AU2018101513A4

AU2018101513A4 - Comprehensive Stock Prediction GRU Model: Emotional Index and Volatility Based

Info

Publication number: AU2018101513A4
Application number: AU2018101513A
Authority: AU
Inventors: Bo HUI; Jingtian Lin; Zifeng Xie; Zhoufan Yu; Danfeng Zhang; YuHao Zhuang
Original assignee: Individual
Current assignee: Individual
Priority date: 2018-10-11
Filing date: 2018-10-11
Publication date: 2018-11-15
Anticipated expiration: 2026-10-11

Abstract

Abstract Investment behavior, in its core, is a deliberate action which is usually elaborately planned before conducted. According to investment's deliberate character, the market in a great proportion consists of individual investors' subjectivity. The traditional stock prediction model has a crucial limitation that while consuming a substantial amount of calculation resources it does not take into account the emotional factors in investment behavior. In this investigation, we will advance the conventional stock prediction model, model that predicts stock performance solely based on past volatility, by further incorporating additional features: Baidu Index, Emotional Index, amount of commentary on East Money. Namely, in the model we introduced, two influential factors in the stock market, the direction of popular opinion and stock popularity (coverage) will be employed and analyzed utilizing Gated Recurrent Unit (GRU) model to provide a precise and thorough prediction of stock tendency. Commentary , Data Craling Cleaning Preprocessing Word2Vec TextCNN Semantic Model &Vector Training Convertion Stocks Price Classificafion Crawling Stocks Commentary Commentary SearchIndex Volatility PopulaCrawling Multi-GRU Model Training Approximate Prediction (figure 1)

Description

TITLE

Comprehensive Stock Prediction GRU Model: Emotional Index and Volatility Based

FIELD OF INVENTION

The present invention relates to a comprehensive stock trend predicting method based on neural networks, and particularly, a method adding in social economic features to improve the predicting accuracy when making investments.

Background of the Invention

Majority of Market prediction model put forward prediction according to past stock volatility only; the fully connected neural network, for instance, is a popular medium employed in the industry. While solely considering past stock market float, the conventional market model fails to mention the importance of subjectivity in the market, and thus bear many limitations; to conclude, the conventional models are not accurate and incapable of adapting the constantly changing market atmosphere. Inspired by “Investor Sentiment and Stock Returns” (Kenneth L. Fisher and Meir Statman), sentiments and popular consensus show prominent influence in terms of stock investment that the sentiments, though differentiated between individual investors, characterized one’s investment behavior. Yumo Xu and Shay B. Cohen in “Stock Movement Prediction from Tweets and Historical Prices” also indicate a correlation between popular judgment and market volatility, and created an unprecedented stock price prediction model processing emotional context from tweets. With the recognition of the significance of emotion, processing, comprehending, and incorporating commentary component from the stock forum can be therefore especially essential for improving prediction accuracy.

SUMMARY OF THE INVENTION

Data Medium

The investigation will be considered unprecedented, targeting the Chinese stock market, in specific, computer industry. China, as the second greatest economic entity, embraced numerous investors, and recorded, on average, over 50 billion trade value daily. To verify a possible correlation and making the result as accurate, a massive amount of data must be analyzed to train the model and therefore minimize the margin of error; the atmosphere of the Chinese stock market, having such a tremendous amount of trades, is a suitable object for the research

Baidu

Baidu, serving as the largest research engine in China, owns 300 million users. The Baidu index reflects the search frequency and the importance of a typical stock. It is one important indicator suggesting how popular is a stock and how much attention investors have on that stock.

East Money

East Money is the largest stock market forum in China or perhaps worldwide; providing users with credible, professional, and timely news, East Money occupied the attention of a variety of different investors including retail investors, influential individuals, foundation, and professionals. The commentary on East money is most credible and comprehensive containing opinions and thoughts from all dimension and perspective. As the top stock forum, East Money is also incredible in its size; it possesses over 50 million of users among the total of 100 million of stock investors in China which is the amount equivalent to half of the investor population. With 219 million visitors, the views on the website are also frequently visited and commented on a daily basis. Such quantity of data is the key to improve the accuracy of the model; by incorporating the data from East Forum we wish to train the most comprehensive model.

Summary

The model created consists of 4 critical features considering the popularity of stock, polarity of commentary, and past stock volatility. The model is well rounded and comprehensive having taken into account all the important indication factors. The system we utilized is both advanced and novel. The TextCNN embrace high accuracy can precisely interpret the meaning of a sentence. The Multi-GRU model, compared with traditional LSTM, calculate faster, more precise, and less amount of required processing.

