Forecasting the nearly unforecastable: why aren’t airline bookings adhering to the prediction algorithm?

Using 27 million flight bookings for 2 years from a major international airline company, we built a Next Likely Destination model to ascertain customers’ next flight booking. The resulting model achieves an 89% predictive accuracy using historical data. A unique aspect of the model is the incorporation of self-competence, where the model defers when it cannot reasonably make a recommendation. We then compare the performance of the Next Likely Destination model in a real-life consumer study with 35,000 actual airline customers. In the user study, the model obtains a 51% predictive accuracy. What happened? The Individual Behavior Framework theory provides insights into possibly explaining this inconsistency in evaluation outcomes. Research results indicate that algorithmic approaches in competitive industries must account for shifting customer preferences, changes to the travel environment, and confounding business effects rather than relying solely on historical data.