| Abstract [eng] |
Internet search advertising is a huge worldwide industry. In 2011 about 19 billion dollars were spent on it in China alone. The ad buying process is auction based therefore the main problem every advertiser encounters is how to optimize bids for certain keywords as to maximise the expected profit. Research aim. The main goal is to create an algorithm/model which could evaluate the market situation on every bid in consideration and predict the ones which should maximise expected profit. Literature overview. Even though the scope of the articles, analysing budget optimization problem, is quite wide, it could be categorized into three main branches based on the models, used in them. Those three branches can be called MDP’s (Markov Decision Process), stochastic models and knapsack secretary. All of them have some benefits, compared to others, however at the same time they all lack correspondence towards a real life situation, which is experienced by advertisers everyday. Here the budget optimization problem is being solved by only using the data, available to an ordinary advertiser. Data. Data from Google Search auctions was used in model creation. Data was collected from 2016.06.01 to 2016.12.10. Data consists of statistics of 31 different keywords. Results and conclusions. After applying machine learning algorithms, a few elements were distinguished as important to a final profit estimation. Those are average ad position, number of impressions, click-through-rate and average conversion rate. A predicting models were constructed to evaluate each variable. Results of the model showed that the bids, used in reality, were extremely off - the average daily profit was negative. The model created two possible bid scenarios, where optimal bids were calculated on hourly and daily basis. The later more closely resembles a real life situation. Each of those scenarios predicted a positive daily profit with marginal differences between them. Keywords: Internet auctions, budget optimization, internet advertising. |
