Factor investing
Factor investing — the academic framework that powers most modern quant strategies. Posts here explain each factor in detail with real-world examples and link to the live ticker scores.
- DUNS is done: why the federal exit ended the gold-standard claimThe U.S. federal government quietly retired the DUNS number as its entity identifier in April 2022. The pay-to-play credit profiles, the stale trade tape, and the methodology nobody can audit had already done the damage. Here's the eulogy for the so-called gold standard.
- Do stock factors actually work? Testing momentum, low volatility, and reversal on 5 years of S&P 500 dataAfter three posts on credit-side modeling, here's the equity-side test. 619,040 daily price rows. 474 names. Five years. Three factors. Momentum +3.9% annualized, low vol -6.4% (wrong sign), reversal +2.8%. The honest version of factor investing.
- Do stock factors work in crypto? Testing momentum, low volatility, and reversal on 1,050 tokensAfter testing factors on 5 years of S&P 500 data and finding nothing significant, we pointed the same machinery at 1,050 cryptos over 6 years. Momentum IC went from +0.016 to +0.111 (t-stat +3.80). Low-vol flipped sign. Reversal flipped sign and lost 20%/year. The honest crypto version of factor investing.
- How credit scoring models actually work: a data-driven breakdownMost explanations of credit scoring stop at "lenders look at your income and credit history." That's not wrong — it's just not useful. So we pulled 32,437 real loan applications, trained a working model, and looked at what the numbers actually say.
- Predicting loan defaults: what the data tells us banks missThe previous credit-risk dataset was generous: AUC 0.87 with simple logistic regression. This one isn't. 67,463 loans, 35 features, and even random forest gets to AUC 0.527 — barely above random. Data quality beats model choice, every single time.
- Detecting credit card fraud: when 99.8% accuracy means your model caught nothingA model that predicts "not fraud" for every transaction in this dataset is right 99.83% of the time and catches zero fraud. We trained two real fraud models on 284,807 transactions and looked at which evaluation metric actually tells the truth.