Insight · forecast method
Probabilistic electricity forecasting: a forecast should show uncertainty.
For power trading, a point forecast without a band and backtest discipline is dangerous. Alvo reads a forecast as decision support: p50 for planning, p10/p90 for risk and diagnostics for trust.
Insights
In brief
The median scenario: a useful planning center, not a clearing-price guarantee.
The uncertainty band: a range, not a market floor or cap.
Walk-forward evaluation must use only data known before gate closure.
Giacomini-White significance reduces the risk of promoting a random model edge.
Why a point forecast is not enough
Electricity prices can move because of weather, imports, outages, cap regime or liquidity shocks. If the interface shows only one number, the trader cannot see how fragile the decision is.
A probabilistic forecast adds a band: p50 as the center, p10 and p90 as uncertainty edges. A wide band is not automatically a bad model; it may be the honest answer when the market is less predictable.
Baseline first
A strong seasonal-naive baseline is surprisingly hard to beat. Alvo therefore does not score a model in isolation: rMAE is read against the naive forecast on an out-of-sample window, not against a polished in-sample chart.
LEAR/Lasso is a good V1 backbone because it is transparent, fast and hard to fake. A more complex model must pass the same backtest and significance gate before it should replace the simpler one.
Walk-forward protects against leakage
In a proper backtest, each day repeats the live process: train on the past, forecast the next day, then roll forward. Weather actuals, late corrections and post-gate facts must not be inserted into the feature snapshot as if they were known in the morning.
This is where many forecast demos overclaim. If the model saw the future, its score is not evidence. Alvo treats leakage as a product bug, not a metric footnote.
Significance and calibration are different questions
Giacomini-White asks whether the model's edge over a baseline is statistically credible under daily recalibration. Calibration asks whether the band covered realized prices at the rate it promised.
A model can have an attractive p50 and still a poor band; a band can be wide enough while p50 has no actionable edge. You need both layers, and both need denominators.
Insight · forecast method
FAQ
Why not MAPE?
MAPE behaves badly near zero and for negative prices; power prices are better read through MAE/RMSE/rMAE and proper quantile diagnostics.
Is a narrow band always better?
No. A narrow band without coverage can be overconfident. A band must be calibrated, not merely attractive.
What if the GW gate is weak?
Use the forecast as a warning layer or require manual review; do not sell it as model edge.
Sources and next steps
Workspace
Want the operating checklist?
The short tutorial shows how to read p10/p50/p90 before a human decision.
Open tutorial