NEWS-DRIVEN STOCK PREDICTION USING NOISY EQUITY STATE REPRESENTATION

Abstract

News-driven stock prediction investigates the correlation between news events and stock price movements. Previous work has considered effective ways for representing news events and their sequences, but rarely exploited the representation of underlying equity states. We address this issue by making use of a recurrent neural network to represent an equity state transition sequence, integrating news representation using contextualized embeddings as inputs to the state transition mechanism. Thanks to the separation of news and equity representations, our model can accommodate additional input factors. We design a novel random noise factor for modeling influencing factors beyond news events, and a future event factor to address the delay of news information (e.g., insider trading). Results show that the proposed model outperforms strong baselines in the literature.

1. INTRODUCTION

Stock movement prediction (Ding et al., 2014; 2015) is a central task in computational quantitative finance. With recent advances in deep learning and natural language processing (NLP), event-driven stock prediction has received increasing research attention (Xie et al., 2013; Ding et al., 2015) . The goal is to predict the movement of stock prices according to financial news. Previous work adopts a relatively simple model on the stock movement process, casting price change as a response to a set of news. The prediction model can therefore be viewed as variation of a classifier that takes news as input and yields a movement direction output. Investigations have focused on news representation, where bag-of-words (Kogan et al., 2009 ), named entities (Schumaker & Chen, 2009) , event structures (Ding et al., 2014) or neural representation features (Ding et al., 2015; Xu & Cohen, 2018) are considered. Intuitively, news events carry information on important changes of company management, market, revenue and other factors, which can affect the fundamental values of equities, and thereby can consequently impact the stock price, as shown in Figure 1 . Properly representing news events is key to modeling such impact on the market. However, the stock market movement can also be influenced by accumulated effects of fundamental changes over time, the overall market sentiment, and other volatile factors, which can be considered as noise to analytical models. These factors have been relatively less modeled by existing work on event-driven stock prediction. For example, although there has been work modeling long-term event impacts by representing event sequences (Ding et al., 2015) , little work has considered representing fundamental values directly. To address these issues, we consider representing the equity state directly using a recurrent neural network over time and propose the stock movement prediction network using Noisy Equity State representation (NES). At each time step, the equity state reflects the current stock price trend, and can be used directly for predicting the next movement. The advantage of separating news representation from equity state representation is that factors beyond news can be modeled as additional input in the recurrent state transition process. Although such factors can be calculated using external tools such as sentiment classification over tweet data, we simply treat them as a random noise factor. The reason is two-fold. First, in practice, noise is inevitable in stock prediction and no single mathematical model can perfectly fit the stock price movement distribution. Second, for fair comparison with existing work on news-driven stock prediction, no additional input should be used on top of standard benchmark input settings. Figure 1 : The stock price volatility and the news impact on 3M Company. Over the first and the second periods, there was only one event. And in the third period, there were two events affecting the stock price movements simultaneously. The input to each recurrent equity state transition consists of a news factor and a noise factor. The news factor consists of three components, namely past news within thrity days, present news within a trading day, and future news within seven trading days. We use real historical news for the future news component for training, and predicted future news according to the current equity state during testing. The motivation for modeling future news is to address the negative effect of delayed news release and insider trading on the prediction accuracy. We represent each news event using contextualized representation of the news title, and aggregate news representations by using the current equity state as a query to conduct attention. The noise factor is integrated into the model by using the current equity state to derive a normal noise distribution specific to the trading day, and then sampling a noise vector. Compared with existing work, our model has three main potential advantages. First, the relative importance of news can be easily visualized using attention. Second, insider trading effect is explicitly handled. Third, noise effect is integrated into the model. All three benefit result directly from the direct representation of equity state. Experiments over the benchmark of Ding et al. (2015) show that our method outperforms strong baselines, giving the best reported results in the literature. To our knowledge, we are the first to explicitly model both events and fundamental equity states for news-driven stock movement prediction, and the first to consider noise factors using a neural random sample module.

2. RELATED WORK

Modeling Price Movements Correlation Most existing work on event-driven stock prediction treats the representation of news events independently using bag-of-words (Kogan et al., 2009) , named entities (Schumaker & Chen, 2009), semantic frames (Xie et al., 2013) , event structures (Ding et al., 2014 ), event embeddings (Ding et al., 2015) or knowledge bases (Ding et al., 2016) . In contrast, work on time-series based stock prediction (Levine & Zervos, 1996; Amihud, 2002; Xu & Cohen, 2018; Zhang et al., 2018a) aims to capture continuous movements of prices themselves. There has also been work modeling the correlations between samples by sparse matrix factorization (Wong et al., 2014 ), hidden Markov model (Zhang et al., 2018a) and Bi-RNNs (Xu & Cohen, 2018; Yang et al., 2019) using both news and historical price data. Some work models the correlations among different stocks by pre-defined correlation graph (Peng & Jiang, 2016) and tensor factorization (Zhang et al., 2018b) . Our work is different in that we use only news events as inputs, and our recurrent states are additionally designed to accommodate noise. Explainable Prediction Rationalization is an important problem for news-driven stock price movement prediction, which is to find the most important news event along with the model's prediction. Factorization, such as sparse matrix factorization (Wong et al., 2014) and tensor factorization (Zhang et al., 2018b) , is a popular method where results can be traced back upon the input features. Our attention-based module achieves a similar goal yet has linear time complexity on feature size. Chang et al. (2016) use attention for characterizing the influence of individual news within a trading day for predicting the cumulative abnormal return in a three-day window. Yang et al. (2019) apply duallayer attention to predict the stock movement by using news published in the previous six days. Each day's news embeddings and seven days' embeddings are summed by the layer. Our work is

