Stock market prediction on high-frequency data using generative adversarial nets.
Stock price forecasting is an important problem in the financial world as it helps in developing effective stock trading strategies. In this paper, we propose a general framework employing long short-term memory (LSTM ) and convolutional neural network (CNN) for adversarial training to predict high-frequency stock markets. This model takes the publicly available index provided by the trading software as input, avoids complex financial theory research and difficult technical analysis, and provides convenience for ordinary traders who are not financial professionals. Our research simulates the trading patterns of actual traders, and adopts the method of rolling training set and test set to analyze the impact of model update cycle on forecasting performance. A large number of experiments show that the method we propose can effectively improve the accuracy of stock price direction prediction and reduce prediction errors.
Predicting stock prices is an important goal in the financial world [1-3] because reasonably accurate predictions have the potential to generate high financial returns and hedge market risks. With the rapid development of Internet and computing technology, the frequency of performing operations on the stock market has increased to a fraction of a second [4, 5]. BM&F Bovespa (Brazilian stock exchange) started high-frequency operations in 2009, and the number of high-frequency operations increased from 2.5% in 2009 to 36.5% in 2013. Aldridge and Krawciw [6] estimate that in 2016 HFT initiated on average 10%‒40% of stock trading volume and 10%‒15% of foreign exchange and commodity trading volume. These percentages indicate that the high frequency stock market is a global trend.
In most cases, forecast results are evaluated from two aspects: one is the forecast error between the actual price and the forecast value (mainly RMSE (Root Mean Square Error) or RMSRE (Root Mean Square Relative Error)); is the directional prediction accuracy, i.e. the percentage of correctly predicting the direction of the price series, such as upward and downward movements, is really important for decision making. Even small improvements in predictive performance can lead to large gains [7, 8].
However, due to the complexity and chaotic dynamics of the market and the many undecidable, non-stationary random variables involved , predicting stock prices is not an easy task [9]. Many researchers from different fields have studied the historical patterns of financial time series and proposed various methods to predict stock prices.
In order to achieve considerable performance, most of these methods need to carefully select input variables, use professional financial knowledge to build predictive models, and use various statistical methods for arbitrage analysis, which makes it very difficult for people outside the financial field. Difficult to predict using these methods. Predicting stock prices [10-12].
Goodfellow et al. introduce generative adversarial networks (GANs). [13], where image patches are generated from random noise using two networks trained simultaneously. Specifically, in GANs, the discriminative network learns to distinguish whether a given data instance is real or not, while the generative network learns to confuse by generating high-quality data. While this approach has been successfully and applied to a wide range of domains, such as image inpainting, semantic segmentation, and video prediction [14-16], to the best of our knowledge, it has not been used for stock prediction. ´ This work uses basic technical indicator data as input variables, which can be obtained directly from trading software, making it easy for people outside the financial field to predict stock prices through our method. This study introduces a prediction error loss and an orientation prediction loss, and shows that generative adversarial training [13] can be successfully used to combine these losses to produce satisfactory prediction results. We refer to this prediction architecture as GAN FD (GAN for minimizing forecast error loss and direction prediction loss). In order to conform to the actual trading practice, this work performs a rolling split on the training set and test set of the original data , which we will describe in detail in the experimental section.