Ekonometrická analýza efektu TV reklam na přirozenou webovou návštěvnost

Do people react to TV ads or do they just watch them? To answer that, data about organic webpage visits with a minute frequency from Google Analytics and data about TV spots from Nielsen is combined. Since the webpage visits have a strong diurnal pattern, it is decomposed by Kernel smoothing and the residuals are passed into the extraction of significant deviations caused by ads. Since ads, in general, have a lagged effect, the time series have to be treated as a dynamic model. However, the Koyck transformation which is commonly used to establish the dynamic link between advertising and the following response has some limitations, therefore a method eliminating them is proposed. The two methods are compared and the proposed one looks promising. The results of the estimation confirm the ability of TV to drive an immediate response which means that people are willing to turn to their phone to look up some information communicated in the TV spot. On contrary, most of the ads do not have a significant effect on the website traffic which opens another question – which ad qualities cause the highest incremental uplift? The estimated ad effects strongly depend on the amount of impressions caused by the ad. Hence, before further research, the impact of impressions is separated with a linear regression (which proves to be a robust solution). Thereafter, a random forest is built to reveal the complex and non-linear dependencies. As it is difficult to interpret almost all machine learning models, it is necessary to use the method of plotting partial dependencies to understand the outputs from the random forest. Apart from the strong effect of the number of impressions generated by the ad, the motive, TV channel and hour were found significant ad qualities affecting the ability to generate the website traffic. Based on this knowledge it is possible to optimize the TV buying process in order to get the most traffic out of the ads.