Questions before models
Six questions investors should ask quantitative models
Each answer begins in plain language, then opens into evidence, application, limitations, mathematics, and reproducible code. Read only as far as the decision requires.
- Decision briefs
- 06
- Executed studies
- 06
- Automated tests
- 204
- Reproduction
- Frozen snapshots
Choose a question
Begin with the assumption closest to the decision
01 · Portfolio construction
Can an “optimal” portfolio stay optimal?
Short answer
Not reliably here. The fitted winner lost much of its apparent advantage in later data.
1.12 → 0.44 Sharpe
Follow the answer →02 · Model risk
Does a 95% risk estimate fail only 5% of the time?
Short answer
No. Breaches were too frequent and clustered, so the forecast failed calibration tests.
135 vs 107 breaches
Follow the answer →03 · Regime risk
Will bonds diversify stocks in the next stress?
Short answer
Not necessarily. The stock–bond relationship changed sign across the studied regimes.
−0.35 → +0.37
Follow the answer →04 · Factor investing
Does a factor fund deliver the exposure on its label?
Short answer
The advertised exposures appeared, but no tested alpha survived multiple-testing correction.
0 alphas at 5% FDR
Follow the answer →05 · Derivatives
Can one volatility number describe every option?
Short answer
No. Volatility, tail behavior, and option prices contradicted the tested constant-volatility model.
≈3% → 93% realized vol
Follow the answer →06 · Numerical methods
How many simulations are enough?
Short answer
Enough is defined by the decision’s error tolerance, not by a conventional path count.
Error ∝ N⁻¹ᐟ²
Follow the answer →Consistent reading path
Simple first; technical only when useful
- 01AnswerPlain-language conclusion.
- 02EvidenceData, estimates, uncertainty.
- 03ApplicationDecision implications.
- 04LimitsBoundaries of the claim.
- 05Method and codeTests, mathematics, source.
Evidence standard
Every result includes its uncertainty and documented limitations.
A refreshed dataset is treated as a new experiment. Earlier failures, revisions, source commits, and executed artifacts remain available in the research record.