A Dynamic Prediction Model For Online Auction Price Based On Functional Regression

Recently, with the rapid growth of e-commerce, the scholars have developed a great interest in this field. Lots of empirical researches occurred. However, most researchers regard the online auction data as cross-sectional, and ignore influence of the ever changing dynamics.In this paper, we study the online auction from another aspect and expand the object of interests to the evaluation of price curves and their dynamics. Besides, we develop a dynamic forecasting system to predict the on-going prices. By dynamics, we mean the model can predict the price in a whole auction process (not only the final price), and can update based on new arriving information.Prediction on online auction price is a big challenge, because price data has some special features:1) it’s the combination of the longitude data and cross-sectional data,2) the distribution of the arriving for the price is not uniform,3) the dynamics for the data are changing all the time,4) e-commerce data can be easily got and so you can get enormous data in one research, which is a big burden for your machine to run on it. The dynamic forecasting model in this paper is based on functional data analysis, thus is very suitable to solve these problems.The main work done and conclusions gain from this research are:1) Estimate the price curve and its first-order and second-order derivatives and uses them to make prediction. We use the functional methods to describe the price curves and their dynamics and compare the prediction of our model to that of ARIMA, and find our model is obviously better than the traditional one.2) Analyze the influence on price curve by each independent variable using functional regression and find the influences change during the auction process. For the impact strength, most independents’impact get weak as the auction goes, but minimum increase influences more at the end.3) Departure the dynamic part from our model and find the result turns bad. So we find the evolvement of the dynamics can improve the prediction. For the same samples, the overall MAPE increases from7%to15%, without considering dynamics.4) Compare statistical characteristics of Treasures Area and Volkswagen Area and find auctions in different areas distinct from each other.