Model dossier
State of the art, and it shows its working.
Every grade this product prints comes from one small, inspectable model — and the numbers behind it are shown to you. Not a black box with a confident voice: a regression you can read, a calibration that is measured, and a scoreboard that is published whether or not it flatters us.
-
What it is
A logistic regression over eleven scale-free price and volume features, refit at request time on that instrument's own trailing history for the horizon you asked about. Because every feature is dimensionless, a ₹9 stock and a ₹3,000 stock are read on the same terms — and nothing is copied in from an instrument that is not yours.
-
How it is checked
Raw scores are mapped onto observed frequencies using a chronological hold-out the fit never saw. The whole pipeline is then walked forward through history to score Brier skill against a naive base-rate forecast — the baseline any model has to beat before it means anything.
-
Where it refuses
When there is too little independent evidence — few non-overlapping windows for the horizon asked for — the fit is refused and a transparent moving-average rule answers in its place, labelled as such. A model that always answers is a model that guesses.
What lands here
- The feature set, and the fitted coefficient on each, per horizon
- The reliability curve and the measured calibration error
- The walk-forward scoreboard: hit rate, Brier, and skill against the naive rate
- Effective sample size per horizon — rows are not observations
The full dossier ships with the server migration. Until then the model's working is shown inside every analysis: coefficients, factor waterfall and probability surface travel with the grade.
{disclaimers and disclosures}