rebase-energy/emflow-submission-example

emflow submission — example predictor

A reference submission for the emflow swedish-temperatures:ar benchmark. This is what a model submission should look like: a single submission.py that defines an emflow.Predictor and exposes it as a module-level model.

📌 For interns: copy submission.py, rename ExamplePredictor, and replace the train / predict bodies with your own model. Keep the same contract.

The contract

submission.py just needs a module-level emflow.Predictor instance:

model = MyModel()          # a FRESH, UNTRAINED model

(You can instead expose a get_model() -> emflow.Predictor factory — the verifier accepts either form.)

Your emflow.Predictor subclass implements:

• train(self, train_df) — fit on the wide temperature DataFrame (hourly UTC index, one column per station). Use only what's passed.
• predict(self, input_df) — return a DataFrame/Series whose entry at the last timestamp of input_df is your 1-step-ahead forecast for that hour. Use only past lags; the value at the last timestamp is intentionally NaN.

How it's scored (leak-proof)

The evaluator trains your model itself on the official pre-2026 split, then scores it with strict walk-forward on the held-out 2026 hours: to forecast hour t, the model sees only the series strictly before t. You never see the test targets and cannot train on them, so the score can't be inflated by leakage.

This example model — a least-squares AR on the last 3 hours — scores MAE ≈ 0.29 °C, comfortably beating the persistence baseline (≈ 0.31 °C).

Verify it

# from a clone of emflow, with the submissions extra installed
uv sync --extra submissions

# straight from this repo on the Hub (pin a commit for reproducibility)
python scripts/verify_submission.py hf://rebase-energy/emflow-submission-example/submission.py \
    --revision <commit-sha>

Or locally: drop submission.py into submissions/ and run python scripts/verify_submission.py submissions/submission.py.