What is being presented is purely a research project and not a recommendation for trading:
So far, I have experimented with various models to predict Bitcoin prices using on-chain data. I've utilized 373 features from the CryptoQuant platform spanning from 2012 to the present day. Since I employ a sliding window technique, classical machine learning models, which typically work with 2D data, are not suitable for my data. Instead, I use deep learning techniques based on tensors, which enable processing of 3D data.
Among the different models I've tried in recent months, the best results have been obtained with N-Beats and WaveNet models. The N-Beats model is developed in TensorFlow, and the model accuracy is MAPE: 31.9849. The performance of this model on train, validation, and test data is visualized in the image A. Based on this, the forecast of the N-Beats model for the next 30 days is shown in the chart B.
The second model that has provided acceptable results so far is the WaveNet model. The loss values for this model have been measured by Negative Log-Likelihood, with a loss value of 2.88. This model also used the same data as the previous model. The image C shows its performance in predicting prices for the past month. And the image D shows the Bitcoin price prediction for the next month based on the WaveNet model.
Based on the WaveNet model, with a confidence interval of 50%, the Bitcoin price is likely to fluctuate within the same range it has experienced over the past few months in the upcoming month.
Written by CryptoOnchain
