Cryptosentiment: A Dataset and Baseline for Sentiment-Aware Deep Reinforcement Learning for Financial Trading

Deep Learning (DL) models have been applied in several studies to solve financial trading problems. Most approaches handle these problems as classification or reinforcement learning problems with the objective of developing profitable strategies. Recent works have demonstrated that supplying financial trading agents with sentiment information can lead to improved performance. However, most of these works focus on collecting sentiment in a coarse-grain manner, which is not always appropriate for making fine-grained trading decisions, e.g., on a minute basis. In this paper, we introduce a fine-grained cryptocurrency sentiment dataset, called CryptoSentiment, which contains 235,907 sentiment scores for 14 cryptocurrency assets, gathered by various online sources. Moreover, we provide Deep Reinforcement Learning (DRL) baselines using the collected dataset, investigating the impact of multi-modal features on cryptocurrency trading.