04 — method
Day-ahead prices are a short-horizon time series with strong seasonality and spikes. This page is only about that statistical choice.
01 — The series
Hourly day-ahead auction prices show a strong weekly cycle (same hour last week is often hard to beat), a daily shape, and occasional extreme spikes. The forecast target is a full next-day curve (many lead hours at once), not a single one-step-ahead point.
02 — Why these models
Seasonal naive (168h) and hour-of-week means are the right baselines: they capture calendar structure with almost no risk of overfitting. SARIMAX is the classical linear seasonal ARIMA reference with calendar-style drivers. Prophet is an additive multi-seasonality check — useful for contrast, usually weak on spikes. LightGBM multi-horizon is the nonlinear tabular approach common in electricity price forecasting: one model per lead hour, using lags and known calendar (and residual-load style drivers when present). LSTM is a sequence baseline for comparison; on this problem tabular boosting typically wins.
03 — What fits best
LightGBM multi-horizon is the primary production choice. It fits nonlinear interactions (level × hour-of-week, residual load, wind/solar when available) better than linear SARIMAX, targets each lead hour of the day-ahead curve directly (no recursive multi-step error pile-up), and usually beats seasonal naive on walk-forward MAE. Pure classical models remain as transparent references, not the champion.
04 — Roles in the stack
Baselines set the floor. Classical models explain linear seasonal structure. LightGBM carries the multi-zone path. Prophet and LSTM stay as gallery checks so a stronger result is judged against more than one family — not as a how-to for rebuilding the system.