A grocery retailer needs to avoid empty shelves without delivering far more stock than it sells. Predicting average demand alone may not match that decision: a 95% service-level target instead points to the 95th quantile of predicted demand. A quantile loss function is asymmetric, so minimizing it targets a chosen quantile rather than the expected value. section 5.3.1
The target quantile determines what the forecast is for. Quantile forecasts can feed stock-control strategies; their practical value is then judged against both the total stock position and out-of-stock events. Keeping the training loss aligned with the evaluation metric makes that choice explicit. Candidate demand targets are the 1.5th, 25th, 50th, 75th, 95th, and 99th percentiles. Metrics can also weight particular products or groups, for example by item price. section 5.3.1
Start simply, then ask whether extra complexity earns its place:
- Estimate each product’s quantiles from historical sales, using a yearly window for an advanced constant baseline. A simpler seasonal baseline predicts from the same day one week earlier. section 8.6.1
- Fit linear regression with quantile loss using lagged sales and rolling aggregates; price, revenue, and customer counts are possible additional inputs. section 8.6.1
- Compare more complex models with the baselines’ quantile losses and residual bias. For gradient boosting, inspect convergence and whether added lags or narrower windows improve error without overfitting. section 9.4.1
A lower quantile loss is useful evidence, but it is not the entire inventory decision. Check whether improvements hold against simple baselines and help balance delivered stock with the cost of running out. section 5.3.1 section 9.4.1
