Product / Predict

A model for one column, with a range you can trust

Choose what you want to predict. CLAIR trains a model, often before you even ask, and gives you a calibrated range next to the point estimate, plus a what-if calculator to test scenarios one input at a time.

What it does

Predict trains a model for one outcome column and lets you explore it: a what-if calculator to change one input at a time and watch the prediction move, and a batch-predict path to score every row in an uploaded file at once. Once a dataset's columns are classified, CLAIR quietly pre-trains a quick baseline for the first couple of outcome columns in the background, so Predict often opens on a model that is already there instead of a blank page.

How it works

Predict's own fast tier, Quick-look, trains on a genuine held-out 20% of your data, which is what lets it report an exact calibrated range around its prediction rather than an approximation. A three-step wizard walks you from picking the target column to reviewing the setup CLAIR chose; Automatic mode picks sensible defaults, Manual mode lets you choose the model families and features yourself.

A category column, like admission type or discharge disposition, or a true/false column becomes a usable Predict input, or a predictable outcome, automatically the moment CLAIR's Smart Encoding has processed it. A column still sitting as raw text is named outright, with a link to where to convert it, instead of just vanishing from the list. Predict can train several outcome columns independently in one pass; each is its own model, so a problem with one target's artifact never affects the others.

Honest limits

  • Calibrated ranges are regression only today. A multiclass target or a stacked champion has no result CLAIR can vouch for the same way, so Predict reports "no calibration" for those rather than a number built on a mapping it cannot confirm.
  • The automatic background baseline covers at most two outcome columns per dataset, trained on a 25-second budget with three-fold cross-validation. It is a first look, not a substitute for running Deep Analysis on the same target.
  • If the trained model's file fails to save, for example a full disk, Predict still shows the accuracy it measured but turns off What-if and file scoring for that run, and says why, rather than pointing you at a model nothing can find.