REST: Relational Event-driven Stock Trend Forecasting
Wentao Xu, Weiqing Liu, Chang Xu, Jiang Bian, Jian Yin, Tie-Yan Liu
RREST: Relational Event-driven Stock Trend Forecasting
Wentao Xu , *, Weiqing Liu , Chang Xu , Jiang Bian , Jian Yin , , Tie-Yan Liu School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China Microsoft Research Asia, Beijing, China School of Artificial Intelligence, Sun Yat-sen University, Zhuhai, China Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou, China{xuwt6@mail2,issjyin@mail}.sysu.edu.cn{weiqing.liu,chanx,jiang.bian,tyliu}@microsoft.com
ABSTRACT
Stock trend forecasting, aiming at predicting the stock future trends,is crucial for investors to seek maximized profits from the stock mar-ket. Many event-driven methods utilized the events extracted fromnews, social media, and discussion board to forecast the stock trendin recent years. However, existing event-driven methods have twomain shortcomings: 1) overlooking the influence of event informa-tion differentiated by the stock-dependent properties; 2) neglectingthe effect of event information from other related stocks. In thispaper, we propose a relational event-driven stock trend forecasting(REST) framework, which can address the shortcoming of existingmethods. To remedy the first shortcoming, we propose to modelthe stock context and learn the effect of event information on thestocks under different contexts. To address the second shortcoming,we construct a stock graph and design a new propagation layerto propagate the effect of event information from related stocks.The experimental studies on the real-world data demonstrate theefficiency of our REST framework. The results of investment sim-ulation show that our framework can achieve a higher return ofinvestment than baselines.
KEYWORDS
Computational Finance, Stock Trend Forecasting, Event-driven,Graph-based Learning
ACM Reference Format:
Wentao Xu , *, Weiqing Liu , Chang Xu , Jiang Bian , Jian Yin , , Tie-Yan Liu . 2021. REST: Relational Event-driven Stock Trend Forecasting.In Proceedings of the Web Conference 2021 (WWW ’21), April 19–23, 2021,Ljubljana, Slovenia.
ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/3442381.3450032
Among various investment channels, the stock market has beendemonstrating its significant profit potential in the long run. Stocktrend forecasting, aiming at forecasting the future price trends ofstocks, plays as one of the fundamental techniques and has attracted *Work done while Wentao Xu was an intern at Microsoft Research.This paper is published under the Creative Commons Attribution 4.0 International(CC-BY 4.0) license. Authors reserve their rights to disseminate the work on theirpersonal and corporate Web sites with the appropriate attribution.
WWW ’21, April 19–23, 2021, Ljubljana, Slovenia © 2021 IW3C2 (International World Wide Web Conference Committee), publishedunder Creative Commons CC-BY 4.0 License.ACM ISBN 978-1-4503-8312-7/21/04.https://doi.org/10.1145/3442381.3450032 soaring attention from human wisdom [34]. According to the Effi-cient Market Hypothesis [24], the stock price can truly and instantlyreflect the stock value, and the significant price fluctuation indicatesthe reaction to emerging important stock-related information, usu-ally in the form of events. Motivated by this phenomenon, togetherwith a rising variety of sources to obtain published event informa-tion, such as news [7–9, 15, 23, 26, 36], social media [39, 41, 43, 45],and discussion board [21, 27, 46], many efforts had been investigat-ing how to mine the valuable patterns from the event informationfor stock trend forecasting [20].However, most of the existing event-driven methods overlooktwo implicit but critical characteristics of the event information:
Stock-dependent Influence as well as
Cross-stock Influence : • Stock-dependent Influence
An event’s effect on a stock isnot only determined by the event but also could be differentiatedby the stock-dependent properties. As a result, similar eventswould have different effects on different stocks. For example,the event of CEO’s resignation is undoubtedly a negative signalfor a fast-growing company. On the contrary, it is inclined tobe interpreted as a positive sign for a company struggling for along time to find new growth points. However, most existingevent-driven methods pay little attention to such crucial stock-dependent event influence. • Cross-stock Influence
Besides directly-associated stocks, anevent could yield a similar or even more significant influence onthose stocks with no explicit connection. For example, as shownin Figure 1(a), given an event of published performance growthabout a company 𝑠 , it is a positive signal to the stock price of 𝑠 ;meanwhile, since this event may indicate a prosperous marketenvironment of the entire industry 𝑠 belonged to, it could exertthe similar positive influence to another stock 𝑠 within thesame industry. Nonetheless, most of the existing event-drivenmethods omit such implicit but significant cross-stock influenceinvoked by events.To address these shortcomings of existing event-driven meth-ods, in this paper, we propose a novel relational event-driven stocktrend forecasting (REST) framework. Specifically, to consider stock-dependent influence, besides directly modeling the event represen-tations from its detailed textual information, we propose to explorethe stock-dependent context of the event for directly related stocksat the same time. This way, we can capture the same event’s diverseinfluence on various stocks given stock-dependent contexts. Moreconcretely, to model the stock context, we take the historical eventsof this stock into account and consider the subsequent reaction of a r X i v : . [ q -f i n . S T ] F e b WW ’21, April 19–23, 2021, Ljubljana, Slovenia Wentao Xu , *, Weiqing Liu , Chang Xu , Jiang Bian , Jian Yin , , Tie-Yan Liu Stock Trends:Stock Trends:Events:Stocks:
Industry
Violation
BusinessDownstream
Events:
Stocks:
Performance growth
Industry (a) Effect of event from related stock.
Performance growth (b) Different relations have different event effect.(c) The strength of the event effect are dynamic.
Performance growth
Industry (d) The multiple hops propagation of event effect.
Downstream
Figure 1: The observations in the propagation of event effect. stock prices to historical events. By capturing the fluctuation ofone company’s stock price and transaction volume after each of itsrelated events, we can model the investors’ feedback to this com-pany’s historical events, therefore enabling our REST frameworkto differentiate similar events’ influences different companies.Furthermore, to facilitate modeling the event’s cross-stock influ-ence, we construct a stock graph, where the nodes in the graph arestocks, and the edges are the relations between stocks. This graphcan propagate the effect of event information directly related toone stock to other inter-connected stocks. A straightforward wayto propagate event information on the stock graph is to leveragethe Graph Convolutional Networks (GCN) [17]. However, basedon our observations, the propagation of the event effect on thestock graph is complicated, and the traditional GCN can not ap-propriately model it. Thus, in our REST framework, we design anew propagation layer to model the complex cross-stock influenceinspired by a couple of characteristics observed in the real-world:(O1)
An event has variant effects on the stocks connectedwith different relations.
As shown in Figure 1(b), when com-pany 𝑠 was announced that there is a violation in its companyoperation, its downstream company 𝑠 may bear the same riskof a stock price drop, while its competitor company 𝑠 maycontrarily translate this violation exposure event as a positivesignal. Therefore, our new propagation layer separately mod-els the effect of event propagation through different relationsand combines them for the trend prediction.(O2) The propagation strength of the event effect betweentwo stocks is dynamic.
