Datasets:

---

Real-TSF

/

TIME-Output

Like 1

Follow

Real-TSF 9

Evaluation results of ci-proto on all 98 TIME tasks, for the leaderboard.

ci-proto is a time-series foundation model from Continuous Intelligence, with 167M parameters (code, weights).

• Scores with scripts/compute_local_leaderboard.py, normalised by Seasonal Naive: MASE 0.634, CRPS 0.532.
• Protocol: the targets of a multivariate series are forecast jointly (to_univariate=False), and the single-target series of a window together, as a cohort; the context is the history before each window, up to 16,384 steps; the model's nine quantiles are TIME's deciles; one forecast per window. Zero-shot: no TIME data in the training data.
• Code: experiments/ci_proto.py and scripts/run_ci_proto.sh in https://github.com/zqiao11/TIME/pull/19, run with continuousintelligence==0.1.1 and the weights at revision ee3031351a151e3f6febf8c62b53e6ea1008bef6 on one NVIDIA H100 (bfloat16).
• Files: results/ci-proto/<dataset>/<freq>/<term>/ with config.json, metrics.npz and predictions.npz for each task.