Abstract
In the modern digital world, the rise of online advertising has brought attribution modelling to the digital advertising forefront. Attribution is the problem of assigning credits to different marketing activities that lead users to perform desirable actions and is considered as one of the biggest challenges that the online advertising industry is currently facing. There is an imperative need for data-driven tractable models that will overpass the inadequacies of the currently used rule-based and ad-hoc attribution models, such as the “last-touch” attribution model (LTA). A data oriented methodology will gain insights towards attributing credits to different marketing interventions based on the observed individual user online behaviour.
In this report, two data-driven and statistically oriented attribution models are developed; a Second-order Probabilistic Model (SPM) and a Bagged Logit Model (BLM). The analysed dataset corresponds to an automotive advertising campaign of “A Fortune 500 Automobile Manufacturer” that took place in the United States from June 15, 2009 to August 23, 2009. Due to the limited amount of available data, the recorded third-party websites serving the client’s advertisements are handled as different marketing channels. That is, the proposed data-driven attribution models estimate the contribution of each channel (website) to the conversion probability at user-level in order to attribute the recorded conversions among the used channels. The attribution results of SPM and BLM are compared to the ones provided by the LTA; the latter seems to partly overestimate both statistically and economically insignificant channels and underestimate efficient ones. An additional predictive random forest model is generated, having the power to accurately predict 97.47% of targeting outcomes (conversion or not) given a user’s funnel path and other explanatory features, such as the total advertising exposure time.
The contribution of this report can be summarised in the insightful information provided to marketers through the proposed attribution methodologies regarding the efficiency of different marketing channels. Furthermore, the developed models, which are easy both to implement and to interpret, can be easily extended to incorporate additional marketing formats and features in order to capture the marketing activities that move users down the funnel path towards conversion and optimise marketing budget allocation among the most efficient channels.