The research question
Does the model add information, or reproduce what the rest of the modelling process already knows?
A machine-learning project for the Numerai setting, where raw predictive correlation and marginal contribution are distinct evaluation objectives.
The technical work
Model development, feature handling, neutralisation and objective-aware evaluation.
The challenge is to connect the economic idea to a precise, reviewable experiment. Data transformations, the moment of a decision and the interpretation of the resulting return series all belong to that specification.
What to challenge
These are evaluation considerations for this domain, not a checklist presented as already passed.
- Time-aware splits
- Marginal contribution versus raw correlation
- Model and dataset version tracking
- Adaptive search counted in the evaluation
The data boundary
Versioned modelling datasets with documented targets, eras and platform evaluation rules.
Descriptions are intentionally high-level. Parameters, execution rules and non-public source data are not published here. Withholding an implementation is different from overstating its evidence.
The interesting question is not whether a backtest looks good. It is what would have to be true for the result to mean something.
For the broader framework, see validation. For the engineering context, see the research stack.