A deep learning approach with extensive sentiment analysis for quantitative investment

Section 1: News Sentiment Analysis

Inside the news_sentiment_analysis folder, including:

• News collecting from East Money (www.eastmoney.com)
• News contents summarizing with pretrained Pegasus
• Sentiment model finetuning with ChnSentiCorp dataset
• Sentiment analyzing for merged news titles and summarized contents

Section 2: Deep Learning Model Building

Inside the modeling folder:

1. Deep Learning Model for Experiments

There are six subfolders, each corresponding to one of the six experimental groups, as follows:

	LSTM	Transformer
Title + Content	MAS_lstm_enhancement/	MAS_transformer_enhancement/
Title	MAS_lstm_enhancement_title/	MAS_transformer_enhancement_title/
Vanilla	MAS_lstm/	MAS_transformer/

2. Subfolder and File Meanings for Each Experimental Group

Inside each experimental group subfolder:

• backtrader_sequence_model.py: Code for deep learning model to predict stock price movements.
• run.py: Script to run backtrader_sequence_model.py.
• output/: Results of the deep learning model predictions.
• backtrader_mystrategy.ipynb: Code for backtesting strategies.
• results/: Results of strategy backtesting.
• plot.py: Code to plot strategy returns and benchmark comparisons, with image paths in figs/.
• procedure/: Records training process metrics for the top 50 stocks.

3. Additional Files

Inside the MAS-2023-code/ path:

• train_procedure.py: Code for plotting metrics for the top 50 stocks.
• evaluate.py: Code for calculating metrics for the top 50 stocks.