Prediction of Conversion Rates in Online Marketing - A study of the application of logistic regression for predicting conversion rates in online marketing.

This thesis was written in collaboration with an anonymous European automotive company, Company X, which uses online marketing as a part of their business model. In online marketing it is of inrest to estimate conversion rates, that is the quota of a population at an initial state that will go on to perform a certain action. The action could be, but is not limited to, clicking on an advertisement, interacting in a certain way with the advertisers webpage, or buying a product. If the advertiser can estimate the value of the performed action, and the conversion rate to the action, the advertiser can then calculate the value of the initial state. In extension, is means that if a company knows the life time value of a customer, and can estimate the conversion rate from someone clicking on one of their advertisements to becoming a customer, they can calculate the value of that click. Generally online marketing space is sold through auctions. Different companies bifor the same given advertising space depending on the expected value of the space and pay for exposure. Exposure is either measured in how many users that has seen the ad (impressions) or how many users that have interacted with the ad (usually measured in clicks). Due to this, if a company can improve the precision of how they estimate the value of an impression or click they can spend their online marketing budget more effectively. Considering the size and rapid growth of the online marketing market, this is of high interest. In this thesis a logistic regression modeling appach was compared to a group average approach for predicting conversion rates. The group average approach is based on grouping different advertisements that have few observations into bigger populations and then using the average of the bigger population. The thesis finds that in most cases logistic regression models seems preferable. However, when the variance of the conversion rates is large, the Group average model can be prefereble.