Rosetta framework

Context

Quantitative research generates candidate features faster than it can judge them. Any sufficiently long search over a price series will produce something that looks predictive.

Constraint

The failure modes are systematic and well known: lookahead bias, sensitivity to execution delay, instability across time, and the multiple-comparison problem that makes an impressive in-sample result the expected outcome of a large enough search. A feature that survives none of these still backtests beautifully.

Contribution

I built Rosetta for MenthorQ, a platform that engineers candidate features, scores them for alpha, collinearity, and signal-to-noise, and then puts them through validation designed to catch exactly those failure modes.

Approach

Validation is the centre of the tool rather than an afterthought: lookahead detection, delay sensitivity, temporal stability, and correction for multiple testing all run before a feature is trusted. Modelling starts linear, and complexity is added only where it earns its place.

Outcome

A feature is accepted or rejected on evidence that anticipates how it will fail, rather than on the strength of its backtest.

Evidence

The platform belongs to MenthorQ and is held in a private repository containing the feature engineering, the quality audit, the validation suite, and the modelling layer. What is described here is method, not the signals it evaluates.