We observed that the event effect’spropagation strength is highly dependent on the dynamicallychanging contexts of the corresponding pair of related stocks.For example, as shown in Figure 1(a), once company 𝑠 an-nounces a remarkable performance growth in the annual re-port, the public tends to feel optimistic about company 𝑠 , inthe same growing industry, if the annual report of the com-pany 𝑠 has not released yet. On the other hand, if 𝑠 hasalready published its annual report before 𝑠 , the performancegrowth of 𝑠 may have very limited influence on 𝑠 ’s stock price, as shown in Figure 1(c). Hence, our new propagationlayer models the dynamic propagation strength of influenceby taking each stock pair’s dynamic context into account.(O3) The effect of the event information could take a multi-hop propagation.
For example, as shown in Figure 1(d), anemerging event in terms of brilliant performance growth onstock 𝑠 could imply a positive effect to the stock 𝑠 , which isin the same industry as stock 𝑠 , as well as another positivesignal to stock 𝑠 , which is the downstream company of stock 𝑠 and one-more step further to 𝑠 . In this paper, we can natu-rally model the multi-hop influence of events by stacking thepropagation layer we designed.To evaluate our proposed REST framework, we conduct extensiveexperiments over the real-world data, and experimental resultsshow that our framework can outperform the existing stock trendforecasting methods. Moreover, we simulate the stock investmentusing a trading strategy, and the investment results show thatour framework can achieve a higher investment return than thebaselines. Finally, we conduct the sensitive tests to further studythe effect of different components in our framework and the casestudy to explain why our framework outperforms the baselines.In summary, the contributions of our work include: • We proposed constructing the stock context with the historicalevents and the corresponding market feed-backs, which isefficient for modeling the various influences of similar eventsto different stocks. • The REST framework we proposed can learn the effect of eventinformation from other related stocks. It can adequately modelsome essential characteristics we observed in the propagationof event effects. • We conducted both experimental evaluation and investmentsimulation on real-world data, and the results and analysisverify the validity and rationality of our framework.
In recent years, stock trend forecasting has attracted much attentionbecause of its vital stock investment role. We can categorize mostof the existing stock trend forecasting work into two categories:the
Event-driven Methods and the
Technical Analysis . According to the efficient market hypothesis [24], people know thatan event that happens on a stock would change the stock informa-tion of this stock, affecting its stock price. There are many efforts tomine the event information from various sources, such as news [7–9, 15, 23, 26, 36], social media [33, 39, 41, 43, 45], and discussionboard [21, 27, 46]. These methods can discover the implicit rulesgoverning the market beyond price data. News is a type of wide-spread event’s source for stock trend forecasting task. [15] proposedto mine news sequence directly from the text with hierarchical at-tention mechanisms; [9] utilized a target-specific abstract-guidednews document representation model to learn the signal in thenews thoroughly; [7] proposes a novel knowledge-driven temporalconvolutional network (KDTCN) to tackle the problem of stocktrend prediction and explanation with abrupt changes. Besides, so-cial media (e.g., Twitter) and discussion board (e.g., Yahoo! Finance
EST: Relational Event-driven Stock Trend Forecasting WWW ’21, April 19–23, 2021, Ljubljana, Slovenia forum) are also important sources of event information for stocktrend forecasting. [43] proposed a methodology to extract socialsentiment from influential Twitter users within a financial commu-nity and provided a more robust predictor of financial markets; [41]mined the information in tweets data and used a deep genera-tive model for stock movement prediction; [39] proposed a novelCross-model attention based Hybrid Recurrent Neural Network(CH-RNN), which can leverage stock price trend representationsto attend daily social text representations through a cross-modalattention interaction; [27] predicted stock price movement based onthe sentiment analysis of event information in the Yahoo! Financeforum.Although there were many efforts in exploiting the event infor-mation for stock trend forecasting, existing event-driven methodsignore the stock-dependent influence and the cross-stock influenceof event information. Our REST framework can overcome the short-ages of previous work and learn the stock-dependent influence andcross-stock influence of event information compared with existingmethods.
Technical analysis [10] is another category of methods for stocktrend forecasting, which is orthogonal to event-driven methods. Thetechnical analysis predicts the stock trend based on the historicaltime-series of market data, such as trading price and volume. Thistype of approach aims to discover the trading patterns that can lever-age for future predictions. Autoregressive (AR) [19] and ARIMA [2]models are the most widely used model in this direction, both forlinear and stationary time-series. However, the non-linear and non-stationary nature of stock prices limits the applicability of AR andARIMA models. With the recent rapid development of deep learn-ing, some studies attempted to apply deep neural networks to catchthe intricate patterns of the market trend [28, 35]. To further modelthe long-term dependency in time series, recurrent neural networks(RNN), especially Long Short-Term Memory (LSTM) network [14],had also been employed in financial perdition [1, 3, 13, 30]. Specifi-cally, [44] proposed a new State Frequency Memory (SFM) recurrentnetwork to discover the multi-frequency trading patterns for stockprice movement prediction; [18] presented a multi-task recurrentneural network with high-order Markov random fields (MRFs) topredict stock price movement direction; [11] leveraged adversarialtraining to simulate the stochasticity during model training. Be-sides, to improve the performance of technical analysis, some recentefforts [6, 12, 16, 25] leveraged Graph Neural Networks [17, 38] tocapture the relationships between different stocks.However, technical analysis is not sensitive to the abrupt changesin stock price caused by external event information of stock [7],limiting its performance on the stock trend forecasting.
In this section, we will introduce some concepts in our proposedrelational event-driven stock trend forecasting framework and for-mally define the problem of stock trend forecasting.
Definition 1. Stock Context.
The stock market is dynamic, andthe stock context of a stock is also dynamic. To better represent the
Table 1: An example the event’s feedback.
Open Close High Low Volume VWAP .
34 12 .
46 12 .
93 12 .
26 364400 12 . .
49 13 .
01 13 .
13 12 .
32 297200 12 . Feedback .
22% 4 .
41% 1 .
55% 0 . − .
44% 2 . stock context, we define the stock context as the combination ofstock’s historical events and these events’ feedback. Definition 2. Event’s Feedback.
An event’s feedback is the rela-tive change of price and transaction volume on the stock that thisevent happened.
Example 1.
Table 1 shows the example of the feedback of a per-formance growth event on the stock
Changan Automobile , and thisevent happened on Jan. 28th, 2016. There are stock price andtransaction volume data on a day, which are opening price (Open), closing price (Close), highest price (High), lowest price (Low), tradingvolume (Volume) and volume weighted average price (VWAP) [4].Thus, this event’s feedback on the stock Changan Automobile isa -dimensional vector, which is the relative change of price andtransaction volume data. The event’s feedback can differentiatethe influences of similar events on different companies, which issignificant for us to learn the distinct effect of event informationfor each stock. Definition 3. Stock Graph.
We define the stock graph as a di-rected graph 𝐺 = ⟨S , R , A⟩ , where S denote the set of stocks inthe market and R is the set of relations between two stocks. A isthe set of adjacent matrices. For an adjacent matrix 𝐴 𝑟 ∈ A ( 𝑟 ∈ R and 𝐴 𝑟 ∈ R |S |×|S | ) of relation 𝑟 , 𝐴 𝑟𝑖 𝑗 = means there is a relation 𝑟 from stock 𝑠 𝑗 to stock 𝑠 𝑖 and 𝐴 𝑟𝑖 𝑗 = indicates there is no a relation 𝑟 from stock 𝑠 𝑗 to stock 𝑠 𝑖 . Definition 4. Stock Price Trend.
