Representation Learning for Financial Time-Series Forecasting

Table of Contents

• Introduction
• Features
• Installation
• Contents
• Results

Introduction

This repository contains the implementation of my MSc project titled Representation Learning for Financial Time-Series Forecasting. The primary goal of this project is to enhance the accuracy of financial time-series forecasting by leveraging Contrastive Predictive Coding (CPC) to generate feature embeddings.

Features

• Contrastive Predictive Coding (CPC): Utilised for feature extraction from financial time series, improving downstream model performance.
• Baseline Models: Includes Persistence, Zero, Mean, Linear Regression and Buy-and-Hold models for performance benchmarking.
• LSTM Model Implementation: For comparison with CPC-enhanced models.
• Sharpe Ratio Optimization: Evaluates model performance using Sharpe ratios.

Installation

To get started with this project, clone the repository and set up the environment using conda and pip.

Step 1: Clone the Repository

git clone https://github.com/ese-msc-2023/irp-agk123.git
cd irp-agk123

Step 2: Set Up the Conda Environment

Create a new Conda environment using the provided environment.yml file (note the environment is mac users on arm architecture who require tensorflow-macos):

conda env create -f environment.yml
conda activate cpc_financial

Contents

• deliverables: Deliverables for my project. Inlcudes the project plan, final report and presentation.
• images: Images utilised in the deliverables.
• other_notebooks: Tried and failed attempts and different approaches.
• src: Where the main notebook lives. Utils folder with utils functions.
• tests: Tests for my utility functions and CPC data generator that are automatically run through the CI github action.

Results

The below results all beat the classic benchmark of buy-and-hold. For more details, visit the CPC notebook.

USDJPY: Sharpe of 1.3 versus 0.53 using Buy-and-Hold Benchmark

USDSGD: Sharpe of 0.98 versus -0.35 using Buy-and-Hold Benchmark

EURGBP: Sharpe of 0.74 versus -0.12 using Buy-and-Hold Benchmark