Interactive model explorer

What’s behind a prediction?

Explore a customer’s predicted churn in the next 30 days.

SYNTHETIC DATA · DEMONSTRATION

Choose a customer

Synthetic customer

Explore a what-if

Adjust the profile to recalculate the prediction.

Inspect the prediction

Predicted churn probability

—

Original customer—No changes

Why this prediction? SHAP contributions

Blue lowers the model’s score; orange raises it. Larger bars mean a larger contribution for this profile.

Contributions add up in log-odds, not percentage points. They explain the model relative to its background data; they do not establish causes.

Choose the tradeoff

600 held-out synthetic customers

Lower thresholds flag more customers, including more false alarms. Higher thresholds miss more customers who churn.

Flag churn risk at or above 50%

Flag everyoneFlag no one

Changing one customer’s what-if profile leaves this test cohort unchanged. The threshold affects both its classifications and the selected profile’s flag.

Confusion matrix · positive = churn
	Predicted stay	Predicted churn
Actually stayed	—True negatives	—False alarms
Actually churned	—Missed churn	—True positives

How to read this demo

Which features matter overall?

Mean absolute SHAP contribution across the 600 test customers, in log-odds. This measures model reliance, not causal importance.

A transparent demonstration

A logistic regression is fitted to 2,400 synthetic customers and evaluated on 600 separate synthetic customers. Outcomes and relationships are simulated; these scores are not MLAIA client results.

Model and explanation methodology

Five independently generated inputs, a fixed random seed and a Bernoulli churn outcome. Standardization is fitted on the training set only. The browser uses exported fitted coefficients for inference.

For this additive linear log-odds model, exact interventional SHAP is coefficient × (input − training-background mean). The baseline is the mean training log-odds, not the mean churn probability. All five contributions plus the baseline equal the model log-odds, which is converted to probability with the logistic function.

The observed outcome belongs only to the original test customer. Modified scenarios have no observed outcome. Real deployment needs representative data, temporal validation and calibration assessment.

SHAP methodology · Explore ExplainerDashboard

This is a custom browser implementation; it does not run the ExplainerDashboard package.

Explore this with your business data →Back to the practice