Real-time forecasting of online auctions via functional K-nearest neighbors

Forecasting prices in online auctions is important for both buyers and sellers. With good forecasts, bidders can make informed bidding decisions and sellers can select the right time and place to list their products. While information from other auctions can help forecast an ongoing auction, it should be weighted by its relevance to the auction of interest. We propose a novel functional K-nearest neighbor (fKNN) forecaster for real-time forecasting of online auctions. The forecaster uses information from other auctions and weights their contributions by their relevance in terms of auction, seller and product features, and by the similarity of the price paths. We capture an auction's price path by borrowing ideas from functional data analysis. We propose a novel Beta growth model, and then measure the distances between two price paths via the Kullback-Leibler distance. Our resulting fKNN forecaster incorporates a mixture of functional and non-functional distances. We apply the forecaster to several large datasets of eBay auctions, showing an improved predictive performance over several competing models. We also investigate the performance across various levels of data heterogeneity, and find that fKNN is particularly effective for forecasting heterogeneous auction populations.