Department of Information Systems, Business Statistics & Operations Management
Deep Learning for Stock Prediction
Supervisor: KWOK James S H / ISOM
Student: JI Xiayan / ECOF Course: UROP1100, Summer
It has been a prevailing trend in recent years to apply machine learning algorithm in a financial context. With randomness in moving direction and abundance in available information, stock price and market trend prediction have attracted considerable attention. Given its gorgeous performance in recognizing the patterns of various kinds of input variables and project the future performance of stocks and markets, artificial neural network, comprised of many different categories of networks, constitutes an interesting direction to look at. In this report, a Multilayer Perceptron (MLP) and a Long Short-term Memory network (LSTM) were constructed to preliminarily generate some insights with the purpose of further improving the prediction accuracy and address some questions came across during the research process.
Deep Learning for Stock Prediction
Supervisor: KWOK James S H / ISOM
Student: JUNAEDI Victoria Amanda / ECOF Course: UROP1100, Summer
Using deep learning methodologies in trading is an increasingly popular field. Numerous researchers have tried to predict the stock market by using sentiment analysis, time-series data, as well as other fundamental or technical indicators. I am currently exploring the possibilities of predicting stock prices using textual analysis and the respective historical prices, specifically in Hong Kong Stock Exchange. Particularly, this report will show the preliminary study of using deep learning for natural language processing (NLP) and time-series data, on top of which will be trained another set of neural networks to classify the stock price movement. In addition, this report will compare different types of neural network to use and discuss which works best for both textual and numerical analysis.
Risk Management
Supervisor: SO Mike Ka Pui / ISOM
Student: HE Jianyao / QFIN Course: UROP1100, Spring
This project is about implementing Bayesian analysis into latent space model to study network relation. The
invisibility of actors in network model makes it difficult to implement in daily life and latent space model is used as it can visualize the position of actors enabling us to study actors' interaction by actors' distributions and ties
(interactions). To study the possible changes in actors' position after updating current information, Bayesian analysis offers such a way to estimate the unknown parameters of latent space models. Over the past few weeks, I read papers about the latent space model and then further read on Bayesian analysis using Markov chain Monte Carlo
Department of Information Systems, Business Statistics & Operations Management simulation. I also performed model fitting for classic network data using R.
Risk Management
Supervisor: SO Mike Ka Pui / ISOM
Student: WANG Ziyi / RMBI Course: UROP1100, Spring
UROP2100, Summer
The research goal of this stage is to estimate the parameters in the proposed cross-sectional vine copula factor model. Professor Mike So has proposed a cross-sectional vine copula model to incorporate correlations among high dimensional high frequency stock return series and market factors. Based on the rationale behind and supervisor’s guidance, I continued to test the model by doing simulation studies. I began from low dimensional data and assume covariance and dependencies are constant, and then expand to high dimensional data, finally further involve the dynamic characteristic of the parameters. The report will summary my research progress in mainly three parts:
Understand the model and estimation algorithm in a theoretical base, practical execution, and further improvement of the model.
Application of Deep Learning (Artificial Intelligence) on Natural Language Processing Supervisor: ZHANG Xiaoquan / ISOM
Student: CHUNG Jihoon / COMP Course: UROP1100, Spring
This research shows possibilities of using deep learning to predict stock price. This research uses two different models to predict HS300 index. First model has one hidden layer, and uses stock history. Second model uses text data from Chinese stock-related forum. First model has shown 86.6% of accuracy, where second model has showed 55.4%.
Both results are then combined with linear regression to see the result, which was 83.4%. It was possible to see that deep learning could be used to predict stock price. However, deep learning model with text data has to be
significantly improved to have meaningful result.
Application of Deep Learning (Artificial Intelligence) on Natural Language Processing Supervisor: ZHANG Xiaoquan / ISOM
Student: HUANG Chen / QFIN Course: UROP1100, Spring
The project aims to find arbitrage situation where stock price jumps or falls due to catalyst event caught by social media including Weibo, Facebook, Twitter. These online social media provide platforms for users to express their feelings including their attitude towards a company or a specific stock. We do not aim to make a long/short decision based on media user’s opinion of a stock, we use their underlying opinions, their sentiments implicated by the verbs and adjectives, instead. Based on data crawled from these social platforms, we use text analysis to transfer these text data into 40 different sentiment indexes. By observing the trend of the index and the degree they change, we find
Department of Information Systems, Business Statistics & Operations Management out if there is a sudden unnatural jump or fall of the data. These unnatural patterns can indicate a big hit or an event that triggers the users’ emotions that they may take actions and meanwhile affect the whole stock market. There have already been some cases before which will be mentioned in the related studies and analysis section. We try to use the 2-year data of social media to find out more about similar cases to test our assumption. And based on the assumptions, we build up models to visualize the fluctuate pattern of users’ sentiments.
Application of Deep Learning (Artificial Intelligence) on Natural Language Processing Supervisor: ZHANG Xiaoquan / ISOM
Student: SRA Jai Singh / COMP Course: UROP1100, Spring
This report describes the progress attained by me in trying to implement topic detection in Chinese documents via Latent Dirichlet Allocation (LDA), an unsupervised learning technique that looks at relative frequencies of words across documents and accordingly puts them in clusters. The report looks into common problems encountered in pre-processing of Chinese data to use in the model, making sense of the results outputted by the model and measuring performance in unsupervised learning tasks.