Machine Learning Deal Grouping

Machine learning deal grouping sorts open opportunities into cohorts that resolve the same way, then predicts timing and outcome per cohort. It is the step before scoring, and getting it wrong makes every downstream prediction worse.

Why Grouping Comes First

A single opportunity carries almost no statistical weight. It has one amount, one stage, one age, and no outcome yet. There is nothing to fit a model against.

Grouping solves that by pooling. Once a deal sits with hundreds of comparable resolved records, the model has enough evidence to estimate how long that kind of deal takes and how often it converts. ORM works this way: each opportunity is grouped by a machine learning model, and for each group the system predicts a curve for how long the deal will take to close.

Learned Groups Versus Drawn Segments

Most revenue teams segment by hand. Enterprise and SMB, by region, by product line. Those splits encode what leadership believes drives behavior, and they are often close enough to be useful and wrong enough to hide things.

A learned grouping is judged by a different test: did members of the group actually resolve alike. That test surfaces combinations no analyst would have proposed. A mid-market deal from a partner referral may behave like an enterprise deal, while a large deal from a cold outbound sequence behaves like a long shot. Hand-drawn segments put those in the wrong buckets and average away the difference.

What the Groups Produce

Each group gets its own expected timing profile. In ORM's models these curves run from 1 to 80 weeks, with most of the expectation before week 12 and very few groups carrying meaningful expectation past 52 weeks.

That per-group timing is what makes aging rules meaningful. A deal is not old because a calendar says so. It is old relative to the group it belongs to. A ninety-day-old deal in a fast group has already missed its window, while the same age in a slow group is unremarkable.

What Breaks Grouping

Inconsistent definitions, not messy data. ORM's position is that every company believes its data is uniquely bad and that this belief is wrong. Garbage in does not have to equal garbage out, because consistent data still supports accurate prediction.

The real damage comes from a field whose meaning shifted. If opportunity type meant one thing before a CRM migration and something else after, the model groups on a distinction that no longer exists and inherits the confusion.

The second failure is over-splitting. Push the model toward many small groups and each one thins out until its curve is fitted on noise. Fewer, better-populated groups produce steadier predictions.

Using Groups in Practice

Review pipeline by group rather than by stage alone. Stage tells you where a rep says a deal is. The group tells you what deals like it have historically done, which is the harder input to argue with.

Feed group-level conversion into your sales forecasting math instead of a flat stage percentage, and track forecast accuracy per group so you can see which cohorts your model reads well and which it does not.