Optimal initial bid and new keywords suggestion in sponsored search auctions

The suggestion of profitable new keywords and the estimation of the optimal initial bid for each of them, are two crucial problems that every bidding strategy has to address in order to maxime profits in Sponsored Search Auctions (SSA). In this thesis, both problems are solved from the perspective of an economic agent not considered in the research developed so far: the broker. It is an intermediary between search engine and advertisers, purchasing ad slots from the former and selling them to the latters. The optimal initial bid is cast as a regression problem, where regressors are the keyword itself (encoded in a numerical vector) and a set of features known a-priori, provided by the search engine. The dependent variable is the actual bid observed. The complex highly non-linear relationship is learnt by a Fully Connected Neural Network. The observations used for training and testing are those belonging to keywords with a positive average profit, in order to ensure that the relation learnt leads to positive profitability. The new keywords suggestion problem is solved defining a general web scraping method, designed to work on almost every webpage returned from a query, followed by keywords extractors algorithms to retrieve relevant terms. A filter is created to clean the extracted keywords, improving quality of suggestions. The proposed neural network reaches a prediction error in the test set of just 0.16$, while performance of the new keywords suggestion tool is assessed by human evaluators, using broadly adopted metrics in the field (Relatedness, Non-obviosness and F-score). The best keywords extractor model reaches an F-score higher than 0.6 for every scrap level and number of suggesions requiered. The synergy between the new keywords suggestion tool, able to provide related but non-obvious (and so cheap) keywords, and the optimal initial bid tool, able to estimate a reliable starting bid, should increase the overall broker’sprofit.