StockTwits Sentiment Classifier

An Application of Random Forest!

Project Description

Introduction

• Objective: Project for my intern at Research Center VERA, Ca' Foscari University of Venice.

• Abstract: 2,045,322 cryptocurrency-related Tweets (~287MB) are retrieved using StockTwits API. The messages are posted from 28/11/2014 to 25/07/2020. Nearly half of those messages are labelled with sentiment (i.e. Bullish/Bearish). Based on the labeled dataset, a Random Forest model is then trained to classify the sentiments of Tweets about cryptocurrencies, resulting in a 74.75% prediction accuracy on test set.

• Status: Completed.

Methods Used

• Text-processing, inspired by Renault (2017) and Chen et al. (2019).
• TF-IDF (for text-vectorization).
• Truncated SVD (for dimension reduction).
• Random Forest.

Dependencies

• Python 3
• numpy==1.18.5
• pandas==1.0.5
• scikit-learn==0.23.2
• requests==2.24.0

Table of Contents

• Project Description
  • Introduction
  • Methods Used
  • Dependencies
• Table of Contents
• Getting Started
  • How to Run
  • Data Storage
• Results
• Read More

Getting Started

How to Run

1. Clone this repo: git clone https://github.com/dang-trung/stocktwits-sentiment-classifier

2. Create your environment (virtualenv):
   virtualenv -p python3 venv
source venv/bin/activate (bash) or venv\Scripts\activate (windows)
   (venv) cd stocktwits-sentiment-classifier
(venv) pip install -e

   Or (conda):
   conda env create -f environment.yml
conda activate stocktwits-sentiment-classifier

3. Run in terminal:
   python -m sentiment_classifier
   Note that due to API limits, it will take several days to fully download all 2m+ cryptocurrencies-related Tweets on StockTwits from 2014 to 2020.

Data Storage

1. Downloaded messages will be stored in data/01_raw.
2. Messages after being processed (so that only information relevant to sentiment) will be stored in data/02_processed.
3. Vectorized text messages are stored in data/03_vectorized (since this file is small compared to the files generated by step 1 and 2, I already included this in the repo.)
4. External files (symbols of cryptos & rules for text-processing) are stored in data/04_external

Results

• Model parameters: ntree=500, max_depth=20, max_samples=0.75
• Confusion matrix of training set

		Actual Classes
		Bearish	Bullish
Predicted Class	Bearish	82,208	8,426
	Bullish	5,269	85,365

• Confusion matrix of test set (~74.75% accuracy)

		Actual Classes
		Bearish	Bullish
Predicted Class	Bearish	59,888	30,747
	Bullish	175,937	551,880

Read More

For better understanding of the project, kindly read the report.