When a forecast
becomes a decision.

A forecast can move closer to the observed outcome while leading to a more expensive action. Following the forecast into the decision makes the reason visible: what can be done, what it costs and where a small error changes the choice.

My public preprint on utility-weighted forecasting examines this problem in financial decisions. It treats a predictive distribution as an input to an objective, trading costs and a set of permitted positions. The resulting decision is evaluated after implementation costs. The January 2026 preprint is the research source for this article.

A closer number can cross the wrong boundary

Consider a deliberately simplified, hypothetical decision. There are two choices: take one fixed-size position for one period, or do nothing. Opening and closing the position costs a total of 10 basis points, measured against its value. One basis point is one-hundredth of a percentage point. Ignore every other cost and risk for this illustration.

Forecast A predicts a gross return of 11 basis points. Forecast B predicts 5. A fixed rule takes the position only when its predicted gross return exceeds the 10-basis-point cost. A therefore triggers a trade; B leaves the position unopened.

Suppose the realised gross return is 9 basis points. A misses by 2 and B by 4, so A has the lower squared prediction error: 4 against 16. Yet A's position returns minus 1 basis point after costs. B's decision returns zero. The prediction-error ranking and the realised-payoff ranking disagree.

This constructed example establishes no forecasting strategy and estimates no market effect. Its purpose is to isolate a boundary. Change the cost to 4 basis points and both forecasts trigger the same position. With the original cost, an error near the trading threshold changes the action; elsewhere, a larger numerical error may leave it untouched.

Choose the score with the task

This connection has a substantial methodological history. Tilmann Gneiting's work on point forecasts explains why a forecast request and its scoring rule need to agree. Asking for a mean, a median or a particular quantile defines different tasks. An evaluation can mislead when it silently substitutes one target for another.

Adam Elmachtoub and Paul Grigas's Smart Predict-then-Optimize framework carries the connection into optimisation. Its decision loss measures the consequence of the choice induced by a prediction. The objective and constraints therefore help determine how prediction models are trained and compared.

Keep the adjustment tied to the decision

My preprint proposes a calibration step between the forecast and the optimiser. It weights probability errors by their significance for the specified decision, including the effects of costs and binding constraints. The theoretical comparison depends on stated assumptions about the objective, costs and feasible choices. It does not establish an unconditional advantage for every recalibration.

The empirical section uses a December 2025 evaluation window. Its reported improvements concern the methods and conditions studied; the paper calls for evaluation across further months and regimes. The public record identifies this version as a preprint. This article adds no replication result or claim of general profitability.

Make the comparison inspectable

A useful follow-on comparison could publish a decision ledger beside its forecast scores. For each case, it would show what information was available, the prediction, the rule applied, the action selected and the cost charged. A reader could then locate the cases where two methods actually chose differently.

The hypothetical example suggests a particularly simple check: move the cost threshold while holding the forecasts and outcomes fixed. Record when the ranking changes and why. That exercise would examine sensitivity to an assumption; it would not create additional observed market data.

The research question is therefore specific enough to inspect. Which forecast differences alter a feasible action, and what follows when they do? Keeping the forecast, decision rule and implementation record connected makes an improvement claim considerably more informative.