Warehouse ML

The same modelled data, run through Snowflake's in-database ML — then compared honestly against the Python models. Two questions: can the warehouse forecast as well as Python? and can it spot unusual market days on its own?

1. Snowflake ML vs Python — same split, same actuals

Both engines trained on identical history (2020-06-01 → 2026-05-24) and predicted the same held-out 7 days, scored against the same actuals. Lower MAPE is better.

The result is a split decision, which is the interesting part. Snowflake ML has the better average (3.34% vs AutoETS 4.01%) and is markedly stronger on the small, volatile regions — in South Australia it more than halves the error. Python's AutoETS still wins the two big stable series, NSW and Queensland. Every model beats the seasonal-naive baseline.

Caveat, stated plainly: this is a single 7-day holdout (35 predictions) — enough to be suggestive, not enough to be conclusive. The rolling-origin backtest on the forecast page is the more rigorous test of the Python models.

2. Anomaly detection — unusual price days

SNOWFLAKE.ML.ANOMALY_DETECTION learned a per-region price baseline from 2020-2024, then scored every day from 2025 on. It flagged 54 of 2,585 region-days (2.1%).

The ranking mirrors the renewables story: South Australia and Victoria — the most weather-exposed grids — throw the most surprises, while NSW and Queensland are stable.

Two patterns stand out, and both are real market events rather than noise:

• 26 Jan 2026 — South Australia cleared $2,457/MWh against an expected $75. A summer heatwave on a public holiday.
• 12 and 26 June 2025 flag in Victoria, South Australia and Tasmania at once — the model independently rediscovered NEM-wide winter evening peaks, which is a good sign it's detecting market physics rather than per-series noise.