Research
Method write-ups on predictive scoring: how to measure a model, where the measurements mislead, and enough detail to rerun the work on public data and get your own numbers.
How to evaluate a predictive lead scoring model
Which measures answer which question, how each one misleads when read alone, and seven traps that make a scoring model test well and fail in production. Demonstrated end to end on a public dataset, with the split, the model configuration and the reproduction steps published.
Method only. Carries no performance results, and the demonstration dataset is consumer bank telemarketing, not B2B.
What goes in this section
- Method: how to measure something, what the measurement is worth, and where it misleads. Written so a reader can apply it to their own data rather than take our word for anything.
- Work on public datasets, where the data is downloadable and every step can be checked by someone who does not work here.
- Undocumented steps named as undocumented rather than filled in with a plausible value, and our own methodology defects disclosed alongside everyone else's.
What does not go here: model performance figures. ax1om does not publish outcome numbers while every account is still on sample data, so these write-ups describe how to measure rather than what was measured. When results come from real customer data, with permission, they will be published as case studies, labelled as such, and they will say whose pipeline they came from.