The stock price trend is usuallydefined as the future change rate of the stock price [15, 29, 42]. Inthis paper, we define the stock price trend for stock 𝑠 𝑖 at date 𝑡 asthe stock price change rate of the next day: 𝑑 𝑡𝑖 = 𝑃𝑟𝑖𝑐𝑒 𝑡 + 𝑖 − 𝑃𝑟𝑖𝑐𝑒 𝑡𝑖 𝑃𝑟𝑖𝑐𝑒 𝑡𝑖 , (1)where 𝑃𝑟𝑖𝑐𝑒 𝑡𝑖 could be specified by different values, such as openingprice , closing price and volume weighted average price (VWAP), andwe use closing price in our work. Problem 1. Stock Trend Forecasting.
Given the stock-specificinformation (e.g., the textual information from news and socialmedia, the historical stock price) of stock 𝑠 𝑖 at date 𝑡 , the goal ofstock trend forecasting is to forecast the stock price trend 𝑑 𝑡𝑖 . In this section, we elaborate on our relational event-driven stocktrend forecasting (REST) framework. Figure 2 is the illustration ofour REST framework. In Section 4.1, we utilize an event informationencoder to learn the representation of event information. To modelthe stock-dependent influence, in Section 4.2, we learn the contextof each stock; in Section 4.3, we use the event information and the
WW ’21, April 19–23, 2021, Ljubljana, Slovenia Wentao Xu , *, Weiqing Liu , Chang Xu , Jiang Bian , Jian Yin , , Tie-Yan Liu … Type-specific
Encoder …… … … LSTMLSTM
LSTM
Concatenate
24 5 Concatenate
Dense
EventHistorical EventsFeedbacks of Historical Events Stock ContextEvent Information Effect of All Event InformationStock Trend Prediction (a) Event Information Encoder.(b) Stock Context Encoder. (d) Learning the Cross-stock Influence.
51 324 51 324 51 3
Effect of Event Information (c) Learning the Stock-dependent Influence.
Figure 2: The illustration of the relational event-driven stock trend forecasting framework. stock context to learn the stock-dependent effect of event informa-tion on a stock. After that, to model the cross-stock influence, inSection 4.4, we propagate the effect of event information on thestock graph with a propagation layer we designed. Our proposedpropagation layer can model the three observations in propagatingthe event effect, which is mentioned in Section 1.
We first learn the representation of event information of a stockon a day. Figure 2(a) shows the event information encoder that weuse to learn the representation of event information, including thetype-specific event encoder and event sequence encoder.
We utilize a type-specific eventencoder to learn the embedding of an event. The events of stockscan categorize into many types according to their content, whichmay describe the company’s performance, the result of litigation,and clarification of rumors. For each event, we have a type embed-ding 𝑡 𝑖 and a sequence of token embeddings [ 𝑤 𝑖 , 𝑤 𝑖 , ...], where 𝑡 𝑖 represents the type of this event, and 𝑤 𝑖𝑥 represents the 𝑥 -th tokenin its content. We need to learn which word contributes most tothe stock trend for each type of event. We introduce a type-specificmulti-head attention mechanism to aggregate the tokens weightedby an assigned attention value, to reward the tokens offering im-portant signal. Specifically, 𝑢 𝑘𝑖𝑥 = LeakyReLU (cid:16) 𝑊 𝑘𝑒 𝑤 𝑖𝑥 + 𝑏 𝑘𝑒 (cid:17) , (2) 𝛼 𝑘𝑖𝑥 = exp (cid:16) 𝑡 T 𝑖 𝑢 𝑘𝑖𝑥 (cid:17)(cid:205) 𝑥 = exp (cid:16) 𝑡 T 𝑖 𝑢 𝑘𝑖𝑥 (cid:17) , (3) 𝑒 𝑖 = Concat (cid:32)∑︁ 𝑥 𝛼 𝑖𝑥 𝑤 𝑖𝑥 , ∑︁ 𝑥 𝛼 𝑖𝑥 𝑤 𝑖𝑥 , ..., ∑︁ 𝑥 𝛼 𝐾𝑖𝑥 𝑤 𝑖𝑥 (cid:33) . (4)We first feed the token 𝑤 𝑖𝑥 through a fully connected layer to get ahidden representation 𝑢 𝑘𝑖𝑥 . We measure the importance of token on this type as the similarity of 𝑢 𝑘𝑖𝑥 with the type embedding 𝑡 𝑖 andcalculate a normalized attention weight 𝛼 𝑘𝑖𝑥 through a softmax func-tion. Thus our model can learn the token importance of a specifictype of event. Furthermore, we employ multi-head attention [37]to jointly attend to information from different representation sub-spaces at different positions. That is, 𝐾 times independent attentionmechanisms execute the calculation of Equation 3 to get the weight 𝛼 𝑘𝑖𝑥 , and then concatenate the weighted sum vector (cid:205) 𝑥 𝛼 𝑘𝑖𝑥 𝑤 𝑖𝑥 oftokens to get the event embedding 𝑒 𝑖 . Due to the influence of an event onthe stock price would last for several days, we should also considerthe events in the past couple of days when we predict the stocktrend 𝑑 𝑡𝑖 of date 𝑡 . For the event embedding sequence [ 𝑒 𝑖 , 𝑒 𝑖 , ..., 𝑒 𝑛𝑖 ] of stock 𝑠 𝑖 in the latest days, we feed them into a one-layerLSTM [14] and take the last hidden state ℎ 𝑡𝑖 as the representationof event information of stock 𝑠 𝑖 : ℎ 𝑡𝑖 = LSTM (cid:16) 𝑒 𝑖 , 𝑒 𝑖 , ..., 𝑒 𝑛𝑖 (cid:17) . (5)Since the LSTM network can capture the long-term dependency inthe events’ sequence, the representation ℎ 𝑡𝑖 contains the informationof any one of the events in the latest days. We can also rewritethe event information of all stocks on date 𝑡 as a matrix ¯ 𝐻 𝑡 , wherethe 𝑖 -th row of ¯ 𝐻 𝑡 is the event information ℎ 𝑡𝑖 of stock 𝑠 𝑖 . The stock market is dynamic, and the stock context would influencenot only the event information’s effect on the directly-associatedstock but also the propagation of event effects on the stock graph.In order to better represent the stock context, we not only usehistorical events [ 𝑒 𝑖 , 𝑒 𝑖 , ..., 𝑒 𝑚𝑖 ] of stock 𝑠 𝑖 in the past days, butalso utilize the corresponding feedbacks of historical events [ 𝑣 𝑖 , 𝑣 𝑖 ,..., 𝑣 𝑚𝑖 ]. We use the same type-specific event encoder in Section 4.1.1to learn the representation of historical events. Moreover, Section 3has introduced the definition of an event’s feedback, which is the EST: Relational Event-driven Stock Trend Forecasting WWW ’21, April 19–23, 2021, Ljubljana, Slovenia relative change of stock price and transaction volume after an eventhappens.As shown in Figure 2(b), we feed the sequences of