The nature of Cryptocurrency markets presents a challenge for Financial Time series forecasting, the regular use of time bars as a source of data to forecast can prove insufficient to predict the movements of the crypto token value. The use of additional data from DeFi sources can be used to create a more robust base in which to use different methods to perform better feature generation and feature selection to use for the prediction models. The use of the Three Barrier Method for labeling the movements of the data is suggested as a way to generate multiclass labeling in which both directions of the prices and magnitude are represented. The proposal of this work is that the use of DeFi data, the adapted use of the three-barrier method, and the use of Genetic Programming could create a dataset that has good predictor capabilities for the multiclass classification prediction of the movement and magnitude of the value of Bitcoin. In this work, a comparison between prediction models is performed using a combination of benchmark models, and the implementation of Random Forest and Multi-Layer Perceptron to construct a multiclass classifier for the price movement of the cross of Bitcoin and USDT from the Binance Exchange using historical data from Binance, Ethereum Blockchain, and symbolic data.
