About DistressSignal
Built by a finance researcher who got tired of finding out about corporate collapses too late.
Founder & Lead Researcher
I built DistressSignal because I needed it. During my MFin at IE Business School, I spent a semester studying corporate restructurings — Hertz, Revlon, Bed Bath & Beyond. Every time I dug into the 10-K filings from 18 months before the Chapter 11, the signals were right there. Leverage ratios in the 90s. Interest coverage below 1. Equity short interest at 30%+. None of it was a surprise in retrospect — but it was invisible in real time.
The Origin Story
My MFin thesis at IE Business School focused on early-warning systems for corporate credit deterioration. The question was simple: can a machine learning model outperform traditional Altman Z-Score models in predicting Chapter 11 filings with a 12-month horizon?
The short answer, after training on over 600 historical defaults from 1998 to 2025: yes, substantially. The Altman Z-Score — a linear five-factor model designed in 1968 — correctly identifies roughly 72% of defaults in out-of-sample testing. Our LightGBM classifier, trained on 24 features including short interest, equity volatility, and EBITDA margin, achieves 91.4% ROC-AUC on the same validation set (2018–2025).
The thesis became the model. The model became DistressSignal.
How This Was Built
Studying Corporate Defaults, Case by Case
Deep-dived into restructuring filings of Hertz, Carvana, Revlon, and Chesapeake Energy. Identified repeating structural patterns: excessive leverage, negative interest coverage, accelerating equity short interest. This became the feature engineering foundation for the LightGBM model.
Building the Inference Pipeline
Wrote Python scripts to programmatically extract balance sheet figures directly from SEC EDGAR 10-K and 10-Q filings. Fed into a LightGBM classifier trained on Compustat/CRSP historical data with strict walk-forward validation to eliminate look-ahead bias.
Backtesting on Real Bankruptcies
Ran retrospective predictions on Hertz (flagged 12 months before filing), Carvana (flagged 9 months before restructuring), and Tupperware (flagged 8 months before Chapter 11). The model's 91.4% ROC-AUC and 3.8% false positive rate validated the approach as institutionally credible.
Opening the Model to Researchers & Analysts
DistressSignal is the public-facing layer on top of that research — a weekly model run, published reports, and real-time alerts for analysts who want institutional-grade default signals without institutional-grade subscription fees.
Education
Master in Finance (MFin)
IE Business School, Madrid
Specialization: Corporate Finance & Capital Markets. Thesis: ML-Based Early-Warning Systems for Corporate Credit Default.
Research Focus
Special situations investing, distressed credit analysis, and gradient-boosted classifiers applied to financial statement analysis.
600+ corporate defaults studied · 1998–2025 training set.
Technical Stack
Python · LightGBM · SEC EDGAR API · Supabase · Paddle · GitHub Actions · Hugo.
Solo-built and maintained. All inference code runs on open-source tooling.
Why Not Just Use Bloomberg?
Bloomberg Terminal subscriptions run $24,000/year. Credit research from major banks is gated behind institutional relationships. The signals I'm publishing — leverage ratios, interest coverage, short interest — come from publicly available SEC filings. The only thing missing for most retail analysts and small fund managers is the time and tooling to process them systematically.
DistressSignal exists to close that gap. A $9.99/month subscription gives you the same early-warning signals that took months of thesis research to build — delivered weekly, directly to your inbox.
Get in Touch
Research inquiries, methodology questions, API access, or press:
sunmoon4cap@gmail.com
Data privacy or account issues: sunmoon4cap@gmail.com