Where forecasting models actually earn their keep in day-ahead trading

Day-ahead desks buy a lot of forecasting. Load, price, congestion, renewable output — each gets its own model, and every model has a roadmap of accuracy improvements waiting to be funded. The question worth asking before the next investment is narrower than “can we make it more accurate?” It’s “where does accuracy convert into a decision we’d actually make differently?” Those are not the same question, and confusing them is how forecasting budgets get spent on error the market never pays you to remove.

Accuracy only counts where it changes a position

A forecast earns money at the margin where it flips a decision — bid or hold, hedge or run, buy the block or wait. Shave error inside a band where the position is the same either way and you’ve bought precision the P&L never sees. The improvements worth funding are the ones that move outcomes at that decision boundary:

• around the bid/no-bid line, where a few dollars of price error changes whether an offer clears
• at the tails, where the scarcity events and price spikes live — the intervals that dominate P&L and that average-error metrics quietly under-weight
• where forecasts feed each other — a load miss that propagates into a price model, compounding downstream
• at the horizon you can still act on, because a better number that lands after the bid is submitted is an academic result, not a trading edge

Measure the forecast in P&L, not in RMSE

The trap is optimizing a model against a statistical loss function that isn’t the desk’s loss function. RMSE and MAPE reward accuracy evenly across every hour; the desk is paid unevenly, concentrated in a handful of high-stakes intervals. Backtesting a model on the decisions it would have driven — and the P&L those decisions would have produced — is a different exercise than minimizing error, and it routinely reorders which improvements are worth the money.

The point isn’t to forecast less. It’s to spend the next model dollar where accuracy becomes a decision, and to stop paying for precision in the hours the market was never going to reward. Tie the forecast to the position it drives, and “how accurate is it?” gives way to the better question: “what would we do differently if it were?”