Transformer fonksiyon yaklaşımcısını kullanarak derin Q-Ağı tabanlı kripto para yatırım stratejileri

Cryptocurrencies have recently started to become more apparent as an alternative investment class, as they are mainly decentralized and transparent. Since cryptocurrencies have unique features and are more volatile, we need to develop novel approaches for more accurate price prediction and more profitable investment strategies. Deep learning methods including deep reinforcement learning and transformers have been attracting remarkably more attention. Here, we come up with Reincrypt which applies Deep Q-Network (DQN) with a vision transformer (ViT) neural network function approximation, where we transform one-dimensional time-series data to 2D grayscale image-like data by using a set of important technical analysis indicators. Such transformed data will be taken as input in our framework to predict various cryptocurrency prices. To stabilize the learning process of the model and enhance its performance, we employ techniques like parameter freezing and experience replay. Our detailed findings show that the proposed method is successful in forecasting changes in cryptocurrency prices. Overall, our method yields profit in different cryptocurrencies, not just Bitcoin. When we analyze the portfolios formed as a result of our method's output, we obtain approximately 0.5% return per transaction before considering transaction costs for the cryptocurrencies. Additionally, we can predict future cryptocurrency prices even when we train our method on a different cryptocurrency. In this case, we can train our method on highly-traded and liquid cryptocurrencies such as Bitcoin, and test its performance on relatively less liquid alternative cryptocurrencies. According to these results, deep learning-based cryptocurrency price prediction methods could be utilized for alternative cryptocurrencies, even though we have fewer data for these assets.