The Sum and Its Parts: A Behavioral Investigation of Top-Down and Bottom-Up Forecasting Processes

Abstract:     
Operations planning tasks require demand forecasts at different levels of aggregation. This leaves firms with a choice. Top-level forecasts can be made based on top-level data or, instead, in a bottom-up fashion, where lower-level forecasts are summed up to create a top-level forecast. Lower-level forecasts can be based on lower-level data, or can instead be computed in a top-down fashion, where a forecast is generated based on top-level data and then apportioned to the lower level. Our study investigates the performance of such forecasting processes through a behavioral lens. A key performance driver across the investigated forecasting processes is the correlation structure among lower-level time-series. We identify two relevant biases that affect the relative performance of direct and indirect forecasting approaches: a propensity to neglect distal time-series information in lower-level forecasts, and a propensity for random judgment errors. We investigate how these biases aggregate as the forecasting process “moves” between levels of aggregation, which allows us to characterize the demand environments in which an organization should favor one forecasting process over the other.

Number of Pages in PDF File: 37

Keywords: forecasting process, exponential smoothing, covariation detection, behavioral operations, sales and operations planning

working papers series