Prediction of car license auction prices using an LSTM neural network for Guangzhou, Shenzhen, Hangzhou, and Tianjin

Abstract

Car license auction price prediction has been identified as a novel practical problem in the economics field for several large Chinese cities. However, making predictions based on the corresponding time-series data is generally regarded as challenging due to the noise and volatility of car license auctions. The prediction capacity of the time-series model is likely to decay over time. Long short-term memory (LSTM) is regarded as a powerful approach for the price prediction of time series. Here, we build LSTM models for predicting the average price of car licenses auctioned in Guangzhou, Shenzhen, Hangzhou, and Tianjin using LSTM neural networks. Our LSTM neural network based on a time series is constructed considering the time characteristics, cities, license quotas, lagged lowest price, lagged average price, price announcements of license auctions, and short-term price trends. We propose a dynamic leave-one-out cross-validation method to address the difficulty in time-series prediction.. The optimal models achieve suitably accurate performance; the average prediction errors of car license auctions are found to 861.71 CNY, 1648.57 CNY, 689.29 CNY, and 689.29 CNY for Guangzhou, Shenzhen, Hangzhou, and Tianjin, respectively.