Structure

In this advanced prediction model, four features, Commentary Polarity, Commentary Popularity, Stock Volatility, Search Index, will be eventually processed by the Multi-GRU Model to result with an approximate prediction of stock tendency. (Figure 1)

These values are crawled from the two major websites, Baidu and East Money. Among these values, Commentary Polarity and Commentary Popularity are produced via a series of complex models. The commentary information will be initially crawled from the East Money websites, going through preprocessing 3 processing terms and TextCNN before it can be inputted to analyze. In specific, the commentary which in sentence form first decompose by the jieba, separating into segments of a word or word group. The word group then be transformed by Word2Vec into vector form. The vector representing the meaning of each sentence will then be inputted into TextCNN. The TextCNN comprehend the emotional context in the commentary and produce the Commentary Polarity (log[(l+P)/(l+N)]) and commentary popularity (P+N).

DESCRIPTION OF THE DRAWINGS

The appended drawings are only for the purpose of description and explanation but not for limitation, wherein:

Fig.l is an approximate prediction of stock tendency

Fig.2 is The Scrappy engine first encoded the stock code, generating URL for each stock information.

Fig3. is the proportion of positive emotion compare to negative emotion will then be narrowed down using the log equation.

Fig4. is these 4 dimensions are all preprocessed via the standardized procedure, inputting the stock fluctuate value.

Fig5. is The emotional index is an extraordinary innovation; adding emotion shows a prominent improvement in the accuracy of stock tendency prediction.

Fig6. is Testing with two different models, one only takes into account the past stock volatility while the advanced prediction model takes into account in addition Baidu index, emotional polarity index, and emotional popularity index, the one considering emotional context shows a higher accuracy in terms of predicting the data of 10 stocks selected.

Fig7. is The improvements are especially remarkable in the stock 600718.

DESCRIPTION OF PREFERRED EMBODIMENT

In order that the present invention can be more readily understood, reference will now be made to the accompanying drawing to illustrate the embodiments of the present invention. I. Crawling

The Scrappy engine first encoded the stock code, generating URL for each stock information.

The scrappy Engine proceeds to send the requests to the downloader which further request from East Money forum. After sending the request, the commentary resources will be downloaded and organized into a response pack. The crawler then resolves the data pack and identify the essential information needed for further evaluation. The processed information item and the commentary data will be handled to the pipeline to further dispose and save as Excel documents. The data will thus be incorporated in the model. (Figure 2) II. Preprocessing

The commentary will be crawled from East Money in the form of Chinese letters. This semantic information will then going through preprocessing stage turning into vectors.

In preprocessing, the major goal is to convert semantic information, the enormous amount of commentary, into vectors, the form that computer can comprehend. The commentary, initially download in Chinese form, will first be separated into segments utilizing Jieba for the computer to separately processing long or large pieces of semantic information; for instance, a sentence in Chinese “GUPIAOYAOZHANGJIA” (stock will climb) will be categorized into several segments “GUPIAO” (stock) “YAO” (will) “ZHANGJIA” (climb). Then segments of semantic information will be converted into vectors via Word2Vec, each vector symbolizing a meaning, a word or word group. The vector then is inputted into Semantic &Vector Conversion which aggregates the vector of each individual word groups and put into a new vector representing the meaning of an entire sentence. The major achievement in the preprocessing is converting the sentence into meaningful vectors and utilizing Word2Vec to interpret it. III. TextCNN Model Training

After converting semantic information into number form, Text CNN then trains the model to further understand the emotional context inside each and every comment.

The TextCNN model help to interpret the vector imported and outputs a number between 0 and 1 representing the possibility of the vector represented text been positive in terms of its emotion. The output will be categorized into a number either 0 or 1 following the equation:

The index N=1 represents that the view is positive and N=0 represents that the view is negative.

We first artificially differentiate and recognize the emotional context of 20000 comments whether positive or negative and feed 15000 commentary information to TextCNN; TextCNN then adjusts the correlation between the input and output to minimize the loss. After the margin of error has been minimized into the point where it cannot be lower, the model is then relatively accurate. Our trained model, tested by the rest of 5000 of commentary information, is proved to be trustworthy with the accuracy up to 90%±5%. The processed index then converted to an emotion index using the equation: E = log2[(l + P)/(l E in the equation representing an emotional index, P represents the amount of positive emotion and N represent the amount of negative emotion. The equation (1+P)/(1+N) represent the comprehensive emotional orientation of comments. To plus one for both P and N, it modifies the value to be smoother. For better application and further calculation, the proportion of positive emotion compare to negative emotion will then be narrowed down using the log equation. (Figure 3) IV. Multi GRU Model Training GRU stock tendency prediction model inputs 4 different indexes, the stock volatility, search index, emotional polarity, and commentary popularity. Before the data are inputted to the GRU model, it has to be preprocessed. Besides the two emotional index from East Money forum which will be processed by TextCNN and presented in

