AU2018101512A4

AU2018101512A4 - A comprehensive stock trend predicting method based on neural networks

Info

Publication number: AU2018101512A4
Application number: AU2018101512A
Authority: AU
Inventors: Xun DONG; Yunhua Pei; Xi Peng; Yu Shi; Ruiyun Wang; Ye Yuan
Original assignee: Dong Xun Miss; Shi Yu Miss; Wang Ruiyun Miss
Current assignee: Dong Xun Miss; Shi Yu Miss; Wang Ruiyun Miss
Priority date: 2018-10-11
Filing date: 2018-10-11
Publication date: 2018-11-15
Anticipated expiration: 2026-10-11

Abstract

A stock trend predicting method with social economic features based on neural networks is disclosed.Stock movement prediction is crucial to stock market analysis. The accuracy of existing methods may not achieve investors increasing requirements to predict the stock market. This invention provides a new method to predict the stock moving trend. It provides users with a valid reference to have a prediction to the stock movement. In this invention, we use Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) which belong to Artificial Neural Network (ANN) as main computing methods and blend in stock-related information factors and the social economic features. This invention provides investors with a clear reference on whether the investment will make profits in the next five days. reaeJeoo i 4 nf rnai nf a u e dat prfmesn Figure I

Description

TITLE A comprehensive stock trend predicting method based on neural networks

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

Nowadays, more and more people are addicted to stock. Under the circumstance of information explosion, the fluctuation of stock becomes increasingly severe. Stock forecasting is a kind of non-linear analysis. Stock-related factors are essential to analyze the stock market. At the same time, social economic features, such as West Texas Intermediate (WTI) price, US Dollar Index (USDX), gold price based on US dollar and WTI futures price can also influence it. In order to make a successful investment, investors have to take a lot of factors under their considerations to know the future situation of the stock market.

The major method that has been used to predict the stock movement is Artificial Neural Network (ANN). ANN is a computing system that is vaguely inspired by biological neural network that constitutes animal brains. It abstracts networks of human neurons, establishes a model, and makes up different networks according to different ways of connection. Because of the limitations of traditional economic statistics, it’s hard to make a scientific prediction for the stock movement. However, ANN can easily deal with incomplete, vague and irregular data. Hence, it can be used to predict the stock movement.

CNN and Recurrent Neural Network (RNN) are two categories within ANN. The prediction result of CNN is stable. Moreover, GRU and LSTM are two algorithms developed from RNN. LSTM, which contains an input gate, a forget gate and an output gate, does well in storing and accessing information. As a variant of LSTM, GRU, which consists of an update gate and an output gate, is more straightforward.

Previously, a lot of studies of stock prediction based on ANN have been performed. In 2017, Selvin et.al studied on forecasting stock prices by CNN, RNN and LSTM. They inputted factors including day stamp, time stamp, transaction id, stock price and volume of stock sold in each minutes. As a result, CNN appeared to be the most accurate method among the three. Xu and Cohen conducted a research on predicting stock movement by GRU in 2018. Date from Twitter and historical stock price listings are used to complete the prediction. In conclusion, they achieved an accuracy of 0.58 and MCC of 0.080796. Roondiwala, Patel and Varma conducted a research in 2017 by using LSTM to predict the rise or fall percentage of stock prices. In addition, they tested the accuracy of prediction by adding more and more factors, such as opening price, closing price, the highest price and the lowest price, into the parameters. After the experiment, they found that the more factors they took under consideration, the more accurate their model would be. In order to make the stock trend prediction more accurate, we synthesize CNN, GRU and LSTM to be our main methods and select some stock-related factors and social economic features as our main data. The following sections of the patent description will explain more details about the method that is used in our research.

SUMMARY OF THE INVENTION

Instead of the stock price, this invention predicts the trend of stock movement in the future five days. Data chosen for this invention not only includes stock-related features, but also takes social economic features under consideration. To provide investors with the most suitable method, three algorithms are evaluated in this invention. The invention effectively deals with the difficulty of prediction in the fluctuant stock market. By applying this invention to their investments, investors benefit from obtaining a more accurate reference and diminishing their loss. For example, if the invention returns a positive prediction result, investors will be certain that they can buy the stock and make profits in the following five days.

