Nordic electricity price forecasting

A ten-model comparison for day-ahead Nord Pool electricity prices, ranked on a Pareto accuracy-versus-compute basis. The repository assembles the data, engineers a leakage-free feature matrix, runs an expanding-window backtest across the 60-minute and 15-minute market regimes, and reports which models sit on the accuracy-per-compute frontier.

Quick start

Requires Python 3.11+ and uv.

uv sync
uv run streamlit run app.py

The committed snapshot ships every matrix and result table, so the app and the headline ranking run offline with no API key. See Replication for the offline rebuild and the full from-scratch run.

Result

Mean across the hourly (three folds) and 15-minute (one fold) regimes over the 2021-01-01 to 2026-08-13 window; lower is better on every accuracy column:

model	family	CRPS	MAE	pinball (P50)	total compute
chronos2	foundation	9.61	15.45	7.73	3061.6 s
catboost	gbdt	9.64	15.50	7.75	1.8 s
lgbm	gbdt	9.72	15.37	7.68	0.5 s
xgboost	gbdt	10.63	15.70	7.85	1.8 s
nbeats	deep	14.39	18.40	9.20	38.6 s
timesfm	foundation	14.72	21.33	10.66	17.5 s
sarima	classical	14.99	21.69	10.85	166.6 s
tft	deep	20.12	30.20	15.10	640.6 s
lear	ml	26.29	26.29	13.15	2.5 s
ets	classical	33.40	52.81	26.41	147.8 s

The gradient-boosted trees (CatBoost, LightGBM, XGBoost) dominate the accuracy-per-compute frontier: they reach CRPS within 0.1–1.0 of Chronos-2 at roughly a thousandth of the compute. Chronos-2 edges the trees on mean CRPS by about 0.5%, but a Diebold–Mariano test on the three hourly folds does not find that edge significant, and the edge costs roughly three orders of magnitude more compute. Four models (SARIMA, TFT, LEAR, ETS) score below a seasonal-naive baseline, so they are worse than repeating last week's price. The full metric set (CRPS, MAE, the P10/P50/P90 pinball decomposition, the skill score, per-regime results, and the pairwise Diebold–Mariano tests) is in REPORT.md.

App

app.py is a single-page Streamlit explorer of the snapshot. It renders the headline ranking, the Pareto scatter, a per-model drill-down, and the feature selection results.

Overview and Pareto frontier	Ranked table and drill-down	Feature exploration

Run it with uv run streamlit run app.py.

Key findings

• Feature selection: forward selection finds cross-border flow to be the dominant exogenous group (it cuts CRPS by 1.68); weather and hydro add less, and carbon does not help. The efficient set is the base group plus cross-border, and a transfer check confirms it earns its cost on every full-features arm.
• Tuning: a secondary pass tunes the best model per family. None of the tuned configurations beat the pinned default on the hourly regime.
• Cross-zone: the frontier (the three trees) plus Chronos-2 generalise across SE1–SE4; Chronos-2 leads every zone.

Methodology

• Data: ENTSO-E (cross-border, hydro), Open-Meteo (weather), ECB (EUR/SEK), and EEX (EUA carbon), joined at a single assemble_data seam with a Parquet cache so a re-run does not re-fetch.
• Features: seven groups on a no-leakage as-of timing rule, because a model that sees a value published after its forecast time looks good in a backtest and fails in production. Boundary masking zeroes lags and rolling windows that cross a regime switch.
• Backtest: expanding walk-forward with yearly refit and a 7-day purge, which reproduces how a forecaster is actually used: train on the past, forecast the next day, never across a structural break.
• Roster: four families (classical, gradient-boosted, deep, foundation) and ten models, every arm pinned to seed 42 so the comparison is reproducible.

The full methodology, metric definitions with formulas, and the decision record are in REPORT.md and docs/adr/.

Project structure

• src/forecast_pipeline/: data ingestion, feature engineering, models, backtest, and scoring.
• notebooks/: exploratory analysis and the comparison and Pareto narrative.
• reports/: result tables from the full run.
• data/snapshot/: committed matrices and tables for offline reproduction.
• docs/adr/: the architecture decision record.
• app.py: the Streamlit explorer.

Replication

The snapshot holds every matrix and result table needed to reproduce the figures offline, with no API key.

uv sync
uv run python -c "from forecast_pipeline.snapshot import load_results, headline_ranking; \
print(headline_ranking(load_results()['results']).to_string(index=False))"

To rebuild the snapshot from the cached data window (no API key):

uv run python -m forecast_pipeline.snapshot --build

To run the full comparison from scratch, an ENTSO-E API key is required in .env:

set -a && source .env && set +a
uv run python -m forecast_pipeline.bakeoff --zones SE3 --start 2021-01-01 --end 2026-08-13

License

MIT. See LICENSE.

Full narrative

• REPORT.md: the complete benchmark report.
• notebooks/01_data_features_regimes.ipynb: data, features, regime detection, and feature selection.
• notebooks/02_backtest_comparison_pareto.ipynb: comparison, tuning, cross-zone, and the Pareto frontier.
• app.py: the Streamlit explorer.