the form of E=log2[(l+P)/(1+N)] and P+N, the stock volatility also have to be

preprocessed before it can be utilized. After downloading the closing price of the stock daily, the data will be further disposed of following the equation: F represents the fluctuating value of the price equivalent to the Pn meaning the closing stock price that day minus Pn_i representing the closing stock price the previous day and decided by 5 * Pn_i + 0.5. The equation effectively converts the stock price change into a value in between 0 and 1; after conversion, F represents the trends of the stock that if F > 0.5 representing the stock is bullish on the day and if F < 0.5 it symbolizes decline. The value undergoes the process can be the utilized.

For search Index, we have collected the search frequency for both stock code and stock name, and adds up the value to produce an overall search index; as the attention and popularity of the stock increases, B increases, reflecting the general interests on the stock.

After encoding the data into the GRU model, the model provides an index in between 0 and 1 meaning the possibility that the stock will climb. The value then is converted into either 1 or 0 following the equation:

The result of the equation represents whether the stock is going up or going down; 0 symbolize decline and 1 symbolize climb.

GRU model is similar to the LSTM model in many ways; in fact, it is one of the varieties of LSTM. The traditional LSTM model consists of 3 functions for different purposes: the input gate, forget gate, and output gate. The GRU model advanced on the base of LSTM, simplifying the calculation process; GRU only consists of 2 gates, the reset gate and update gate. Reset gate determine the how previously inputted information synthesis with the current input, and update gate determine the proportion of data been reserved and further evaluated. If setting the reset value as 1 and update value as 0, the model will degenerate into traditional RNN model.

For stock prediction problems, a massive amount of data is required to be embedded; nevertheless, with the Chinese market that is not yet mature enough, stocks we evaluated normally contain only data for the past 5 to 10 years, which is approximately thousands of data per stock. In such circumstances where the training data is limited, the GRU model consumes relatively less parameter, and therefore the training speed and quality will be advanced on the traditional model. To comprehensively consider, the project decided to employ the GRU model, and each hidden layer is seated to have 6 neurons. GRU stock prediction model import data from 4 different dimensions; these 4 dimensions are all preprocessed via the standardized procedure, inputting the stock fluctuate value, search index, emotional index, for previous k day, and connect the last hidden layer with a fully connected layer of two output representing the percentage of climb and decline for the next day. (Figure 4)

Evaluation

The emotional index is an extraordinary innovation; adding emotion shows a prominent improvement in the accuracy of stock tendency prediction. (Figure 5) Testing with two different models, one only takes into account the past stock volatility while the advanced prediction model takes into account in addition Baidu index, emotional polarity index, and emotional popularity index, the one considering emotional context shows a higher accuracy in terms of predicting the data of 10 stocks selected. (Figure 6) The controlled model without index shows an average accuracy of 53.01% while the model with index shows an accuracy of 56.74 % which stably improved approximately 4 % accuracy compared to the conventional model. The performance of the enhanced model shows advantages in most of the stocks predicted. Though for 600571, 600718, and 600588, the conventional model achieves a higher accuracy, the enhanced model still maintains a stable performance and retain over 50% of accuracy. The improvements are especially remarkable in the stock 600718 (Figure 7). The accuracy for the most time outmatches the conventional model, it is obtrusive that the line for the enhanced model constantly staying above the conventional model and is about 20% more accurate in its peak. The example shows convince evidence showing the progression. The stable and reliable accuracy of the advanced prediction model is sufficient to convince that it is capable of serving as an indicator assisting investors to conduct the best investment strategy.

Claims

Claim

1. An enhanced RNN GRU model function to predict the stock tendency on a daily basis, forecasting stock tendency based on 4 inputs, commentary polarity (symbolizing pessimistic or optimistic), Baidu search frequency index (frequency of a stock been searched), East Money commentary popularity (amount of comments on a stock), past stock volatility (stock trends).
2. The model processes a substantial amount of commentary crawled from two major websites, East Money (largest stock forum in China) and Baidu (most frequently used research engine in China), comprehending the polarity of comments and converts into values between 0 and 1 via TextCNN representing the polarity of the text. The commentary polarity, along with the popularity of the stock which represented by the commentary popularity and research index and stock volatility, combined and processed via RNN GRU model to forecast an index between 0 and 1 meaning the possibility of stock activity, climb or decline; the predicted index will then be classified as either 0 or 1, following the equation that if n< 0.5, n=0 and if n > 0.5, n=l, to achieve the approximate prediction.