Applied experimentation

Prediction Sandbox

Train a model, enter a new case, inspect its probability and local feature effects, and compare it with nearby training observations.

Prediction sandbox Code ↔ visual Challenge mode

Lab concept guide

What to observe while you experiment

A prediction is produced by applying the complete fitted pipeline to one new case. Its probability/value should be interpreted with the model’s validation, threshold/calibration and local feature context—not as certainty.

MechanismEnter a new raw case, pass it through the same preprocessing/model, inspect probability/prediction, local effects and nearby training examples.

Failure modeFeeding manually preprocessed values that do not match the training pipeline or treating one probability as a causal/guaranteed outcome.

VerificationRe-enter a known training/validation-like case, compare its transformed values/prediction, and perturb one feature to check direction and local sensitivity.

Experiment deliberately

Create two nearby cases differing in one feature. Predict which probability/local contribution should change and verify the model response.

The trained model stays fixed while you change the new case. That makes movement across the decision separator visible without silently retraining the model.

ModelThreshold0.50Seed

Ready.

New case

Move one feature at a time and watch the predicted probability and star position change.

Open full feature-importance analysis →

Signal A0Signal B0Noise feature0Redundant A0

Model separator and new case

Local feature effects

Probability change when each feature is replaced by its training median.

Nearest training observations

Similarity is calculated in the standardised feature space. Nearby examples provide context, not causal explanations.