historicalevents [ 𝑒 𝑖 , 𝑒 𝑖 , ..., 𝑒 𝑚𝑖 ] and feedbacks of historical events [ 𝑣 𝑖 , 𝑣 𝑖 , ..., 𝑣 𝑚𝑖 ] to two different LSTMs, and concatenate the last hidden states ℎ 𝑒𝑖 and ℎ 𝑣𝑖 of these two LSTMs as the stock context ℎ 𝑐𝑖 : ℎ 𝑒𝑖 = LSTM (cid:16) 𝑒 𝑖 , 𝑒 𝑖 , ..., 𝑒 𝑚𝑖 (cid:17) ,ℎ 𝑣𝑖 = LSTM (cid:16) 𝑣 𝑖 , 𝑣 𝑖 , ..., 𝑣 𝑚𝑖 (cid:17) ,ℎ 𝑐𝑖 = Concat (cid:0) ℎ 𝑒𝑖 , ℎ 𝑣𝑖 (cid:1) . (6)It is crucial to notice that when we forecast the stock trend 𝑑 𝑡𝑖 ondate 𝑡 , we do not know the stock future price and transaction volumeat date 𝑡 + , so the feedbacks of events happened on the date 𝑡 isunknown. Therefore, the historical events and the correspondingfeedbacks we use to learn the context ℎ 𝑐𝑖 do not include the eventsthat happened on the date 𝑡 . We will explore the influence of thestock context in Section 5.5.1. To learn the effect of event information on the stocks differentiatedby the stock-dependent properties (stock-dependent influence), weutilize the representation of event information and the stock contextto learn the effect of one stock’s event information on this stock.In other words, we learn the strength of event information’s effecton a stock. Formally, we learn the strength of event information’seffect from the stock context ℎ 𝑐𝑖 and the event information ℎ 𝑡𝑖 : 𝐷 𝑡𝑖𝑖 = LeakyReLU (cid:16) 𝑎 T (cid:0) ℎ 𝑐𝑖 || ℎ 𝑡𝑖 (cid:1)(cid:17) , (7)where 𝑎 T is a single-layer feed-forward neural network and || represents concatenation. We use the concatetion because it canretain complete information of the event and stock context. 𝐷 𝑡 isa diagonal matrix where 𝐷 𝑡𝑖𝑖 is the effect’s strength of stock 𝑠 𝑖 ’sevent information to stock 𝑠 𝑖 . After that, we left multiply eventinformation ¯ 𝐻 𝑡 by the strength of event effect 𝐷 𝑡 to learn the effectof event information: 𝐻 𝑡 = 𝐷 𝑡 ¯ 𝐻 𝑡 , (8)where 𝐻 𝑡 is the effect of event information and its 𝑖 -th row 𝐻 𝑡 𝑖 isthe effect of stock 𝑠 𝑖 ’s event information on stock 𝑠 𝑖 . To address the limitation of existing event-driven work that cannot learn the effect of events from other related stocks (cross-stockinfluence). We construct a stock graph defined in Section 3, wherethe nodes in the graph are stocks, and the edges are stock relations,such as the industry relation and the downstream relation. We willintroduce the details of the stock relations we used in Section 5.1.After that, we design a novel propagation layer to propagate theeffect of event information from related stocks in the stock graph.When we design the propagation layer, we pay close attention tomodeling the following observations in the propagation of event ef-fect: (O1) different relations result in different event effects; (O2) thedynamic strength of the event effect between two stocks; (O3) thepropagation of event effect on multi-hop path. Given a matrix 𝐻 𝑡 of the effect of event information on date 𝑡 , we progressively designour propagation layer to propagate the effect of event information 𝐻 𝑡 and realize to model these three observations. We first utilize the aggregator inGraph Convolutional Networks (GCN) [17] to propagate the effectof event information from related stocks: 𝐻 𝑡 = ˜ 𝐴𝐻 𝑡 = 𝐷 − 𝐴𝐷 − 𝐻 𝑡 , (9)where ˜ 𝐴 = 𝐷 − 𝐴𝐷 − and 𝐴 is the adjacent matrix for all relations,that is, 𝐴 𝑖 𝑗 = indicates there is an arbitrary relation from stock 𝑠 𝑗 to stock 𝑠 𝑖 . 𝐷 is a diagonal degree matrix with entries 𝐷 𝑖𝑖 = (cid:205) 𝑗 𝐴 𝑖 𝑗 ,and 𝐷 − is used to normalize 𝐴 .After the GCN propagation, the 𝑖 -th row of 𝐻 𝑡 contains all theeffect of event information of stock 𝑠 𝑖 ’s neighbor stocks. Then wecan learn the effect of events of related stocks from the matrix 𝐻 𝑡 . Different relationsbetween two stocks result in different event effects, but the straight-forward GCN propagation can not distinguish the impacts of dif-ferent relations. To remedy this problem, we adopt the idea in therelational GCN [31] to learn the influence of different relations: 𝐻 𝑡 = ∑︁ 𝑟 ∈R ˜ 𝐴 𝑟 𝑊 𝑟 𝐻 𝑡 = ∑︁ 𝑟 ∈R 𝐷 − 𝐴 𝑟 𝐷 − 𝑊 𝑟 𝐻 𝑡 , (10)where 𝐴 𝑟 is the adjacent matrix of relation 𝑟 , and 𝑊 𝑟 is the mappingmatrix for relation 𝑟 . In the relational GCN propagation layer, wemap the effect of event information 𝐻 𝑡 with a relation-specificmapping matrix 𝑊 𝑟 before propagating the event effect, and 𝑊 𝑟 𝐻 𝑡 represents the event effect propagated with relation 𝑟 . Thus, therelational GCN propagation layer can learn the propagation ofevent effects under different relations. The GCN propagationlayer and relational GCN propagation layer use a fixed weight ˜ 𝐴 𝑖 𝑗 or ˜ 𝐴 𝑟𝑖 𝑗 to propagate the event effect from stock 𝑠 𝑗 to stock 𝑠 𝑖 .However, the real-world stock market is dynamic, and the strengthof the event effect between two stocks is also dynamic. To model thisobservation, we propose to use the dynamic propagated weights topropagate the effect of event information: 𝐻 𝑡 = ∑︁ 𝑟 ∈R ¯ 𝐴 𝑟 𝑊 𝑟 𝐻 𝑡 , (11)where ¯ 𝐴 𝑟 is the matrix of dynamic propagated weights.We utilize the stock context to learn the dynamic weights ¯ 𝐴 𝑟 .We learn the dynamic propagated weight ¯ 𝐴 𝑟𝑖 𝑗 of relation 𝑟 from thestock 𝑠 𝑖 ’s neighbor stock 𝑠 𝑗 ( 𝐴 𝑟𝑖 𝑗 = ) to 𝑠 𝑖 with the context of thesetwo stocks: ¯ 𝐴 𝑟𝑖 𝑗 = 𝑓 (cid:16) ℎ 𝑐𝑖 , ℎ 𝑐𝑗 (cid:17) = LeakyReLU (cid:16) 𝑏 T 𝑟 (cid:16) ℎ 𝑐𝑖 || ℎ 𝑐𝑗 (cid:17)(cid:17) . (12)The function 𝑓 (·) can have many different forms, in this paper,we adopt the form of function 𝑓 (·) shown above, where 𝑏 T 𝑟 is asingle-layer feed-forward neural network for the relation 𝑟 and || is the concatenation. If a stock 𝑠 𝑘 is not the neighbor of stock 𝑠 𝑖 ( 𝐴 𝑟𝑖𝑘 = ), we set the propagated weight ¯ 𝐴 𝑟𝑖𝑘 as . WW ’21, April 19–23, 2021, Ljubljana, Slovenia Wentao Xu , *, Weiqing Liu , Chang Xu , Jiang Bian , Jian Yin , , Tie-Yan Liu The effect of event