The Pricing Committee Pack - Model Lift

The Decision Question

This edition continues the Pricing Committee Pack series with a review of the model lift view.

Does the proposed pricing model identify risk better than the current rating plan, and does that difference matter enough to action?

A lift view belongs between the technical model build and the selected pricing action. It is the evidence page that shows whether the model is finding economically useful segmentation on data it did not train on.

What Lift Means

Broadly speaking, model lift is the value created when a model separates better risks from worse risks more accurately than an alternative model or the current rating plan. In insurance pricing, that value is usually framed as reduced adverse selection: charging closer to the actuarially fair rate by risk segment.

A simple lift chart sorts policies by a model prediction (expected loss cost for each policyholder), groups them into equal-volume or equal-exposure buckets and compares observed results across those buckets. If the model has useful segmentation power, the observed results should move meaningfully from low-risk buckets to high-risk buckets. If the chart is flat, the model has not separated risk.

A double-lift chart goes one step further. It compares two models directly by sorting policies on the ratio of proposed model prediction to current model prediction. The first and last buckets are therefore the places where the two views disagree the most. Within each bucket, the chart compares actual experience with both model predictions. A model whose predictions follow the observed pattern more closely across the deciles has stronger evidence of incremental segmentation value.

The Default Practice

A standard double-lift chart has three lines: current plan prediction, proposed model prediction and actual result, usually by decile of the proposed/current prediction ratio.

This does something important: it shows whether the proposed model is directionally right exactly where it creates the largest dislocation from the current plan. That is a much better governance question than asking only whether the model has a better aggregate error metric.

The weakness is that a standard double-lift chart can be hard for a committee to read. It asks the audience to compare three similar lines, infer which one is closer to actuals, remember the exposure behind each point, and translate the result into a pricing decision. That is a lot of cognitive admin for one chart.

Suggested Upgrade

The Differential Lift Chart extends the view by adding two practical dimensions: the change in rate index and the exposure affected.

The chart brings together three measures:

• Observed relative loss cost shows the actual experience within each decile, relative to the portfolio average of 1.00.

• Change in rate index shows the proposed model rate index less the current-plan rate index. Negative values indicate a lower relative rate under the proposed model, while positive values indicate a higher relative rate.

• Exposure % shows how much of the portfolio falls within each decile.

  
  

In this example, the proposed model identifies lower relative risk in the first deciles and higher relative risk in the later deciles.

The observed loss-cost pattern is directionally consistent with the model proposing lower relative rates in the early deciles and higher relative rates in the later deciles.

The exposure bars provide the additional context needed to judge how much of the portfolio would be affected by the change.

This allows the chart to answer three related questions in one view:

Does observed experience support the proposed model’s differentiation?

How large is the change from the current rate structure?

How much business is affected?

The result is a more decision-ready extension of the standard double-lift chart, linking model evidence to pricing consequence and portfolio materiality.

Decision Boundary

Use this chart when the committee is choosing the overall rate action. It is designed for the executive pricing result: the decision about the selected portfolio-level or class-level action.

Do not use it as evidence for segmentation, model performance, rating factor selection, individual customer impact or governance. Those belong in later pages of the Pricing Committee Pack.

Use this chart when…	Don’t use this chart when…
Comparing a proposed pricing model with the current rating plan.	Assessing the predictive performance of a model in isolation.
You need to show whether observed experience supports the model’s proposed re-ranking of risk.	You need detailed calibration, discrimination or model-validation metrics.
The decision requires both the direction of the rate-index change, and the amount of exposure affected.	You need to identify the specific customer or risk groups within each decile.
Supporting a pricing committee discussion on whether the proposed model introduces meaningful, evidence-based differentiation.	The model indication cannot be separated from subsequent judgement, smoothing, caps or commercial adjustments.
The proposed and current indices are calculated on a consistent basis and can be meaningfully compared.	The current and proposed indices use materially different definitions, coverage or normalisation bases.

The Differential Lift Chart shows where the proposed model changes the pricing signal, whether observed experience supports that change, and how much exposure sits in each decile.

Next edition: Who sits behind the lift?

Profiling the segments concentrated in the areas of greatest proposed rate increase and decrease.