Description:
Despite its high risk, cryptocurrency has gained popularity as a successful trading option. Cryptocurrencies are digital assets that fluctuate dramatically in a market that operates for 24 hours. Developing trading bots using machine learning-based artificial intelligence (AI) approaches has recently received considerable attention. Previous studies have used machine learning techniques to predict financial market trends or make trading decisions, primarily using numeric data extracted from candlesticks. However, these data often ignore the temporal and spatial information present in candlesticks, resulting in a poor understanding of their significance. In this study, we used multi-resolution candlestick images that contain temporal and spatial information on prices. The goal of this study was to compare the performance of raw numeric data and candlestick image data to optimize trading strategies and maximize returns. We used deep reinforcement learning algorithms (Deep Q-Networks (DQN), Dueling-DQN, and Proximal Policy Optimization (PPO)) to generate trading signals for opening a long or short position, closing a position, or staying idle. The trading signal was generated using a multiagent weighted voting ensemble approach. We tested the ensemble automated trading approach on two BTC/USDT datasets, a 30-day bullish market, and a 15-day bearish market. Our findings showed that models using candlestick image data outperformed those using numeric data and other baseline models. Additionally, we used a visual representation of candlestick images to depict the results from attention-based deep reinforcement learning algorithms, highlighting its advantages over other models