informationwill propagate through the multi-hop path. Therefore, we stackmore propagation layers to gather the neighboring stocks’ eventeffect with a distance of more than one hop. More formally, wepropagate the event effect 𝐻 𝑡𝑙 − from the neighbor stocks of 𝑙 − hops to gather the event effect from the neighbor stocks of 𝑙 hops: 𝐻 𝑡𝑙 = ∑︁ 𝑟 ∈R ¯ 𝐴 𝑟 𝑊 𝑟 𝐻 𝑡𝑙 − . (13)As shown in Figure 2 (d), after performing 𝑙 propagation layers,we obtain the event effect from the neighbor stocks of different dis-tances: [ 𝐻 𝑡 , 𝐻 𝑡 , ..., 𝐻 𝑡𝑙 ]. To jointly learn the effect of original eventinformation 𝐻 𝑡 and the event effect of the neighbors of differentdistances, we leverage the layer-aggregation mechanism [40] toconcatenate the effect of all event information of each stock: 𝐻 𝑡 ∗ = Concat (cid:16) 𝐻 𝑡 , 𝐻 𝑡 , ..., 𝐻 𝑡𝑙 (cid:17) . (14)By doing so, we cannot only learn the event effect from the neigh-bors of different distances but also can control the distance of eventeffect propagation by adjusting the hyper-parameter 𝑙 . We willstudy the effect of the propagating distance 𝑙 in Section 5.5.2. We feed the effect of all event information 𝐻 𝑡 ∗ 𝑖 (the 𝑖 -th row ofthe matrix 𝐻 𝑡 ∗ ) of stock 𝑠 𝑖 to a dense layer, and output our RESTframework’s stock trend forecasting 𝑝 𝑡𝑖 of stock 𝑠 𝑖 on date 𝑡 : 𝑝 𝑡𝑖 = 𝑊 𝐻 𝑡 ∗ 𝑖 + 𝑏. (15) We leverage the stochastic gradient descent (SGD) algorithm tooptimize our REST framework by minimizing the mean squarederror (MSE) loss function with 𝐿 regularization: L = |T | ∑︁ 𝑡 ∈T ∑︁ 𝑠 𝑖 ∈S 𝑡 (cid:16) 𝑝 𝑡𝑖 − 𝑑 𝑡𝑖 (cid:17) |S 𝑡 | + 𝜆 || Θ || , (16)where T is the set of dates in the training period, and S 𝑡 is theset of stocks in date 𝑡 . The 𝑝 𝑡𝑖 is the stock trend prediction of stock 𝑠 𝑖 at date 𝑡 , and the 𝑑 𝑡𝑖 is the ground truth stock trend of stock 𝑠 𝑖 at date 𝑡 . 𝜆 is the regularization parameter; Θ represents all of theparameters in our framework, including the parameters in neuralnetworks, type embeddings, and token embeddings. In this section, we study our relational event-driven stock trendforecasting framework with experiments. We aim to answer thefollowing research questions via the experiments: • RQ1 : How is the performance of our REST framework? • RQ2 : Can our framework achieve a higher investment returnin the investment simulation on real-world datasets? • RQ3 : How is the effect of different components (i.e., the stockcontext, the multi-hop propagation) in our framework? • RQ4 : How does the REST perform in real-world case?
Datasets.
We evaluate our REST framework on the stocks intwo popular and representative stock indexes: CSI 300 and CSI 500.CSI 300 index consists of the largest and most liquid stocks,reflecting the market’s overall performance. CSI 500 index consistsof the largest remaining stocks after excluding the CSI 300Index constituents, reflecting the small-mid cap stocks. We collectedthe stock data, including the event data, stock relations, event’sfeedback, and the label of stock price trend from 2013 to 2018. Wesplit the stock data by time to a training period from 2013 to 2016,a validation period of 2017, and a test period of 2018.
Event Data.
We collected the event information from the com-pany’s announcements , which have predefined types and haveless noise than the event information extracted in the news orsocial media. We extracted its company name, type of content,content (abstract), and publication timestamp as an event for eachannouncement. We clear the content of announcements by convert-ing decimals to integers and removing the tokens whose frequencyis lower than 5; therefore, we can vastly reduce the scale of tokenvocabulary. The event data we collected are from 2013 to 2018, andwe collected and events for the stocks in the CSI300 index and CSI 500 index, respectively. Stock Relations.
To construct the stock graphs for the stocks inCSI 300 and CSI 500 indexes, we collect four types of stock relations ,which are listed as follows:- Industry: there is an industry relation between two stocks whenthey are in the same industry (e.g., Car, Bank, and Electronics).- Business: there is a business relation between two stocks whenthey have the same business. (e.g., wine, phone, and stainlesssteel)- Shareholder: we only consider the top shareholders of eachstock. There is a shareholder relation between two stocks whenthese two stocks have the same shareholder.- Downstream/Upstream: there is a downstream/upstream rela-tion between two stocks when these two stocks’ business has adownstream/upstream relation. Event’s Feedback.
To obtain each event’s feedback, we collectthe daily stock price and volume data of stocks in CSI 300 and CSI500 indexes . There are 6 stock price and volume data on each day,which are the opening price , closing price , highest price , lowest price , volume weighted average price (VWAP) and trading volume . Wecalculate the relative change of stock price and volume data afteran event happened as the event’s feedback. Thus, the feedback ofeach event is a 6-dimensional vector. Label of the Stock Price Trend.
We use the stock price trend de-fined on Equation 1 as the stock trend label on each day. For bettertraining our framework, we apply normalization on the originallabels of the same date. Specifically, we calculate the mean and stan-dard deviation of stock labels on the same date. Then, we subtract We collect the stock relations from a publicly available API tushare: https://tushare.pro/. We collect daily stock price and volume data from http://xueqiu.com.
EST: Relational Event-driven Stock Trend Forecasting WWW ’21, April 19–23, 2021, Ljubljana, Slovenia the mean and divide by the standard deviation for each stock’s labelon this date.
