Precision at the top

Precision at the top 20% is the closest number on the model card to what a rep experiences.

What it represents

Take the top 20% of scored records. Precision is the share of them that actually converted.

Where AUC describes ordering and lift describes improvement over baseline, precision describes the raw hit rate inside the slice your team works. If precision at the top 20% is 10%, then one in ten records in that slice converted.

Precision alone is not a quality signal. A 10% hit rate is strong against a 4% baseline and unremarkable against a 9% one. That is why the card prints the multiple underneath, and why the two numbers should never travel apart.

How it’s calculated

ax1om sorts the held-out evaluation records by score, takes the first 20%, and averages the real outcomes in that slice. The baseline underneath it is the conversion rate across the whole held-out set.

Lift at the top 20% is the same measurement expressed as a ratio: precision divided by baseline. They are not two independent findings. If precision is 10% and baseline is 4%, lift is 2.5x by definition.

This also fixes the ceiling. When your baseline is below 20%, the highest precision the top fifth can reach is baseline divided by 0.2, because that is the point where every conversion in the dataset has been pulled into the slice. At a 4% baseline that ceiling is 20% precision, or 5x lift. Nothing scores higher.

What it means for you

Worked example. Baseline 4%, precision at the top 20% is 10%, so 2.5x baseline. Against a 20% ceiling, the model reaches about half the concentration available. Put another way, 2.5x across a fifth of the list means half of every conversion in the dataset sits inside that top fifth. Your team reaches half the conversions in a fifth of the calls.

Two readings to avoid. Precision is measured on held-out history, so treat it as an estimate of future hit rate, not a promise of one. And a low precision with a healthy lift is not a broken model, it is a rare-conversion dataset. When the baseline is 1%, even a 4x model puts you at a 4% hit rate, and the honest framing to your team is that they are working a much better list, not a warm one.