High-Frequency CSI300 Spot and Futures Price Predictions via the Neural Network

The financial price index prediction is an important issue to market participants and policy makers. We focus on the price prediction problem for the CSI300 spot, its nearby futures, and its first distant futures using high-frequency 1 min data for the time frame from the launch of the futures market to about two years after all its constituent stocks becoming shortable, a period with continuously expanding trading of these financial indices. We employ the neural network to model these complex price time series and attempt to answer the following research questions: (1) can the prices be predicted by their own lags, and if so, how well; (2) can other two series help predictions of one series and by how much; and (3) how complex do the models need to be and how robust can they be? We find that these three price indices can be accurately predicted through relatively low complex models with five hidden neurons based on their first-order lags, leading to robust performance with R 2 ’s above 99.99% and root mean square errors in the range of 1.5–3 as compared to average prices in the range of 2597–2612. Our results also show that incorporating other series could sometimes improve predictions of one series by modest magnitude, without increasing model complexity, shedding light on existence of information feedback among the three series. We also find that the spot is the easiest to be predicted, followed by the nearby futures, and then the first distant futures. Findings here should be of interest and use to design trading strategies, monitor portfolio risk, and form index price predictions.