Applying Improved Deep Capsule Network to Analyze Limit Order Books for Day Trading

碩士 === 國立交通大學 === 資訊管理研究所 === 108 === Day trading allows investors to avoid the risk of holding shares overnight, and only needs to pay transaction fees and the difference of price between the entry and exit point when losing money, without actually bearing the stock price, so it has high rewards ca...

Full description

Bibliographic Details
Main Authors: Chang, Ya-Ting, 張雅婷
Other Authors: Chen, An-Pin
Format: Others
Language:zh-TW
Published: 2019
Online Access:http://ndltd.ncl.edu.tw/handle/s5b28v
id ndltd-TW-108NCTU5396004
record_format oai_dc
collection NDLTD
language zh-TW
format Others
sources NDLTD
description 碩士 === 國立交通大學 === 資訊管理研究所 === 108 === Day trading allows investors to avoid the risk of holding shares overnight, and only needs to pay transaction fees and the difference of price between the entry and exit point when losing money, without actually bearing the stock price, so it has high rewards caused by high leverage. These people who conduct day trading are called day traders. But day trading also has high risks. Once day traders lose the opportunity to buy and sell, they may have greater losses. Therefore, it is very important to predict the short-term stock price movement more accurately. However, the factors affecting the short-term stock price are quite complex, including the influence of the same industry stocks, investor sentiment, and the behavior of institutional investors, etc., so it is difficult to find out the rules of the stock price movement. Day traders may not only lose the optimal entry and exit price but also be affected by the spoofing traders, thus making the wrong trading strategy. With the electronic trading system, limit order books (LOBs) records a lot of information about high-frequency unexecuted limit orders, which is the driving force for short-term price fluctuations. Therefore, some studies have attempted to use LOBs to analyze the trading intentions of investors in the market. However, statistical models are difficult to handle data with high dimensional, complex, and nonlinear characteristics. In recent years, with the breakthrough development of computer computing strength, machine learning approaches using algorithms to analyze data has risen a lot. The machine can quickly process large amounts of data and automatically analyze the rules from it. Machine learning methods increase the possibility of short-term price forecasts, and more and more research has attempted to apply machine learning to the financial domain. At present, some studies have used Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) to model LOBs. However, they did not deal well with the spatial relationship and temporal relationship between the trades and unexecuted limit orders that have a large impact on short-term prices. Besides, the max-pooling method used in CNNs to aggregate information will cause the loss of spatial information and the lack of equivariance, leading to misclassification of financial time series. The CapsNet is an innovative network proposed by Sabour et al.. The output of the capsule is a vector rather than a scalar, and retention of feature spatial information and equivariance are achieved by the dynamic routing process. The performance of CapsNets through the prediction agreement of the dynamic routing is better than that of CNNs, and its output as a vector can have better feature representation ability. This thesis proposes a network architecture designed with the characteristics of trades information and LOBs. The CNNs and LSTM are used to extract the spatial relationship and time dependencies of the input features. Because the financial data is quite complex, we use capsules to encode features, which can improve feature representation. On the other hand, because of the high variability of financial time-series, the dynamic routing process of CapsNets can model the input time-series one by one to achieve the agreement of prediction. In addition, the temporal regulation is added during the training process to conform to the characteristics of the financial time-series, so that the model has better prediction performance in the classification of the bid-ask spread price spread crossing. The experimental results show that our proposed model MT-CapsNets can achieve an improvement of precision up to 3.39% compared with the best baseline model. TR-MT-CapsNets, which is based on adding a temporal regulation term to loss function of original CapsNets for the financial time series, contributes to the improvement result of the MT-CapsNets, increasing 2.74% in precision.
author2 Chen, An-Pin
author_facet Chen, An-Pin
Chang, Ya-Ting
張雅婷
author Chang, Ya-Ting
張雅婷
spellingShingle Chang, Ya-Ting
張雅婷
Applying Improved Deep Capsule Network to Analyze Limit Order Books for Day Trading
author_sort Chang, Ya-Ting
title Applying Improved Deep Capsule Network to Analyze Limit Order Books for Day Trading
title_short Applying Improved Deep Capsule Network to Analyze Limit Order Books for Day Trading