We use three standard evaluation met-rics of regression to evaluate the results of stock trend forecasting,which are
Root Mean Squared Error (RMSE) , Mean AbsoluteError (MAE) and
Median Absolute Error (MedAE) . RMSE iscalculated by:
RMSE ( 𝑦, ^ 𝑦 ) = √︃ 𝑛 (cid:205) 𝑛𝑖 = ( 𝑦 𝑖 − ^ 𝑦 𝑖 ) ; MAE is the meanvalue of absolute error between labels and predictions: MAE ( 𝑦, ^ 𝑦 ) = 𝑛 (cid:205) 𝑛𝑖 = | 𝑦 𝑖 − ^ 𝑦 𝑖 | ; MedAE is the median of absolute error betweenlabels and predictions: MedAE ( 𝑦, ^ 𝑦 ) = median (| 𝑦 − ^ 𝑦 | , ..., | 𝑦 𝑛 − ^ 𝑦 𝑛 |) . We repeat the testing procedure ten times for all the experi-mental results and report the average value to eliminate the fluctu-ations caused by different initialization. In the type-specific text encoder,we set the dimension of type embeddings and token embeddingsto , and the number of head 𝐾 in multi-head attention is ;thus, the dimension of event embedding 𝑒 𝑖 is . The number ofhidden units in the three LSTMs of our framework is . We setthe distance 𝑙 of event propagation on CSI 300 and CSI 500 to and , respectively. The regularization parameter 𝜆 is × − and thenumbers of training epoch is . We compare our REST framework withthe following stock trend forecasting methods: • ARIMA [5]: ARIMA uses the historical stock trend’s series ofeach stock as input to forecast the stock future trend. • TGCN [25]: TGCN leverages a variant of graph convolutionalnetworks [17] to propagate the historical price informationon the stock graph for stock trend forecasting. It only useshistorical price data and does not use the event information. • HAN [15]: A model that utilizes the hierarchical attention mech-anism over media information for stock trend forecasting. • Event-driven : The Event-driven method directly feeds the orig-inal event information ¯ 𝐻 𝑡 in Section 4.1 into a dense layer andoutput the prediction of the stock price trend. • Event-driven_sd : The Event-driven_sd method directly feedsthe effect of event information 𝐻 𝑡 in Section 4.3 into a denselayer and output the prediction of the stock price trend. TheEvent-driven_sd method can learn the stock-dependent influ-ence compared with the Event-driven method. • GCN [17]: GCN model uses the GCN propagation layer in Sec-tion 4.4.1 to propagate the effect of event information on thestock graph. • Relational GCN [22]: This method uses the Relational GCNpropagation layer in Section 4.4.2 to propagate the effect ofevent information. It can learn the different event effects underdifferent relations compared with the GCN model. • REST 𝑙 = : This method uses the dynamic propagation layer inSection 4.4.3 to propagate the effect of event information onthe stock graph. Nevertheless, it can not propagate the effect ofevent information through the multi-hop path. Table 2: Experimental results on CSI 300 and CSI 500.
Method CSI 300 CSI 500RMSE MAE MedAE RMSE MAE MedAE
ARIMA [5] . . . . . . TGCN [25] . . . . . . HAN [15] . . . . . . Event-driven . . . . . . Event-driven_sd . . . . . . GCN [17] . . . . . . Relational GCN [22] . . . . . . REST 𝑙 = . . . . . . REST . . . . . . Table 2 shows the results of our REST framework and the comparedmethods on CSI 300 and CSI 500. Our REST framework can achievethe lowest RMSE, MAE, and MedAE compared with the previousmethods. The Event-driven_sd method is better than the Event-driven method, thus considering the stock-dependent influence ofevent information is essential. The GCN model has better resultsthan the Event-driven_sd, which demonstrates that propagatingthe effect of event information can improve the performance of thestock trend forecasting. When we compare the GCN model and theRelational GCN model, we can find that learning the propagatedevent effect under different relations is vital. The comparison be-tween the results of the Relational GCN model and REST 𝑙 = modelshows that the dynamic propagated weights learned from the stockcontext are better than the fixed propagated weights. Finally, ourREST’s results are better than the REST 𝑙 = model, which verifiesthe effectiveness of the multi-hop propagation. To further evaluate our framework’s effectiveness, we conduct atrading strategy to simulate the stock investment. In detail, we rankthe stocks on date 𝑡 from high to low according to their stock trendpredictions, then select the top 𝑘 stocks to evenly invest and sellthe holding shares of the stocks which rank behind 𝑘 . We considera transaction cost of . for the buying shares and . forthe selling shares for approximating real-world trading. We usetwo financial metrics to evaluate the investment simulation result:the Annual Return and Sharpe Ratio [32]: • Annual Return (AR) is a standard profit indicator in finance,which is the return that an investment provides over a year. • Sharpe Ratio (SHR) measures the performance of an invest-ment compared to a risk-free asset, after adjusting for its risk. Itis defined as:
SHR = E [ R a − R b ] √︁ var [ R a − R b ] , where 𝑅 𝑎 is the returnof given portfolio and 𝑅 𝑏 denotes the risk-free return. E [·] isthe expected value, and var [·] is the variance.The number of stocks we select to invest evenly is , , , and ,respectively. Figure 3 shows the simulation results of the Annual Re-turn and the Sharpe Ratio. Our REST framework can gain the high-est Annual Return and Sharpe Ratio. The Event-driven_sd methodhas a higher investment return than the Event-driven method, sothe stock-dependent influence of event information is vital for stocktrend forecasting. The GCN model has a higher investment return WW ’21, April 19–23, 2021, Ljubljana, Slovenia Wentao Xu , *, Weiqing Liu , Chang Xu , Jiang Bian , Jian Yin , , Tie-Yan Liu TGCN
10 20 30 40
Top k (a) The Annual Return on CSI 300.
10 20 30 40
Top k (b) The Annual Return on CSI 500.
10 20 30 40
Top k (c) The Sharpe Ratio on CSI 300.
10 20 30 40
Top k (d) The Sharpe Ratio on CSI 500.
Figure 3: The results of investment simulation.Table 3: Effect of the stock context.
Stock Context CSI 300 CSI 500RMSE MAE MedAE RMSE MAE MedAE
REST
Event 𝑙 = . . . . . . REST
Feedback 𝑙 = . . . . . . REST 𝑙 = . . . . . . Table 4: Effect of the distance of event propagation.
Distance CSI 300 CSI 500RMSE MAE MedAE RMSE MAE MedAE
REST 𝑙 = . . . . . . REST 𝑙 = . . . . . . REST 𝑙 = . . . . . . REST 𝑙 = . . . . . . than the Event-driven_sd model, which means the propagation ofevent information can make the forecasting more accurate. Theinvestment results of Relational GCN show the effectiveness oflearning the propagated event effect under different relations; theinvestment results of the REST 𝑙 = model demonstrate that dynamicpropagated weights learned from the stock context can achieve ahigher investment return; the comparison between the results of theREST 𝑙 = model and our REST framework verify the effectivenessof multi-hop propagation. To explore the effect of thestock context described in Section 4.2. We observe the performanceof REST 𝑙 = ’s variants: 1) REST Event 𝑙 = merely uses the historicalevents to learn the stock context; 2) REST Feedback 𝑙 = only uses thehistorical events’ feedbacks to learn the stock context. We summa-rize the results in Table 3, and have the following observations: • From the comparison between REST
Event 𝑙 = and REST Feedback 𝑙 = ,we can find that using historical events’ feedbacks to learnthe stock context is better than historical events. It verifies theeffectiveness of the historical events’ feedbacks in the stocktrend forecasting. • Using historical events and historical events’ feedbacks simul-taneously can achieve the best performance. Thus we need to
The net profit growth in the first quarter of 2018 is 38.35%
WLY (000858.SZ)
KCMT (600519.SH)
LZLJ (000568.SZ)
Tongwei (600438.SH)
Industry BusinessDownstreamGround Truth +0.85%
ARIMA -1.83%
Event-driven -0.45%
REST +1.15%
Ground Truth -1.25%
ARIMA +1.13%
Event-driven +0.04%
REST -0.80%
Ground Truth +2.22%
ARIMA -1.68%
Event-driven -0.27%
REST +1.31%
Industry
Figure 4: The ground truth of stock price trend and the fore-casting results of ARIMA, Event-driven and REST methods. jointly use historical events and historical events’ feedbacks tolearn the stock context.
