Financial forecasting using a BERT based sentiment index

This work is based on a Master Thesis by Manuel Tonneau which in turn is based on:

Hiew, Joshua & Huang, Xin & Mou, Hao & Li, Duan & Wu, Qi & Xu, Yabo. (2019). BERT-based Financial Sentiment Index and LSTM-based Stock Return Predictability.

You may use this repository as a foundation to develop alternative approaches and additional functionalities. To run this code, execute run.py or edit the file to configure it to your liking. I recommend using a code editor such as Visual Studio Code with the python plugin or any other python capable editor to work on this project. The application also requires a Dataset which is not public and will have to be requested if you work on this topic as your master thesis.

Requirements

The following package versions have been used to develop this work. Newer versions will likely work as well, older versions might not.

Python 3.6.8

tensorflow-gpu==2.0.0 (you may use regular tensorflow, tf 2.1 should be fine as well)
pandas==0.24.2
nltk==3.4.5
scikit-learn==0.21.2
statsmodels==0.10.2
pytorch-pretrained-bert==0.6.2
torch==1.3.1
fastai==1.0.59
beautifulsoup4==4.8.1

Acknowledgements

Data:

Malo, Pekka & Sinha, Ankur & Takala, Pyry & Korhonen, Pekka & Wallenius, Jyrki. (2013). FinancialPhraseBank-v1.0.

http://iborate.com/usd-libor/ (USD libor)

https://mba.tuck.dartmouth.edu/ (Fama/French 5 Factors)

https://finance.yahoo.com/ (AAPL and USDX)

Code:

Elisabeth Bommes (2016) https://github.com/QuantLet/TXT/tree/master/TXTfpblexical

https://www.machinelearningplus.com/time-series/vector-autoregression-examples-python/