Optimizing Labor Schedules with More Accurate Sales Forecasts

Our client, a leading nationwide restaurant chain, depends on daily sales forecasts to inform many of their operational planning decisions, including labor scheduling. Labor schedules aim to make sure restaurants have enough staff members available to serve customers efficiently and satisfactorily while also avoiding overstaffing. To fully optimize these schedules, our client post-processes their daily sales forecasts, breaking them into 15-minute intervals throughout the day.

These forecasts were intended to help anticipate the natural rhythms of each restaurant’s peak and off-peak periods. But they failed to present the full picture, compromising efficiency and, ultimately, sales. Our client’s existing method relied on recent historical data to break down their daily sales forecasts into 15-minute intervals. However, our team identified scenarios in which these recent sales patterns simply weren’t predictive of future patterns.

For example, holidays like the Fourth of July and Labor Day tend to follow patterns that are more similar to the same day in prior years and less similar to typical days in the recent past. Similarly, during periods of seasonal transition—like the weeks leading from spring into summer—sales patterns from recent weeks may not be representative of future weeks, and previous years’ summertime sales patterns may be more predictive.

For our client, relying solely upon recent sales data also produced poor forecasts when recent weeks’ sales were highly variable or contained outliers.

After identifying these areas for improvement, our team set about developing an alternative decomposition methodology that led to more accurate 15-minute forecasts.

To improve forecast accuracy, we identified the optimal weighting of the data for each day from prior weeks and prior years. We further customized our solution to account for the holidays that can fall on different weekdays each year. We also identified the parameters that optimally tune the forecast-smoothing process, ensuring a balance between capturing the underlying signal and rejecting noise from spikes and outliers.

We then backtested this new approach by applying our model to historical data and observing how it would have performed had it been available in the past. We found that these changes led to a 16 percent reduction in forecast error.

Our client now generates a more accurate set of fine-grained 15-minute sales forecasts that is robust to seasonal and holiday fluctuations. This helps their restaurants more precisely set staff schedules to mitigate the risk of under- and overstaffing as demand ebbs and flows throughout the day.