Exploring the Predictive Power of On-chain Metrics in Bitcoin Price Forecasting

This repository contains a collection of Jupyter notebooks detailing a machine learning project for predicting Bitcoin prices. The project utilizes two datasets, ta_df and oc_ta_df, which include Bitcoin prices, technical indicators, and on-chain metrics.

The main notebooks can be found in: src/python/main-experiment-notebooks.

Content

Here's a brief description of what each notebook does:

1. 1-exploratory.ipynb: This notebook is the initial exploration of our datasets. It contains visualization of data, outlier detection, correlation analysis.

2. 2-experiment_1.ipynb: This notebook trains BiLSTM and BiRNN models on the ta_df and oc_ta_df datasets. The objective is to quantify the impact of including on-chain metrics in the model. It provides initial insights into how on-chain metrics influence the model's predictive performance.

3. 3-hypertuning_tadf_birnn.ipynb and 4-hypertuning_tadf_bilstm.ipynb: These notebooks perform hyperparameter tuning for the BiRNN and BiLSTM models, respectively, on the ta_df dataset without any human intervention. The goal here is to ensure that the results from the first experiment are not merely due to the selection of hyperparameters.

4. 5-experiment_2.ipynb: This notebook repeats the process of 2-experiment_1.ipynb, but now using the hyperparameters obtained from the tuning notebooks (3-hypertuning_tadf_birnn.ipynb and 4-hypertuning_tadf_bilstm.ipynb). The main aim is to validate the robustness of the findings - if the models still perform better on the oc_ta_df dataset even when tuned on the ta_df dataset, it confirms the robustness of our findings.

Usage

To reproduce the findings, run the notebooks in the order they are numbered. You'll need to have Jupyter installed, along with libraries such as pandas, numpy, matplotlib, seaborn, and scikit-learn.

Conclusion

The goal of this project is to evaluate the impact of on-chain metrics for Bitcoin price prediction. On-chain metrics are found to improve model performance in both experiments.