The Impact of Joint versus Separate Prediction Mode on Forecasting Accuracy

Forecasters predicting how people change their behavior in response to a treatment or intervention often consider a set of alternatives. In contrast, those who are treated are typically exposed to only one of the treatment alternatives. For example, managers selecting a wage schedule consider a set of alternative wages while employees are hired at a given rate. We show that forecasts made in joint-prediction mode—which considers a set of alternatives—generate predictions that expect substantially larger behavioral responses than those made in separate-prediction mode—which considers the response to only one treatment realization in isolation. Results show the latter to be more accurate in matching people’s actual responses to interventions and treatment changes. We present applications to managerial decision-making and forecasting of scientific results.