To studythe effect of multi-hop propagation’s distance 𝑙 in Section 4.4.4,we vary the number of propagation layer of the REST frameworkfrom to and observe the results on CSI 300 and CSI 500. Ta-ble 4 summarizes the experiments. Then we have the followingobservations: • Increasing the distance 𝑙 of propagation is capable of boost-ing the performance substantially. When 𝑙 = , the REST 𝑙 = achieves the lowest MAE and MedAE on CSI 500; when 𝑙 = ,the REST 𝑙 = achieves the best RMSE, MAE and MedAE on CSI300, and lowest RMSE on CSI 500. • When we further increase the number of propagation layers to , we observe that our framework’s performance will reduce.It suggests that 𝑙 is a sensitive hyper-parameter, and too muchof the propagation layer can not further improve stock trendforecasting. To illustrate our REST framework’s effec-tiveness, we dig into a specific case study on three companies:
WLY , KCMT , and
LZLJ . While belonging to the same general industry,these three stocks connect yield complicated competitive relations,
EST: Relational Event-driven Stock Trend Forecasting WWW ’21, April 19–23, 2021, Ljubljana, Slovenia making it quite hard to discern the influence of one company’sevents on the others.For example, on April 28, 2018, the stock
WLY released a per-formance announcement that the net profit growth in the firstquarter of 2018 is 38.35%. Such a positive event of
WLY can intu-itively imply a bullish trend of the corresponding industry. However,complicated competition between different companies can lead todiverse stories. Particularly, while the stock
KCMT , whose primaryproducts are distinct as
WLY , can benefit from this positive event,another stock,
LZLJ , having the same type of product thus the morecompetitive position as
WLY , could suffer from the adverse effectsof the positive event of
WLY .We observe the forecasting results of different methods on theclosest trading day after this event happened. Figure 4 illustratesthe ground truth of the stock price trend and the forecasting resultsof ARIMA, Event-driven, and REST. The ARIMA forecasts the stocktrend with historical stock trend series only, and the Event-drivenmethod can not learn the cross-stock influence of event information,so they both fail to make the correct forecastings. However, ourREST framework can propagate the effect of event information withthe propagation layer we designed. Thus, our REST framework canmake the correct forecasting on the
KCMT and
LZLJ . The effectof
WLY ’s event on other related stock is difficult to be found byordinary people unless experts with rich domain knowledge. OurREST framework can make the correct numerical prediction whilethe experts can only forecast a rise or a fall.Moreover, our REST framework can propagate the event infor-mation through multi-hop paths, so our framework can also makecorrect forecasting on the stock
Tongwei , which is the downstreamcompany of
KCMT . Stock trend forecasting is vital for investors to seek maximizedprofits from financial investment. In this paper, we proposed a rela-tional event-driven framework (REST) to forecast the stock trend.Our proposed framework can address some shortcomings of exist-ing event-driven work. The first shortcoming is existing methodsoverlook the influence of event effect differentiated by the stock-dependent properties (stock-dependent influence). The second oneis that existing methods only use one stock’s information to predictits stock price trend, overlooking the effect of event informationfrom other different stocks on the stock trend of this stock (cross-stock influence). The experiment results on real-world stock marketdata demonstrated the effectiveness of our framework. The resultsof the investment simulation show that our REST framework canachieve a higher investment return.In the future, we would mine more abundant and diverse stockinformation from the Web, such as the news, social media, and otherstock related information, and fuse different kinds of stock infor-mation to forecasting the stock price trend. We would also explorethe techniques of unsupervised learning and apply unsupervisedlearning to stock trend forecasting.
ACKNOWLEDGMENTS
Wentao Xu and Jian Yin are supported by the National Natural Sci-ence Foundation of China (U1711262, U1711261,U1811264,U1811261, U1911203,U2001211), Guangdong Basic and Applied Basic ResearchFoundation (2019B1515130001), Key R&D Program of GuangdongProvince (2018B010107005). Jian Yin is the corresponding author.
REFERENCES [1] Ryo Akita, Akira Yoshihara, Takashi Matsubara, and Kuniaki Uehara. 2016. Deeplearning for stock prediction using numerical and textual information. In . IEEE, 1–6.[2] Adebiyi A Ariyo, Adewumi O Adewumi, and Charles K Ayo. 2014. Stock priceprediction using the ARIMA model. In . IEEE, 106–112.[3] Wei Bao, Jun Yue, and Yulei Rao. 2017. A deep learning framework for financialtime series using stacked autoencoders and long-short term memory.
PloS one
12, 7 (2017), e0180944.[4] Stephen A Berkowitz, Dennis E Logue, and Eugene A Noser Jr. 1988. The totalcost of transactions on the NYSE.
The Journal of Finance
43, 1 (1988), 97–112.[5] Robert Goodell Brown. 2004.
Smoothing, forecasting and prediction of discretetime series . Courier Corporation.[6] Yingmei Chen, Zhongyu Wei, and Xuanjing Huang. 2018. Incorporating Corpo-ration Relationship via Graph Convolutional Neural Networks for Stock PricePrediction. In
Proceedings of the 27th ACM International Conference on Informationand Knowledge Management . ACM, 1655–1658.[7] Shumin Deng, Ningyu Zhang, Wen Zhang, Jiaoyan Chen, Jeff Z Pan, and HuajunChen. 2019. Knowledge-Driven Stock Trend Prediction and Explanation viaTemporal Convolutional Network. In
Companion Proceedings of The 2019 WorldWide Web Conference . ACM, 678–685.[8] Xiao Ding, Yue Zhang, Ting Liu, and Junwen Duan. 2016. Knowledge-drivenevent embedding for stock prediction. In
Proceedings of coling 2016, the 26thinternational conference on computational linguistics: Technical papers . 2133–2142.[9] Junwen Duan, Yue Zhang, Xiao Ding, Ching-Yun Chang, and Ting Liu. 2018.Learning Target-Specific Representations of Financial News Documents ForCumulative Abnormal Return Prediction. In
Proceedings of the 27th InternationalConference on Computational Linguistics . 2823–2833.[10] Robert D Edwards, John Magee, and WH Charles Bassetti. 2018.
Technical analysisof stock trends . CRC press.[11] Fuli Feng, Huimin Chen, Xiangnan He, Ji Ding, Maosong Sun, and Tat-Seng Chua.2019. Enhancing Stock Movement Prediction with Adversarial Training.
IJCAI (2019).[12] Fuli Feng, Xiangnan He, Xiang Wang, Cheng Luo, Yiqun Liu, and Tat-Seng Chua.2019. Temporal relational ranking for stock prediction.
ACM Transactions onInformation Systems (TOIS)
37, 2 (2019), 27.[13] Qiyuan Gao. 2016.
Stock market forecasting using recurrent neural network . Ph.D.Dissertation. University of Missouri–Columbia.[14] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory.
