In a recent research project, various models were tested to predict Bitcoin prices using on-chain data from the CryptoQuant platform. The study utilized 373 features spanning from 2012 to the present day. Classical machine learning models were unsuitable due to the use of a sliding window technique, leading to the application of deep learning techniques based on tensors for 3D data processing.

The most promising results were obtained with the N-Beats and WaveNet models. The N-Beats model, developed in TensorFlow, achieved a Mean Absolute Percentage Error (MAPE) of 31.9849. The model's performance on train, validation, and test data is demonstrated in image A, with the 30-day forecast shown in chart B.

The WaveNet model also showed acceptable results, with a Negative Log-Likelihood loss value of 2.88. Image C displays its performance in predicting past month prices, while image D presents the Bitcoin price prediction for the upcoming month. According to the WaveNet model, the Bitcoin price is likely to fluctuate within the same range it has experienced over the past few months, with a 50% confidence interval.