Fine tuning of DeepMoji on StockTwits messages

Introduction

This projects fine-tunes the DeepMoji classifier using the self-reported sentiments of StockTwits messages.

Set-up

The project builds upon the code of Felbo et. al. which can be found here. Detail instructions for setting up their model can be found directly at their GitHub repository.

For this partiuclar project, using Miniconda with Python 3.7, a virtual enviroment of Python 2.7 was set-up:

conda create -n deepmoji python=2.7

In this virtual enviroment, install a theano. Then navigate to the root folder where the DeepMoji files are located and run the installation file:

pip install -e .

Next the backend of keras has to be changed: either change in .keras/keras.JSON the backend from "tensorflow" to "theano" or create a new folder .keras_python27 (as a copy of .keras/) and change there the file keras.JSON. In latter case, go to path_to_miniconda\Miniconda3\envs\deepmoji\Lib\site-packages\keras\backend\__init__.py and change the location of the keras setting file to point to the newly created folder. Finally, if when running the scripts an error about the ifelse function appears, in the file where the error occurs add at the beggining from theno import ifelse and replace in the code theano.ifelse with ifelse.

Project structure

The projects consist of mainly four folders:

• data: this folder contains only the script that converts the StockTwits data to a dictionary with the structure info, texts, val_id, test_id, train_id and saves them as 'pickle'-file. The
  actual data is stored locally on the machine. The folder has the following structure:
  • 00_raw: the raw data and help-data
  • 01_processed: the output of the sentiment prediction, i.e. the daily sentiment
  • 02_models: the estimated models saved as JSON files
  • 03_model_output: csv-file with the model performance (accuracy, F1-score)
• deepmoji: files as in the original DeepMoji repository
• scripts: contains the scripts used to run the estimation and prediction
• model: files as in the original DeepMoji repository

Estimation details

The script 'fit_stocktwits.py' fine-tunes the pre-trained DeepMoji model of Felbo et. al. (2017 on the self-reported bullishness or bearishness of StockTwits messages. The models are estimated using StockTwits messages mentioning one of 360 US stocks of the S&P 500. For the estimation only messages from 2013-06-01 to 2014-08-31 are used; this matches the data estimation window used by Renault (2017). An undersampling technique is then used to create a balanced train data set.

The fine-tuning on StockTwits messages starts by increasing the number of words in the embedding vocabulary. Then, the output layer of the neural network is replaced to account for the fact that we have only two labels in the StockTwits sentiment classification task. The fine tuning is then done by using the 'chain-saw' approach suggested by Felbo et. al. (2017).