Neuralcomputation
9, 8 (1997), 1735–1780.[15] Ziniu Hu, Weiqing Liu, Jiang Bian, Xuanzhe Liu, and Tie-Yan Liu. 2018. Listeningto chaotic whispers: A deep learning framework for news-oriented stock trendprediction. In
Proceedings of the Eleventh ACM International Conference on WebSearch and Data Mining . ACM, 261–269.[16] Raehyun Kim, Chan Ho So, Minbyul Jeong, Sanghoon Lee, Jinkyu Kim, andJaewoo Kang. 2019. Hats: A hierarchical graph attention network for stockmovement prediction. arXiv preprint arXiv:1908.07999 (2019).[17] Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification withGraph Convolutional Networks. In
International Conference on Learning Repre-sentations (ICLR) .[18] Chang Li, Dongjin Song, and Dacheng Tao. 2019. Multi-task Recurrent NeuralNetworks and Higher-order Markov Random Fields for Stock Price MovementPrediction: Multi-task RNN and Higer-order MRFs for Stock Price Classification.In
Proceedings of the 25th ACM SIGKDD International Conference on KnowledgeDiscovery & Data Mining . ACM, 1141–1151.[19] Lili Li, Shan Leng, Jun Yang, and Mei Yu. 2016. Stock Market AutoregressiveDynamics: A Multinational Comparative Study with Quantile Regression.
Math-ematical Problems in Engineering
IEEE Transactions on Knowledge and Data Engineering
30, 2 (2017), 381–399.[21] Qing Li, TieJun Wang, Ping Li, Ling Liu, Qixu Gong, and Yuanzhu Chen. 2014.The effect of news and public mood on stock movements.
Information Sciences
278 (2014), 826–840.[22] Wei Li, Ruihan Bao, Keiko Harimoto, Deli Chen, Jingjing Xu, and Qi Su. 2020.Modeling the Stock Relation with Graph Network for Overnight Stock MovementPrediction. 4491–4497.[23] Qikai Liu, Xiang Cheng, Sen Su, and Shuguang Zhu. 2018. Hierarchical Comple-mentary Attention Network for Predicting Stock Price Movements with News.In
Proceedings of the 27th ACM International Conference on Information and
WW ’21, April 19–23, 2021, Ljubljana, Slovenia Wentao Xu , *, Weiqing Liu , Chang Xu , Jiang Bian , Jian Yin , , Tie-Yan Liu Knowledge Management, CIKM 2018, Torino, Italy, October 22-26, 2018 . 1603–1606.https://doi.org/10.1145/3269206.3269286[24] Burton G Malkiel and Eugene F Fama. 1970. Efficient capital markets: A reviewof theory and empirical work.
The journal of Finance
25, 2 (1970), 383–417.[25] Daiki Matsunaga, Toyotaro Suzumura, and Toshihiro Takahashi. 2019. ExploringGraph Neural Networks for Stock Market Predictions with Rolling WindowAnalysis. arXiv preprint arXiv:1909.10660 (2019).[26] Arman Khadjeh Nassirtoussi, Saeed Aghabozorgi, Teh Ying Wah, and DavidChek Ling Ngo. 2015. Text mining of news-headlines for FOREX market predic-tion: A Multi-layer Dimension Reduction Algorithm with semantics and senti-ment.
Expert Systems with Applications
42, 1 (2015), 306–324.[27] Thien Hai Nguyen, Kiyoaki Shirai, and Julien Velcin. 2015. Sentiment analysison social media for stock movement prediction.
Expert Systems with Applications
42, 24 (2015), 9603–9611.[28] Jigar Patel, Sahil Shah, Priyank Thakkar, and Ketan Kotecha. 2015. Predictingstock market index using fusion of machine learning techniques.
Expert Systemswith Applications
42, 4 (2015), 2162–2172.[29] Yu Qin and Yi Yang. 2019. What you say and how you say it matters: Predictingstock volatility using verbal and vocal cues. In
Proceedings of the 57th AnnualMeeting of the Association for Computational Linguistics . 390–401.[30] Akhter Mohiuddin Rather, Arun Agarwal, and VN Sastry. 2015. Recurrent neuralnetwork and a hybrid model for prediction of stock returns.
Expert Systems withApplications
42, 6 (2015), 3234–3241.[31] Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, IvanTitov, and Max Welling. 2018. Modeling relational data with graph convolutionalnetworks. In
European Semantic Web Conference . Springer, 593–607.[32] William F Sharpe. 1966. Mutual fund performance.
The Journal of business
39, 1(1966), 119–138.[33] Jianfeng Si, Arjun Mukherjee, Bing Liu, Qing Li, Huayi Li, and Xiaotie Deng. 2013.Exploiting topic based twitter sentiment for stock prediction. In
Proceedings ofthe 51st Annual Meeting of the Association for Computational Linguistics (Volume2: Short Papers) . 24–29.[34] Rashmi Sutkatti and D TORSE. 2019. Stock market forecasting techniques: Asurvey.
International Research Journal of Engineering and Technology
6, 5 (2019).[35] Jonathan L Ticknor. 2013. A Bayesian regularized artificial neural networkfor stock market forecasting.
Expert Systems with Applications
40, 14 (2013),5501–5506.[36] Manuel R Vargas, Beatriz SLP De Lima, and Alexandre G Evsukoff. 2017. Deeplearning for stock market prediction from financial news articles. In . IEEE, 60–65.[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is allyou need. In
Advances in Neural Information Processing Systems . 5998–6008.[38] Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, PietroLiò, and Yoshua Bengio. 2018. Graph Attention Networks.
International Con-ference on Learning Representations (2018). https://openreview.net/forum?id=rJXMpikCZ accepted as poster.[39] Huizhe Wu, Wei Zhang, Weiwei Shen, and Jun Wang. 2018. Hybrid Deep Se-quential Modeling for Social Text-Driven Stock Prediction. In
Proceedings ofthe 27th ACM International Conference on Information and Knowledge Man-agement, CIKM 2018, Torino, Italy, October 22-26, 2018 . 1627–1630. https://doi.org/10.1145/3269206.3269290[40] Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichiKawarabayashi, and Stefanie Jegelka. 2018. Representation Learning on Graphswith Jumping Knowledge Networks. In
Proceedings of the 35th International Con-ference on Machine Learning (Proceedings of Machine Learning Research) , Vol. 80.PMLR, Stockholmsmässan, Stockholm Sweden, 5453–5462.[41] Yumo Xu and Shay B Cohen. 2018. Stock movement prediction from tweets andhistorical prices. In
Proceedings of the 56th Annual Meeting of the Association forComputational Linguistics (Volume 1: Long Papers) , Vol. 1. 1970–1979.[42] Linyi Yang, Tin Lok James Ng, Barry Smyth, and Riuhai Dong. 2020. HTML:Hierarchical Transformer-based Multi-task Learning for Volatility Prediction. In
Proceedings of The Web Conference 2020 . 441–451.[43] Steve Y Yang, Sheung Yin Kevin Mo, and Anqi Liu. 2015. Twitter financialcommunity sentiment and its predictive relationship to stock market movement.
Quantitative Finance
15, 10 (2015), 1637–1656.[44] Liheng Zhang, Charu Aggarwal, and Guo-Jun Qi. 2017. Stock price predictionvia discovering multi-frequency trading patterns. In
Proceedings of the 23rd ACMSIGKDD International Conference on Knowledge Discovery and Data Mining . ACM,2141–2149.[45] Zhenkun Zhou, Jichang Zhao, and Ke Xu. 2016. Can online emotions predict thestock market in China?. In
International conference on web information systemsengineering . Springer, 328–342.[46] David Zimbra, Hsinchun Chen, and Robert F Lusch. 2015. Stakeholder analy-ses of firm-related web forums: Applications in stock return prediction.