The accuracy of television network rating forecasts: the effects of data aggregation and alternative models

Denny Meyer, Rob J Hyndman
(2006) Model Assisted Statistics and Applications 1(3), 147-155

 pdf

This paper investigates the effect of aggregation in relation to the accuracy of television network rating forecasts. We compare the forecast accuracy of network ratings using population rating models, rating models for demographic/behavioural segments and individual viewing behaviour models. Models are fitted using neural networks, decision trees and regression. The most accurate forecasts are obtained by aggregating forecasts from segment rating models, with neural networks being used to fit these models. The resulting models allow for interactions between the variables and the non-linear carry-over effect is found to be the most important predictor of segment ratings, followed by time of day and then genre. The analysis differs from those of previous authors in several important respects. The AC Nielsen panel data considered stretches over 31 days, 24 hours per day, 60 minutes per hour, making it necessary for ratings to be appropriately transformed prior to the fitting of the rating models and for non-viewing time periods to be under-sampled when fitting the models for individual viewing. For the first time individual viewing within each 15 minute time period is defined by network choice and proportion of viewing time.

Keywords: aggregation, discrete choice models, neural networks, decision trees, two stage models.