The success of deep learning (DL) in various areas, such as computer vision, fueled the interest in several novel DLenabled applications, such as financial trading, which could potentially surpass the previously used approaches. Indeed, there has been a plethora of DL-based trading methods proposed in recent years. Despite the success of these methods, they typically rely on a very restricted set of information, usually employing only price-related information. As a result, they ignore sentiment-related information, which can have a profound impact and be a strong predictor of various assets, such as cryptocurencies. The contribution of this paper is multifold. First, we examine whether the use of sentiment information, as extracted by various online sources, including news articles, is beneficial when training DL agents for trading. Then, given the difficulty of training reliable sentiment extractors for financial applications, we evaluate the impact of using different DL models as sentiment extractors, as well as employ an unsupervised training pipeline for further improving their performance. Finally, we propose an effective multisource sentiment fusion approach that can improve the performance over the rest of the evaluated approaches. The conducted experiments have been performed using several different configurations and models, ranging from multilayer perceptrons (MLPs) to convolutional neural networks (CNNs) and recurrent neural networks (RNNs), to provide a reliable evaluation of sentiment-aware DL-based trading strategies providing evidence that sentiment information might be a stronger predictor compared to the information provided by the actual price time series for Bitcoin.

No abstract