Analysis
Feature importance is not causation: the mixer problem
Importance scores measure a model's reliance on inputs under correlations — a variable your model ranks first may only be standing in for the real driver.
Home / Analysis
Analysis
Analysis coverage from UZU NEWS.
Analysis
What does it mean for a probability forecast to be calibrated?
Calibration is the match between stated probabilities and observed frequencies — measurable with Brier scores and reliability diagrams, and distinct from…
Analysis
What transaction costs do to a backtest's promised returns
Spread, impact and timing drag compound with turnover — and the strategies they kill are usually the high-Sharape, high-frequency ones.
Analysis
Survivorship bias in fund data: how big, and how to check
Databases of existing funds answer a question nobody asked — how did the survivors do — and the gap is measured in percentage points.
Analysis
How do you detect lookahead bias in someone else's backtest?
A checklist of audit questions — each targeting a specific leak channel — that turns a suspicious track record into a diagnosed one.