This repository contains the source code, data, and manuscript for the Bachelor of Science in Computer Science thesis titled "Algorithmic Trading on Bitcoin with Market Microstructure Resilience Testing."
The research investigates the limitations of static mean-reversion strategies in non-stationary cryptocurrency markets and proposes an adaptive architecture using Hidden Markov Models (HMM) to mitigate catastrophic risk.
This study attempted to develop a profitable mean-reversion strategy for Bitcoin using Volume-Enhanced Bollinger Bands (VEBB). The research hypothesized that volume anomalies (Z-Score > 2.0) could filter false signals in the non-stationary 2025 market. While the volume signal showed significant promise during the 2024 calibration period (improving Profit Factor from 0.90 to 1.79), the strategy failed to maintain profitability during out-of-sample validation (-8.65% return).
An adaptive HMM architecture was introduced to mitigate these losses, successfully reducing maximum drawdown usage by 98.7% but failing to achieve positive returns (-1.30%). These results suggest that while volume anomalies capture meaningful microstructure events, these signals are not consistently tradeable using static execution parameters in high-volatility regimes. The primary contribution of this research is the empirical demonstration of the failure of static parameters to generalize out-of-sample, limiting the scope of the study to variance reduction rather than wealth generation.
The project evolved through two distinct phases:
- Logic: Traditional Mean-Reversion using Bollinger Bands + Volume Z-Score + On-Balance Volume (OBV).
- Implementation: Pine Script v6 (TradingView).
- Performance: Suffered -20.4% capital loss during the 2025 liquidity crisis due to chronic overtrading in trending markets.
- Logic: Unsupervised regime classification using Gaussian Hidden Markov Models.
- Features: Garman-Klass Volatility, Hurst Exponent, Volume Flow.
- Mechanism: A "Circuit Breaker" that actively disables trading during detected "Crisis Regimes" (State 0).
- Performance: Achieved -1.3% drawdown (Survival) vs -20.4% baseline loss. Validated that variance reduction is the primary benefit of regime-switching models in crypto.
Contains the full thesis manuscript and build files.
sp.pdf: The final compiled thesis document (73 pages).chapters/: Source LaTeX files for each chapter.sp.tex: Main document controller.
reactive_vebb.py: The core Python backtesting engine implementing the HMM logic and walk-forward validation.regime_features.py: Feature engineering module for Garman-Klass and Hurst calculations.
- Historical BTCUSDT 15-minute data (2023-2025) sourced from Binance via TradingView.
- Non-Stationarity: Static parameters optimized for 2024 failed in 2025, proving the "Stationarity Trap" hypothesis.
- Volume Anomalies: Volume Z-Scores > 2.0 served as reliable filters for 50th-percentile volatility regimes but failed during liquidity cascades.
- Survival Metric: The primary metric of success was shifted from "Profit Maximization" to "Drawdown Mitigation" after the HMM demonstrated a 92% reduction in trade frequency during crash periods.
- Statistical Integrity: The study acknowledges that while the regime-filtered configuration showed a Profit Factor of 1.50, the sample size (N=24) is insufficient for statistical significance, positioning the result as anecdotal evidence for future research rather than a production-ready strategy.
- [Author Names Redacted for Repository Privacy]
- Department of Computer Science
- Ateneo de Naga University