End-to-end price movement prediction network using incomplete time series data

Abstract

Market data is complex and dynamic data. It is caused by the traders' interactions in the market and creates market trends. We solved the problem of predicting the direction of price changes for one of the market data, Naphtha. Since Naphtha is also market data, there are two difficulties. 1) Market data is determined by the involvement of persons and is influenced by the investor's psychological state or international situation or policy. This is what makes market data very unstable. Therefore, it is impossible to predict future price based only on past price trends. 2) Naphtha is a petrochemical produced during petroleum refining, and many other petrochemicals are produced. Inevitably, the price of Naphtha is challenging to predict because it is heavily influenced by the prices of products such as crude oil and gasoline, a type of petrochemical. In this paper, we propose a network that can use Naphtha's supply/demand data to solve the difficulties. Naphtha's supply/demand data is collected every half month. The proposed network generates a latent vector that shows the characteristics of the supply and demand data using the Autoencoder structure. Traditional time series data imputation networks mostly use Generative Adversarial Networks(GAN). A Generative Adversarial Networks must learn step by step. However, the proposed network can learn at once and has a more straightforward structure. It also has the advantage of being more structurally convenient than other networks when used for classification. The proposed network was able to use the price data collected daily and the supply/demand data collected on a half monthly to help the naphtha price prediction. We use 16 price-related data and 17 supply/demand data. Price data was collected from July 2011 and supply/demand data from January 2016. In this paper, we propose a network that can efficiently utilize price data and supply/demand data of Naphtha. The network is designed separately because the characteristics of the two data are different. The important features of price data were extracted using CNN, and the characteristics of supply and demand data were extracted using the proposed autoencoder-based imputation network. In order to resolve the imbalance in the collection period, we first learned using price-related data and fine-tuned to set a better starting point of price network. Also, the two networks were combined to enable end-to-end learning. In the experiments chapter, we measured and compared the network performance using the actual collected data. The performance of the proposed network was measured by Accuracy and Matthew's Correlation Coefficient, and the proposed network showed better results than the network using other market data.