newtoniano / tradetheevent

Implementation of "Trade the Event: Corporate Events Detection for News-Based Event-Driven Trading." In Findings of ACL2021

TradeTheEvent

This repository contains

1. Official implementation of the following paper:

Zhihan Zhou, Liqian Ma, Han Liu. Trade the Event: Corporate Events Detection for News-Based Event-Driven Trading. In Findings of ACL 2021.

2. The EDT dataset for corporate event detection and news-based stock prediction benchmark.
3. Tools for scraping all the news articles from Reuters back to 2017 and interactively analyze any online news articles (Including event detection and sentiment analysis).

If you use any of them in your work, please cite our paper.

Dataset

We release the EDT dataset for corporate event detection and news-based stock prediction benchmark. Please refer to the data/ folder for access and detailed information about this dataset. The dataset can be found here.

Tool

We shall the tool for scraping news article and interactively analyze news articles on how it may influence the stock market. Please refer to the tool/ folder for access and detailed information about our tool.

Environment

We recommand to use a Python virtual environment with Python >= 3.6. The requirements can be installed with:

git clone https://github.com/Zhihan1996/TradeTheEvent
cd TradeTheEvent
pip install -r requirements.txt

Official Implementation

0. Download Data

Please download data from here and put all the three data folder into a folder. Suppose the data dir is /home/user/data, set the environment variable as:

export DIR_TO_EDT_DATASET=/home/user/data

1. Domain Adaptation

export TRAIN_FILE=DIR_TO_EDT_DATASET/Domain_adaptation/train.txt
export TEST_FILE=DIR_TO_EDT_DATASET/Domain_adaptation/dev.txt
export ADA_MODEL_DIR=models/bert_bc_adapted

python run_domain_adapt.py \
    --output_dir=ADA_MODEL_DIR \
    --model_type=bert \
    --model_name_or_path=bert-base-cased \
    --do_train \
    --train_data_file=$TRAIN_FILE \
    --do_eval \
    --eval_data_file=$TEST_FILE \
    --mlm \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 2 \
    --warmup_steps 500 \
    --learning_rate 3e-5 \
    --evaluate_during_training \
    --eval_steps 500 \
    --num_train_epochs 20 \
    --max_steps 10000

2. Train the Model for Event Detection

export ADA_MODEL_DIR=models/bert_bc_adapted
export DATA_DIR=DIR_TO_EDT_DATASET/Event_detection
export OUTPUT_DIR=models/bilevel

python run_event.py \
    --TASK bilevel \
    --data_dir DATA_DIR \
    --epoch 5 \
    --model_type ADA_MODEL_DIR \
    --output_dir OUTPUT_DIR \
    --bert_lr 5e-5  \
    --per_gpu_batch_size 8 \
    --gradient_accumulation_steps 1 \
    --max_seq_length 256 

3. Detect Events on the Evaluation News

export OUTPUT_DIR=models/bilevel
export PRED_DIR=preds/bilevel
export DATA_DIR=DIR_TO_EDT_DATASET/Trading_benchmark/evaluate_news.json


python run_event.py \
    --TASK bilevel \
    --output_dir OUTPUT_DIR \
    --per_gpu_batch_size 64 \
    --predict_dir PRED_DIR \
    --do_predict \
    --max_seq_length 256 \
    --data_dir DATA_DIR

4. Backtest

export PRED_DIR=preds/bilevel
export DATA_DIR=DIR_TO_EDT_DATASET/Trading_benchmark/evaluate_news.json
export RESULTS_DIR=results/

python run_backtest.py \
    --evaluate_news_dir DATA_DIR \
    --pred_dir PRED_DIR \
    --save_dir RESULTS_DIR \
    --model_type bilevel \
    --seq_threshold 5 \
    --stoploss 0.2 \
    --buy_pub_same_time

The final backtest results will appear at the RESULTS_DIR.

Reference

If you use the codes, tool, or the dataset, please kindly cite our paper.

@inproceedings{zhou-etal-2021-trade,
    title = "Trade the Event: Corporate Events Detection for News-Based Event-Driven Trading",
    author = "Zhou, Zhihan  and
      Ma, Liqian  and
      Liu, Han",
    booktitle = "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-acl.186",
    doi = "10.18653/v1/2021.findings-acl.186",
    pages = "2114--2124",
}