Cryptocurrency Returns Predictor

An Application of Random Forest!

Project Description

Introduction

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

• Abstract: Use sentiment-based features to predict cryptocurrency returns. Models used: Random Forest Classifier, Random Forest Regressor, and VAR time-series model. Analysis timeframe: 28/11/2014 - 25/07/2020.

• Status: Completed.

Methods Used

• Random Forests (Regressor & Classifier)
• Principal Component Analysis
• Vector Autoregression (VAR) model
• Sentiment Indicators (retrieved from my graduation thesis)

Dependencies

• Python 3
• numpy==1.18.5
• pandas==1.0.5
• scikit-learn==0.23.2
• statsmodels==0.12.0
• plotly==4.9.0

Interesting Results to Keep You Reading

Backtesting strategies based on 3 models:

• Generate trading signals: Long as predicted return > 0, short as predicted return < 0, wait otherwise.
• Test period (25% of the dataset): 05/03/2019 - 25/07/2020
• RF Classifier outperforms significantly both strategies and also the simple buy-and-hold strategy.
• Download the interactive version.

Table of Contents

• Project Description
  • Introduction
  • Methods Used
  • Dependencies
  • Interesting Results to Keep You Reading
• Table of Contents
• Getting Started
  • How to Run
  • Dependent Variable/Target
  • Sentiment Measures
  • Features Selection
• Results (Test Period)
• Read More

Getting Started

How to Run

1. Clone this repo:
   git clone https://github.com/dang-trung/crypto-return-predictor

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

   Or (conda):
   conda env create -f environment.yml
conda activate crypto-return-predictor

3. Run in terminal:
   python -m crypto_return_predictor

Dependent Variable/Target

Cryptocurrency market returns (computed using the market index CRIX, retrieved here, see more on how the index is created at Trimborn & Härdle (2018) or those authors' website.)

Sentiment Measures

• Sentiment score of Messages on StockTwits, Reddit Submissions, Reddit Comments
  • Computed using dictionary-based sentiment analysis, lexicon used: crypto-specific lexicon by Chen et al (2019), retrieved at the main author's personal page.
  • StockTwits messages are retrieved through StockTwits Public API, Reddit data are retrieved using PushShift.io Reddit API.
• Messages volume on StockTwits, Reddit Submissions, Reddit Comments.
• Market volatility index VCRIX (see how the index is created: Kolesnikova (2018), retrieved here.)
• Market trading volume (retrieved using Nomics Public API)

Read more on how I retrieve these sentiment measures in my graduation thesis or its Github repo.

Features Selection

• For VAR model: Lagged values of the first principal component of all 9 sentiment measures (up to 5 lags).
• For Random Forests: Sentiment measures' lagged Values (up to 5 lags).

Results (Test Period)

Order by performance (from high to low):

1. Random Forest Classifier:

• Accuracy: 61.86%
• Confusion matrix:

		Actual
		Negative	Unchanged	Positive
Predicted	Negative	145	0	97
	Unchanged	1	0	0
	Positive	96	0	170

• Backtesting daily returns: ~91bps

2. VAR(5):

• Accuracy: 54.62%
• Confusion matrix:

		Actual
		Negative	Unchanged	Positive
Predicted	Negative	57	0	185
	Unchanged	0	0	1
	Positive	45	0	221

• Backtesting daily returns: ~48bps

3. Random Forest Regressor:

• Accuracy: 56.19%
• Confusion matrix:

		Actual
		Negative	Unchanged	Positive
Predicted	Negative	222	0	20
	Unchanged	1	0	0
	Positive	202	0	64

• Backtesting daily returns: ~19bps (just slightly better than holding the CRIX index)

Read More

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