Individual Conditional Expectation Plots

A partial dependence plot shows the average effect of one feature on a model's prediction, holding everything else fixed and averaging across all observations. That average can be misleading if the feature actually affects different observations in opposite ways — those effects can cancel out and look like nothing is happening. An Individual Conditional Expectation (ICE) plot fixes this by drawing one line per observation instead of a single averaged line, showing exactly how the prediction moves for that specific data point as the feature varies.

Each line on an ICE plot represents one company, one loan, or one trading day, holding all its other features fixed at their actual observed values, and tracing what the model would have predicted if only the feature of interest had been different. Overlaying many such lines, and optionally the average of all of them (which recovers the ordinary partial dependence curve), shows both the typical pattern and how much individual cases deviate from it.

Worked example. A default-risk model's partial dependence plot on "debt-to-equity ratio" looks nearly flat, suggesting the feature barely matters. Its ICE plot reveals that for low-revenue companies, predicted default risk rises steeply as debt-to-equity increases, while for high-revenue companies it barely moves at all — the two opposite patterns average out to something that looks flat, even though the feature is actually highly influential for a meaningful subset of companies.

ICE plots are most valuable exactly when a model includes interactions between features, since that's when the average trend and the individual trends are most likely to diverge.