DEEP REINFORCEMENT LEARNING FOR AUTOMATED CRYPTOCURRENCY TRADING

• Main Author: Shan, Elbert
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/66647
• Institution: Institut Teknologi Bandung

This research presents a deep reinforcement learning model for algorithmic trading of cryptocurrencies. The aim of the model is to help traders earn greater profits than traditional strategies. Traditional strategies can be used to gain profits. However, this strategy tends to require more knowledge, experience and time compared to machine learning-based solutions in order to generate optimal returns. Models will be trained to trade on the cryptocurrency market. Model inputs are 1 minute interval candlestick data and technical indicators for BTC/USDT cryptocurrency pair. The output of the model is buy, hold, or sell signals. The model development was carried out based on the CRISP-DM methodology. The model will be created with PPO algorithm and a custom environment that follows gym interface. The transaction fee rate of 0.1% will be taken into consideration by the model in determining the strategy. The model will be trained and tested in constant-sized episodes of 720 timestep minutes (12 hours). The trading positions that can be done by the model are neutral (owning 0 amount of assets) and long. Reward for the model will be calculated based on profit earned on each position. The reward system will also include a certain penalty system to prevent the model from always implementing Buy and Hold strategy or never going long. The performance of the model is improved by optimizing the hyperparameter values using Optuna. The performance of the model is compared to the Buy and Hold strategy. The tests results show that the model produced still cannot beat the Buy and Hold strategy.