Human-Computer Interactions in Demand Forecasting and Labor Scheduling Decisions

Abstract

We empirically analyze how managerial overrides to a commercial algorithm that forecasts demand and schedules labor affect store performance. We analyze administrative data from a large grocery retailer that utilizes a commercial algorithm to forecast demand and schedule labor across all their stores. In total, this data encompasses 29 million shifts covering 50,000 employees in more than 500 stores, totaling more than 800,000 store-date observations. Using an instrumental variables strategy that exploits exogenous events where the algorithm requires human intervention, we find that managerial overrides made to aggregate labor minutes and employee teams increase store performance. We make the broader point that exercising discretion based on private knowledge about demand and employee relationships is one channel by which managers add value to their stores.