Aggregating Web Offers to Determine Product Prices

Historical prices are important information that can help consumers decide whether the time is right to buy a product. They provide both a context to the users, and facilitate the use of prediction algorithms for forecasting future prices. To produce a representative price history, one needs to consider all offers for the product. However, matching offers to a product is a challenging problem, and mismatches could lead to glaring errors in price history. We propose a principled approach to filter out erroneous matches based on a probabilistic model of prices. We give an efficient algorithm for performing inference that takes advantage of the structure of the problem. We evaluate our results empirically using merchant offers collected from a search engine, and measure the proximity of the price history generated by our approach to the true price history. Our method outperforms alternatives based on robust statistics both in tracking the true price levels and the true price trends.