Prédiction de la tendance des actions basée sur les réseaux convolutifs graphiques et les LSTM

As stocks have been developing over decades, the trend and the price of a stock are more often used for predictions in stock market analysis. In the field of finance, an accurate stock future trending can not only help decision-makers estimate the possibility of profit, but also help them avoid risks. In this research, we present a quantitative approach to predicting the trend of stocks in which a clustering model is employed to mine the stock trends patterns from historical stock price data. Stock series clustering is a special kind of time series clustering. We aim to find out the trend types, e.g. rising, falling and others, of a stock at time intervals, and then make use of the past trends to predict its future trend. The proposed prediction method is based on Graph Convolutional Neural Network for clustering and Long Short-Term Memory model for prediction. This method is suitable for the data clustering of unbalanced classes too. The experiments on real-world stock data demonstrate that our method can yield accurate forecasts. In the long run, the proposed method can be used to explore new possibilities in the research field of time series clustering, such as using other graph neural networks to predict stock trends.

Name thesis:degree_name

M. Sc.

Level thesis:degree_level

Maîtrise

Discipline thesis:degree_discipline

Informatique

Grantor dc:publisher

Université de Sherbrooke

Year dc:date.issued

2022

Author dc:creator

• Sun, Mingxuan

Advisor dc:contributor.advisor

• Wang, Shengrui

• Time series
• Clustering
• Finance data prediction
• LSTM
• Graph convolutional network
• Séries temporelles
• Prédiction de données financières

Language dc:language.iso

en

Handle dc:identifier.uri

http://hdl.handle.net/11143/19605

OAI identifier oai:identifier

oai:usherbrooke.scholaris.ca:11143/19605