The thesis introduces an innovative way of predicting the value of the Etherum cryptocurrency using deep learning. Good forecasting is a key to building a successful trading strategy, which is also our long-term goal. We focus on short-term forecasting, in particular, we forecast the time series of six 5-minute increments of the Etherum price.
To build predictive models, we use deep learning which is known for its capacity to capture complex relations in the data. Lately, the vast majority of the state-of-the-art models in sequence-to-sequence predictions have been using deep learning. We employ methods for supervised and unsupervised deep learning. Supervised methods are employed for learning the final predictive model, where we will take advantage of attention mechanism, while unsupervised methods are applied to the task of dimensionality reduction of the input data. For both supervised and unsupervised methods, we add noise to training data to reduce overfitting. The training data is collected from the cryptocurrency broker Kraken. To the transactional data obtained there, we will add different price-smoothing and other trading indicators to increase the predictive performance of the learned models.
The achieved results exceeded our initial expectations, especially since econometrics implies that short-term forecasting is more difficult than long-term forecasting. We have managed to contain the big problems with overfitting. The final, best-performing prediction model is obtained using modified attention mechanism.
|
