Forecasting Time-Varying Intermarket Dependencies Between Cryptocurrencies and Conventional Assets Using Machine Learning

Master's Thesis — Bogdan Babaev M.Sc. Artificial Intelligence · University of Kragujevac, Serbia (2023–2026)

---

Overview

This project investigates whether cryptocurrency market dynamics can be used to forecast time-varying dependency structures between Bitcoin and conventional financial assets — equity indices, precious-metal ETFs, and the U.S. dollar index.

A fully reproducible machine-learning pipeline is developed covering: data acquisition, return construction, rolling-correlation target generation, feature engineering, walk-forward model evaluation, and benchmarking against a leakage-safe DCC-GARCH(1,1) specification.

Key finding: Intermarket dependency is forecastable out of sample. Ridge, AR(1), and HAR form a near-indistinguishable top cluster (avg RMSE 0.0656–0.0659, R² ≈ 0.942–0.943), confirming that serial persistence is the dominant driver. All ML models substantially outperform the DCC-GARCH econometric benchmark (avg RMSE 0.2136).

---

Asset Universe

Symbol	Asset	Type
BTC-USD	Bitcoin	Crypto (base)
ETH-USD	Ethereum	Crypto (reference)
^GSPC	S&P 500	Equity index
^IXIC	NASDAQ Composite	Equity index
GLD	SPDR Gold Shares ETF	Precious metal
SLV	iShares Silver Trust ETF	Precious metal
UUP	Invesco US Dollar Index	Currency

Daily prices 2017–2026 · source: Yahoo Finance

---

Methodology

• Target: rolling Pearson correlation (BTC vs each asset) — windows: 14 / 30 / 60 / 90 days
• Transform: Fisher-z (arctanh) for variance stabilization
• Features: momentum, volatility, return-based predictors derived from the dependency series
• Models: Naive_Last, AR(1), HAR, ElasticNet, Ridge, Adaptive Ensemble, Random Forest, GBM, XGBoost vs DCC-GARCH(1,1) benchmark (10 specifications total)
• Evaluation: walk-forward expanding window (no data leakage)
• Statistical tests: Diebold-Mariano with Newey-West correction
• Signal layer: logistic classifier for investor stress-day detection on traditional assets

---

Project Structure

├── main.py                  # Entry point — runs full pipeline
├── run_all.py               # Full reproducibility runner (pipeline + notebooks)
├── config.yaml              # All settings
├── requirements.txt
│
├── thesis_app/
│   ├── pipeline.py          # Core ML pipeline
│   ├── dcc.py               # DCC-GARCH(1,1) benchmark
│   ├── dcc_walk.py          # Walk-forward DCC wrapper
│   ├── signal_layer.py      # Investor signal layer
│   ├── regime_analysis.py   # Regime detection utilities
│   ├── data_quality.py      # Data validation
│   └── notebook_helpers.py  # Shared notebook utilities
│
├── notebooks/
│   ├── 01_EDA_Dataset.ipynb        # Exploratory data analysis
│   ├── 02_GridSearch.ipynb         # Hyperparameter search (TimeSeriesSplit)
│   ├── 03_Model_Comparison.ipynb   # RMSE/R² across all pairs & windows
│   ├── 04_DM_Tests_Visuals.ipynb   # Diebold-Mariano tests & thesis figures
│   ├── 05_XGB_vs_DCC.ipynb         # XGBoost vs DCC-GARCH deep dive
│   ├── 06_Regime_Analysis.ipynb    # Dependency regime analysis
│   ├── 07_Robustness_Checks.ipynb  # Sensitivity & robustness
│   └── 08_Market_Events_Showcase.ipynb  # Landmark market-event case studies
│
├── data/                    # Auto-created on first run
├── outputs/                 # Auto-created: figures, results, predictions
└── models/                  # Auto-created: saved model artefacts

---

Quickstart

# 1. Clone
git clone https://github.com/b0gdaan/master-thesis.git
cd master-thesis

# 2. Create virtual environment (Python 3.14 recommended)
python -m venv .venv
.venv\Scripts\activate       # Windows
# source .venv/bin/activate   # Linux / macOS

# 3. Install dependencies
pip install -r requirements.txt

# 4. Run full pipeline
python main.py

First run downloads ~10 years of price data from Yahoo Finance (~30 sec). Subsequent runs use the cached CSV in data/raw/.

No GPU? Set xgb_device: "cpu" in config.yaml. No DCC-GARCH? Set use_dcc_garch: false in config.yaml.

---

Notebooks

Run after main.py (outputs must exist):

jupyter notebook

Notebook	Purpose
01_EDA_Dataset.ipynb	Price/return EDA, ADF stationarity tests, correlation overview
02_GridSearch.ipynb	Hyperparameter tuning with TimeSeriesSplit CV
03_Model_Comparison.ipynb	RMSE / R² heatmaps, model ranking, LaTeX tables
04_DM_Tests_Visuals.ipynb	DM significance tests, publication-quality forecast plots
05_XGB_vs_DCC.ipynb	Error analysis, rolling RMSE, scatter diagnostics
06_Regime_Analysis.ipynb	Dependency regime detection and characterization
07_Robustness_Checks.ipynb	Sensitivity to rolling window length and refit frequency
08_Market_Events_Showcase.ipynb	Landmark market-event case studies (COVID, FTX, SVB, tariff shock, …)

---

Configuration

All parameters are in config.yaml:

Key	Default	Description
base_asset	BTC-USD	Base cryptocurrency
rolling_windows	[14,30,60,90]	Correlation window sizes (days)
use_fisher_transform	true	Fisher-z transform on correlation target
use_dcc_garch	true	Include DCC-GARCH(1,1) benchmark
use_xgboost	true	Include XGBoost model
xgb_device	cuda	cuda for GPU, cpu for CPU
min_train_size	800	Minimum training observations (walk-forward)
refit_every	20	Model refit frequency (trading days)
n_parallel_workers	8	Parallel experiments (ThreadPoolExecutor)
enable_signal_layer	true	Run investor stress-warning signal layer

---

Tech Stack

Python 3.14 · pandas 3.0 · numpy 2.4 · scikit-learn 1.8 · XGBoost 3.2 · arch 8.0 · scipy 1.17 · yfinance 1.2 · matplotlib · seaborn · Jupyter

---

Author

Bogdan Babaev M.Sc. student · Artificial Intelligence — University of Kragujevac, Serbia (2023–2026) github.com/b0gdaan · linkedin.com/in/b0gdaan