Using a nested logit model to forecast television ratings

The television environment has become increasingly complex over the past decade, but scant attention has been paid to the modeling and forecasting of television ratings. In this study we use a little-known version of the nested logit model that is suitable for aggregate choice decision data, since television ratings are aggregate measures. We extend this model to include television program random effects, and develop a novel method for predicting program random effects for programs that have not previously been broadcast. Our dataset is comprehensive, spanning the period 2004–2008, and has program ratings for each main broadcaster, as well as some satellite channels, in a market with over 70 channels. We compare our model’s forecasts with those of several other models and show that it markedly outperforms these models.