Researchers Develop DL-GuesS: A Deep Learning and Sentiment Analysis-Based Framework For Cryptocurrency Price Prediction

A cryptocurrency is a string of encrypted data representing a currency unit. It has been hugely successful since money transfers are cheaper and faster, and decentralized systems don’t collapse at a single point of failure. Because of this, academics have become interested in the topic and have attempted to forecast price fluctuations for various kinds of cryptocurrencies. However, this task is challenging given its extreme volatility and reliance on other cryptocurrencies.

The price forecasting of cryptocurrencies has drawn the attention of numerous researchers. Several works proposed to use the price history and algorithms such as multi-layer perceptron, support vector machine, random forest, and long short-term memory (LSTM) to ensure the prediction. In addition, the technique of sentiment analysis, based on natural language processing, was also exploited by combining it with the algorithm cited above. Those research proved that choosing more variables is not a concern; the main challenge is picking the appropriate features to forecast prices and creating a reliable model. In this context, a research team composed of Indian and South African scientists proposed DL-GuesS, a deep learning network based on LSTM and gated recurrent unit (GRU), and a Twitter sentiments-based hybrid model, which targets to predict the price of the cryptocurrency.

DL-GuesS targets to predict the price of a specific currency regarding their price history and tweet sentiments of the other dependent or alternate coins. It specifically considers the window sizes, i.e., 1, 3, and 7 days. The authors also took into account the inter-cryptocurrency dependencies to enhance the efficiency of the suggested model. A correlation study between several currencies has shown that Bitcoin, Litecoin, and Dash are very dependent and that it is wise to use all three in the training phase to be able to predict the price of one of them each time. 

Two types of inputs are used to ensure the training stage: past days’ prices and present-day tweets for each cryptocurrency. Each type of data is first processed by a specific branch. One branch based on the VADER algorithm is made to get the polarity of tweets. The other branch is built by 100 neurons of LSTM, 100 neurons of GRU, and 100 neurons of Dense. It takes the cryptocurrency price data. Then, the outputs of the two streams are merged. This operation is carried out simultaneously through three subunits for the three types of cryptocurrency. The output layer receives the concatenated outputs from the three subunits. Following this strategy, the proposed network is considered a multi-level hierarchical model since the past prices of Dash, Litecoin, and Bitcoin are passed as input features.

The authors perform a comparison study with the traditional prediction model, which takes only one type of currency as input to check the efficiency of DL-GuesS. Three metrics (MSE MAE and MAPE) are utilized to evaluate the models. Two scenarios were done in the experimental study. In the first scenario, the price DASH prediction is performed using traditional and multi-level hierarchical strategies. In the second scenario, the same process is made for BITCOIN-CASH prediction. Results obtained in the two scenarios reveal that the proposed multi-level hierarchical approach performs better than the conventional systems.

In this paper, we have seen an overview of a new hybrid model, DL-GuesS, proposed to forecast cryptocurrency prices regarding both price history and sentiments analysis of recent Twitter. An experiment study demonstrates that the new approach outperforms conventional models. 

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'DL-GuesS: Deep Learning and Sentiment Analysis-Based Cryptocurrency Price Prediction'. All Credit For This Research Goes To Researchers on This Project. Check out the paper.

Please Don't Forget To Join Our ML Subreddit