A selection of research projects, publications, and software systems. Focused on AI, quantitative finance, and the intersection of machine learning with real-world business problems.
Hybrid Random Forest classifier predicting financial distress ("quasi-zombie" status) in 30 SDAX industrial firms. Combining financial ratios, macroeconomic variables, and FinBERT sentiment features: ROC-AUC 0.973, F1 0.711 — outperforming logistic regression baseline by +43.5pp F1.
Full-stack portfolio management (Python/Flask) covering 133+ EUR/US tickers. Black-Litterman optimization, Monte Carlo simulation, regime detection, LSTM alpha signals, and Ledoit-Wolf covariance shrinkage. 3.65-year backtest: 26.0% CAGR, 1.46 Sharpe, –18.5% max drawdown, +9.7% annualized alpha.
Modular multi-agent ecosystem (V8.6) with a central Agent OS enforcing governance across 14 capability scopes and 30+ tools. DCF Agent automates multi-period equity analysis; Personal AI Assistant handles research and task automation. Validated by 1,127+ unit tests across 82 test files.
Estimates and compares three VaR methods (parametric, historical simulation, Monte Carlo) and Expected Shortfall on 1,036 daily log-returns. 1-day 99% Historical VaR: 7.11%, Expected Shortfall: 12.36%. CAPM regression and Sharpe/Treynor/Roy ratios with bootstrap confidence intervals.
Designed a machine-learning scoring workflow using SHAP-based feature attribution to formalize credit officer risk intuition. Analyzed 1,000+ prior fraud incidents; embedded ML insights into the credit decision process for consumer and commercial lending.
SSRN publication. Isolates the predictive contribution of purely geo-spatial features for real estate valuation in the German market (2024–2025). Random Forest for price prediction, XGBoost for valuation intervals; automated geo-score computation pipeline.
Compared Logistic Regression and XGBoost on TF-IDF, VADER, and BERT representations for sentiment classification (bearish/neutral/bullish) of 414 financial news headlines. Best result: TF-IDF + Logistic Regression with 69% accuracy, macro-F1 0.64.
Internal platform built from scratch as Product Owner and backend/database lead. Tracks 10+ property and construction details per asset (contractors, materials, heating systems). Led requirements-gathering meetings and authored specification documentation.