Forecasts Bitcoin prices over a 6-month horizon using ARIMA(2,0,3), optimized via grid search, achieving an RMSE of 4,586 USD. Combines historical price data, Google Trends, and on-chain metrics (0.28 correlation with active addresses) to predict BTC price trends using R.

Developed as part of a study at Università di Padova, this project predicts Bitcoin (BTC) prices using time-series and machine learning models: Multiple Linear Regression (MLR), ARIMA, SARMAX, XGBoost, and Prophet. ARIMA(2,0,3) emerged as the best performer, leveraging monthly aggregated data from 08/2017 to 12/2021. It explores Bitcoin’s volatile, nonlinear behavior with preprocessing techniques like log transformations and stationarity adjustments.

• Top Model: ARIMA(2,0,3) with RMSE 4,586 USD and MAPE 8.43%.
• Data Integration: Historical BTC prices, Google Trends, and on-chain metrics (e.g., Active Addresses, Transaction Count).
• Model Comparison: Evaluates MLR, ARIMA, SARMAX, XGBoost, and Prophet with grid search optimization.

• R: Core language (tested with R 4.3.2).
• Packages:
  • tidyverse: Data manipulation and visualization (dplyr, ggplot2, etc.).
  • forecast: ARIMA and SARMAX modeling.
  • xgboost: Gradient boosting.
  • prophet: Time-series forecasting.
  • tseries: Stationarity tests (ADF).
  • car: VIF for multicollinearity.
  • Metrics, MLmetrics: Performance evaluation.

• R 4.0+ (RStudio recommended).
• Install dependencies:

  install.packages(c("tidyverse", "forecast", "xgboost", "prophet", "tseries", "car", "Metrics", "MLmetrics"))
  

• isoformat-2017.xlsx: Daily BTC prices from OpenDataSoft.
• btc_search.xlsx: Monthly Google Trends for "Bitcoin" from Google Trends.
• btc_onchain.xlsx: On-chain metrics from blockchain APIs.

!! Place datasets in data/ or update paths in .R file

git clone 'https://github.com/oonuroo/BitcoinARIMA.git'
cd BitcoinARIMA

source("install_packages.R")

Rscript main.R

• data/: Store input datasets (manually added).
• install_packages.R: Installs required R packages.
• main.R: Full analysis script (preprocessing, modeling, visualization).
• results/: Output plots (e.g., arima_plot.png, xgboost_importance.png).

• Aggregated daily prices and on-chain data to monthly intervals.
• Applied log transformations (e.g., log_Volume, DiffMean) to reduce skewness.
• Dropped 4.1% missing on-chain data; retained outliers for market insights.
• Decomposed prices into trend, seasonal, and residual components; differenced for stationarity (ADF p-value 0.03).

• Correlation analysis: Active Addresses (0.28), Transaction Count (0.21), Google Trends (0.12).
• VIF to remove multicollinearity, retaining TxCnt, AdrActCnt, TrendScore, etc.

• MLR: Baseline, RMSE 15,758 USD (poor generalization).
• ARIMA(2,0,3): Grid search optimized, RMSE 4,586 USD on test data.
• SARMAX: Grid search applied, but often reverted to ARIMA(0,0,0), failing to converge.
• XGBoost: Grid search over features and parameters (e.g., max_depth=3), RMSE ~17,418 USD.
• Prophet: Grid search with regressors, RMSE 9,197 USD, excels at long-term trends.

• ARIMA(2,0,3): Best performer (RMSE 4,586 USD, MAPE 8.43%), though it overestimates price declines.
• Key Predictors: Active Addresses and Transaction Count moderately enhance accuracy.
• Visualizations: See results/ for ARIMA forecasts, XGBoost feature importance, and more.

• Short-term volatility remains challenging.
• SARMAX struggles with weakly correlated external predictors.
• XGBoost underperforms without extensive tuning.

License MIT License - free to use, modify, and share. Contact Onur Bacaksız | Email (mailto:onurobacaksiz@gmail.com)