Stock-related and social economic features are collected and processed. Then they are run in each of the three networks. The results are shown in four graphs indicating the accuracy and running time. The comparison of the three networks can be gained from the graphs to help determine which one is suitable.

As for the methodology used in this invention, there are advantages in both of the data and algorithm choices. Our data includes stock-related as well as social economic features. The addition of social economic features makes our prediction more well-rounded and accurate. All of the social economic features are carefully selected because they are highly related to finance instead of randomly chosen from a variety of unrelated fields. Other than the data selection, data processing has also gone through careful manipulation. Considering the effect of the extreme values, we normalize our data. Although no adjustments are made to the networks, a comparison of the three networks is conducted to give the investors a clearer reference on which method to choose. Under the comparison, accuracy and running time were plotted on graphs. In terms of CNN, it has the relatively highest accuracy and stability. As for GRU, it is relatively the fastest among the networks. LSTM is neither the fastest nor the most accurate, but it has a relatively smaller fluctuation than GRU.

When applied to real stock movement prediction for investors, this invention benefits investors in many ways. The design of the invention provides the investors with a piece of straightforward advice on whether investing the stock today will make profits in the future five days. The comparison among CNN, GRU and LSTM helps investors determine which network to use based on their need. For instance, if an investor is urgent to make the investment, since this invention shows that GRU has the fastest prediction speed, it’s clear that GRU is the most suitable choice for him or her. Other benefits include an allowance of high volume data, a valid prediction accuracy and so on.

Carefully selected data, distinctively shown comparison and other advantages give this invention the confidentiality to provide investors with a useful reference on stock trend prediction. Hence, investors will be more confident about their choices.

DESCRIPTION OF THE DRAWINGS

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

Fig.l is a flow chart showing the general procedure, according to the present invention.

Fig.2 is a structure diagram showing the algorithm of LSTM.

Fig3. is a structure diagram showing the algorithm of GRU.

Fig4. is a structure diagram showing the algorithm of CNN.

Fig5. is a bar chart showing the average running time of three neural networks.

Fig6. is a bar chart showing the different stocks’ prediction accuracy trained by CNN.

Fig7. is a bar chart showing the comparison of LSTM accuracy with or without social economic features.

Fig8. is a bar chart showing the comparison of GRU accuracy with or without social economic features.

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. STEP A: For data collection, stock-related factors are collected: close price, percent change, transaction amount, and dividend rate. Also, social economic features are gathered: WTI price, USDX, gold price based on US dollar and WTI futures price. Our oil indexes are collected from XOM, SNP, PTR and PBR. STEP B: (1) Clean data: Wipe off useless information such as market closure and stop quotation. (2) In order to eliminate the influence of extreme data on the other data, we normalized our data by the standard deviation normalization formula. (3) Add labels to the data according to rising situation in five days. If there is one day the price rises, a “1” is added as the label, otherwise, a “0” is added instead. Because the ratio of the labels is not equally distributed, we assigned a rising percentage with which it can be tagged as “1”. (4) The data is divided into two parts, training set and testing set. STEP C: CNN and RNN are used in this invention. Since GRU and LSTM are two advanced versions of RNN, they are chosen for our invention. Data is run in each of the network corresponding to the sequence of LSTM, GRU and CNN. In order to get a more effective result, parameters such as input size, output size, max step, learning rate and batch size are adjusted according to the previous training results. For CNN, its convolutional kernel size, pooling layers and fully connected layers are adjusted. STEP D: Each of the four stock-related factors is matched with each of the four social economic features to form 16 pairs of conditions. Each of the 16 conditions is then run in each of the three networks for the four stocks (XOM, SNP, PTR and PBR) to generate the prediction results. For time using, a bar chart (Fig. 5) shows the running time of each of the network. For accuracy, a graph (Fig. 6) shows CNN accuracy for four stocks, which has the same accuracy with or without the 4 features; two graphs (Fig. 7-8) show a comparison of LSTM and GRU accuracy with or without the 4 features.

The above shows the embodiment of the invention’s procedure. This invention provides investors with a new method to forecast the stock trend so that they can make a wise investment to achieve an ideal profit.

Claims

1. A comprehensive stock trend predicting method based on neural networks, which is the stock prediction with not only stock-related factors, but also social economic features based on LSTM, GRU and CNN which belong to ANN as main algrithms, forecasting whether investing a stock will make profits in the next five days.