title_full Applying Improved Deep Capsule Network to Analyze Limit Order Books for Day Trading
title_fullStr Applying Improved Deep Capsule Network to Analyze Limit Order Books for Day Trading
title_full_unstemmed Applying Improved Deep Capsule Network to Analyze Limit Order Books for Day Trading
title_sort applying improved deep capsule network to analyze limit order books for day trading
publishDate 2019
url http://ndltd.ncl.edu.tw/handle/s5b28v
work_keys_str_mv AT changyating applyingimproveddeepcapsulenetworktoanalyzelimitorderbooksfordaytrading
AT zhāngyǎtíng applyingimproveddeepcapsulenetworktoanalyzelimitorderbooksfordaytrading
AT changyating jīyúgǎiliángshìshēndùjiāonángwǎnglùfēnxīxiànjiàwěituōbùfāzhǎndāngchōngjiāoyìcèlüè
AT zhāngyǎtíng jīyúgǎiliángshìshēndùjiāonángwǎnglùfēnxīxiànjiàwěituōbùfāzhǎndāngchōngjiāoyìcèlüè
_version_ 1719296667585347584
spelling ndltd-TW-108NCTU53960042019-11-26T05:16:55Z http://ndltd.ncl.edu.tw/handle/s5b28v Applying Improved Deep Capsule Network to Analyze Limit Order Books for Day Trading 基於改良式深度膠囊網路分析限價委託簿發展當沖交易策略 Chang, Ya-Ting 張雅婷 碩士 國立交通大學 資訊管理研究所 108 Day trading allows investors to avoid the risk of holding shares overnight, and only needs to pay transaction fees and the difference of price between the entry and exit point when losing money, without actually bearing the stock price, so it has high rewards caused by high leverage. These people who conduct day trading are called day traders. But day trading also has high risks. Once day traders lose the opportunity to buy and sell, they may have greater losses. Therefore, it is very important to predict the short-term stock price movement more accurately. However, the factors affecting the short-term stock price are quite complex, including the influence of the same industry stocks, investor sentiment, and the behavior of institutional investors, etc., so it is difficult to find out the rules of the stock price movement. Day traders may not only lose the optimal entry and exit price but also be affected by the spoofing traders, thus making the wrong trading strategy. With the electronic trading system, limit order books (LOBs) records a lot of information about high-frequency unexecuted limit orders, which is the driving force for short-term price fluctuations. Therefore, some studies have attempted to use LOBs to analyze the trading intentions of investors in the market. However, statistical models are difficult to handle data with high dimensional, complex, and nonlinear characteristics. In recent years, with the breakthrough development of computer computing strength, machine learning approaches using algorithms to analyze data has risen a lot. The machine can quickly process large amounts of data and automatically analyze the rules from it. Machine learning methods increase the possibility of short-term price forecasts, and more and more research has attempted to apply machine learning to the financial domain. At present, some studies have used Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) to model LOBs. However, they did not deal well with the spatial relationship and temporal relationship between the trades and unexecuted limit orders that have a large impact on short-term prices. Besides, the max-pooling method used in CNNs to aggregate information will cause the loss of spatial information and the lack of equivariance, leading to misclassification of financial time series. The CapsNet is an innovative network proposed by Sabour et al.. The output of the capsule is a vector rather than a scalar, and retention of feature spatial information and equivariance are achieved by the dynamic routing process. The performance of CapsNets through the prediction agreement of the dynamic routing is better than that of CNNs, and its output as a vector can have better feature representation ability. This thesis proposes a network architecture designed with the characteristics of trades information and LOBs. The CNNs and LSTM are used to extract the spatial relationship and time dependencies of the input features. Because the financial data is quite complex, we use capsules to encode features, which can improve feature representation. On the other hand, because of the high variability of financial time-series, the dynamic routing process of CapsNets can model the input time-series one by one to achieve the agreement of prediction. In addition, the temporal regulation is added during the training process to conform to the characteristics of the financial time-series, so that the model has better prediction performance in the classification of the bid-ask spread price spread crossing. The experimental results show that our proposed model MT-CapsNets can achieve an improvement of precision up to 3.39% compared with the best baseline model. TR-MT-CapsNets, which is based on adding a temporal regulation term to loss function of original CapsNets for the financial time series, contributes to the improvement result of the MT-CapsNets, increasing 2.74% in precision. Chen, An-Pin Huang, Szu-Hao 陳安斌 黃思皓 2019 學位論文 ; thesis 68 zh-TW