revealed 24h AFTER it was used for scoring (as_of is a closed UTC day). Because the pool rotates daily, a released snapshot is always for a round that is already scored — it lets anyone reproduce the leaderboard without exposing the round currently in play.

Layout

snapshots/<YYYY-MM-DD>/          one folder per revealed day (as_of)
  <series_id>.npy                float32, freshest context_length + horizon points
  metadata.json                  {series_id: {freq, seasonal_period, domain, source}}
  provenance.json                build config (context_length, horizon, builder_version)

Scoring is identity-agnostic: MASE uses seasonal_period, CRPS/WQL use the values + a model's quantiles, and the KOTH cluster-bootstrap groups windows by source. Re-cut windows of context_length + horizon, forecast the horizon, and the metrics reproduce byte-for-byte.

Latest snapshot — 2026-08-01

  • series: 3504
  • domains: transport=2545, energy=367, healthcare=227, nature=193, econ_fin=74, sales=66, web_cloudops=32
  • cadences: min=2566, D=337, 30min=258, 15min=134, 5min=122, H=53, 10min=15, 6min=10, 240S=6, 15S=1, 10S=1, 1S=1

This is an evaluation set, not training data. Publishing it to train on would contaminate the benchmark it exists to measure.