Different verticals, the same features
We grouped our underwriting features by platform vertical to start. The cross-vertical experiments are holding up, and the same engineered features keep doing the heavy lifting — whatever the industry.
The same features
keep winning
One of the good things about working at Slate is seeing how differently platforms explore putting capital inside their own products, sometimes with genuinely original ideas.
To refine the underwriting model during integration, we grouped ML features more closely aligned with platform verticals to start. But we are now seeing associations across those groups that are becoming more and more relevant. The cohorts in our cross-vertical feature experiments are starting to hold up.
Honestly, that was not what we expected. Every time a new partner integrates we run a discovery pass, define the target for their book, and let the model tell us what predicts business performance. And the same winners keep showing up. Different industry, different transaction patterns, the same engineered features doing the heavy lifting.
The fun part is what is left after that. The last stretch of AUC almost never comes from the bank transactions. It comes from whatever the platform already knows about its own users. Usually that data adds features to the scoring set, or sharpens the features we are already using.
The next step was automating that. Candidate features get proposed off the partner's own data, they go through the same lift test as everything else, covariance is tested, and only the ones that actually beat what we already score make it into the vertical-specific model.
Same pipeline for every partner now — which means every integration makes the next one sharper.
