Data Science

Full recap of the Fortuna algorithmic trading project: HMM regime detection, ML signal generation, position sizing, and risk parameter analysis across Labs 1–9A.

Building a BTC spike predictor from scratch: z-score as the entry signal, EV tables for regime mapping, and volume as a leading indicator. Logistic regression baseline on 5,744 daily candles.

Fitting a 4-state Gaussian Hidden Markov Model on BTC candle shape features to produce soft regime posteriors γ_k(i). Bull / SoftBull / Choppy / Bear states derived from Baum's 1970 mathematics.

XGBoost classifier augmented with HMM regime posteriors γ_k(i). Bear regime avoidance alone produces +8.21 Sharpe on BTC. SOL simulation: $10k → $14k with regime gating.

The BTC-trained HMM labelled 78% of SOL history as Choppy. Fixed with SOL-specific rolling training window and a volatility ratio feature. Regime distribution now matches SOL's actual market character.

Cron-scheduled signal radar that fetches daily candles, runs the regime model, and emits typed signals to SQLite. Human-in-the-loop Textual dashboard for signal approval and regime monitoring.

Five sizing schemes tested on SOL: flat binary, linear confidence, Kelly, stepped, and regime-switch. Flat binary wins on trending markets; Kelly dominates on capital preservation.

Testing four risk parameters (ρ₁–ρ₄) for their correlation with drawdown. Key discovery: ρ₄ (jump outcome variance) is negatively correlated with drawdown — it is a conviction signal, not a risk measure.

Empirical exit timing study leading to a take-profit + z-score stop system. BTC: $10k → $215k (+2,058%). SOL TP=13%: $10k → $320k (+3,100%). Parameter sweep confirms window=20 optimal for both assets.

A cross-book analysis of two foundational options trading texts, extracting common patterns, the implied volatility framework, and four key trading ideas.