Optimal bidding in Google Shopping

A merchant advertising products on Google Shopping has to bid on his advertisements in order for them to be shown. In this project we present a method to calculate optimal bids. When sufficient historical sales data is available, the optimal bid is calculated by estimating model parameters directly from the historical data. For products with insufficient data, we learn optimal bids using a multi-armed bandit. Furthermore, we investigate if product characteristics can be used to aggregate data such that the profit of the products increases. In our test results, we show that the adjusted UCB1 algorithm outperforms the standard UCB1 algorithm. We find that bidding on buckets of products using aggregated data based on product characteristics is not profitable compared to bidding on individual products based on non-aggregated data. Future research is needed to determine better ways to aggregate data in order to make better predictions of model parameters for products with insufficient information.