Analysis Lab

Model Finder

Pick a target; the finder searches the panel's variables for the best predictive model with deterministic statistics — screened candidates, entity-grouped cross-validation, forward selection, and a cluster-robust final fit. Every exclusion and decision is disclosed.

DomainWhich panel to model — ….

Domain

Target (predicted variable)The variable the model should predict.Effort5 folds, 20 candidates, up to 8 terms — the honest default; minutes on large panels.Family override (optional)Leave on automatic: binary targets go binomial, skewed positive targets run a gaussian-vs-gamma probe.

Universe

Restrict which rows the search may learn from — the same slicers, saved cohorts and period windows the other models use. The chosen universe is recorded on the run.

Geography (tier 1)Asset class (tier 2)Sub-category (tier 3)Reference groupRestrict the sample to a saved cohort (built on the Groups page).Fee methodHow a fund's representative TER is chosen across its fee classes.Quarters (last N)Leave blank for all history.

Exclude factors

Tick anything the search must not consider — a variable you know is contaminated, unavailable in advance, or simply not the question. Every exclusion is listed in the report.