ESTIMATING THE TV AUDIENCE PER DEMOGRAPHIC SEGMENT FROM HOUSEHOLD LEVEL DATA

In TV advertisements, recent developments such as the set-top box have made it pos­sible for media agencies to collect detailed numbers of viewers (also called TV ratings) of the shows they broadcast. Ideally, this information can be used to gain insight in the demographic composition of the target group for each TV channel at each moment in time. However, this can not directly be achieved. The reason for this, is that only household-level data can be obtained, rather than individual-level data. In this work, two methodologies have been developed to obtain TV ratings per demographic segment from household viewing and composition data. In the first method, theories of group utility and choice models are combined to form a household level choice model, using individual utilities. From a number of group utility specifications, the multiplicative group utility (using the product of individuals' utilities as household utility) proved to be the most suitable to apply in a choice model. In the second method, a linear regression model and a LlghtGBM model are estimated on data where segment TV ratings are known, to apply this model on data from a different source. This method uses aggregated household ratings as predictors and is therefore called the aggregated method. To tune the LlghtGBM model, Bayesian Optimization of its hyperparameters is used. The LightGBM model proved to outperform the linear model by a large margin in terms of model fit. The LightGBM model also outperformed the choice model in terms of model fit and computation time. Therefore, we conclude that the LightGBM model of the aggregated method is the most suitable to estimate segment TV ratings. However, the aggregated method can only be applied to household viewing data where segment TV ratings of data from a similar source is available. In case this data is unavailable, one has to apply the choice model.