zubovskiii98 / stock-price-prediction-with-bert Goto Github PK

The post is about how I used BERT model with news articles to predict DJIA's adjusted close prices which range from 2008 to 2016.

There are total of 1989 samples of data and each day contains 25 top news articles, gathered by Aaron7sun and has the following structure.

import pandas as pd

djia = pd.read_csv('data/Combined_News_DJIA.csv')

djia.head()

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

I started preprocessing all the articles as it is in string format and contains unnecessary starting characters b' or b".

Once it is done, I passed them to BERT to get vectors of numeric values which then were fed into a neural network to compute the coefficient indicating how much news articles affect a day's open price and the predicted close values.

I used five different models and following plots are predicted values.

They have high variances but it can be seen that they were able to capture general trends so what one can do to improve the models is to set a threshold so that the close price does not change drastically over a night.

The posts include all the codes I used with explanations and are divided into two separate pages.

1. Part 1
2. Part 2

All models are implemented using Gluon and Tensorflow.