Business-Cost-Aware Forecasting Losses

Business-cost-aware losses evaluate forecasts by the decision cost they create. MAE and RMSE treat errors symmetrically, but many forecasting decisions do not. Underforecasting demand may cause stockouts; overforecasting may create holding cost or waste.

Asymmetric cost

A simple asymmetric loss is:

$$L(y,\hat{y})=c_u\max (y-\hat{y},0)+c_o\max (\hat{y}-y,0)$$

where $c_u$ is the unit cost of underforecasting and $c_o$ is the unit cost of overforecasting. If $c_u>c_o$, the loss favors higher forecasts.

This connects to quantile forecasting. The optimal quantile level for a simple newsvendor-style decision is:

$$\tau =\frac{c_u}{c_u+c_o}$$

where $\tau$ is the target service quantile.

Decision quantile

Suppose underforecasting one unit costs 4 and overforecasting one unit costs 1. Then:

$$\tau =\frac{4}{4+1}=0.8$$

A forecast near the 80th percentile is more aligned with the decision than the median, because shortages are four times as costly as excess.

Business-weighted metrics

Weights can emphasize high-value items, critical regions, peak periods, or service-level-sensitive horizons. A weighted MAE can be written as:

$$\text{weighted MAE}=\frac{{\sum }_i{w}_i|y_i-{\hat{y}}_i|}{{\sum }_i{w}_i}$$

where $w_i$ is a nonnegative business weight. Weights should be documented because they define which errors matter most.

Practical guidance

• Start from the operational decision, then choose the loss or quantile.
• Document underforecasting and overforecasting cost assumptions.
• Evaluate standard accuracy metrics alongside cost-aware metrics.
• Use cost-aware metrics for model selection only when the cost model is credible.
• Revisit costs when business constraints, margins, or service targets change.

Common failure modes

• Encoding business priorities with undocumented weights.
• Optimizing a cost metric that no operational decision actually uses.
• Treating cost parameters as precise when they are rough estimates.
• Ignoring bias after switching to an asymmetric objective.
• Comparing cost-aware scores across datasets with different weight definitions.

Connections

Cost-aware losses connect forecast error metrics to decisions. Asymmetric stockout and overstock costs often imply quantile loss, prediction intervals, or full probabilistic forecasting rather than a single unbiased point forecast.

References

• Hyndman & Athanasopoulos, FPP3: Distributional forecast accuracy
• Romano, Patterson, and Candes, Conformalized Quantile Regression

Nav

Section — Time-Series Forecasting

← Forecast Error Metrics Forecasting Pitfalls and Worked Examples →