apararti/repro-beyond-model-ranking-predictability-aligned-evaluation-for-time-series-forecasting-results

Beyond Model Ranking reproduction results

This dataset repository contains scripts, tests, raw tables, figures, and intermediate predictions for an independent reproduction of Beyond Model Ranking: Predictability-Aligned Evaluation for Time Series Forecasting.

The reproduction follows the algorithms in the authors official repository and uses the four public ETT datasets from the official ETT repository. The original ETT CSVs are not duplicated here; rerun commands fetch them from that source.

Scope

• Claim 1: FFT/Welch runtime benchmark through N=32768.
• Claim 2: jointly Gaussian consistency check plus a boundary mean-shift counterexample.
• Claim 3: four ETT datasets, history/horizon 336, five lightweight forecasters. This is not a full reproduction of the paper trained SOTA matrix.
• Claim 4: every stride-1 ETTh1 test window, all seven channels.
• Claim 5: frequency-resolved LUR on ETTh1 channel 1 with three real lightweight forecasts and two diagnostic controls.

Hugging Face GPU Jobs were attempted but rejected before scheduling with HTTP 402 because the account had insufficient prepaid Jobs credit. All included results are local CPU results.

Rerun

uv sync
mkdir -p data
for name in ETTh1 ETTh2 ETTm1 ETTm2; do
  curl -L "https://raw.githubusercontent.com/zhouhaoyi/ETDataset/main/ETT-small/${name}.csv" -o "data/${name}.csv"
done
uv run pytest -q
uv run python run_reproduction.py --claim all