How Machine Learning Predicts Deal Close Dates

The close date field is the least reliable number in the CRM and the most consequential. Every coverage calculation, every quarterly commit, and every capacity plan runs on dates that a rep typed under pressure. Predicting timing from behavior instead of intent is one of the clearest places machine learning earns its cost in revenue operations.

How does a model predict when a deal will close?

It assigns each opportunity to a group of similar deals, then applies the timing pattern that group produced historically. The prediction is not about the individual deal in isolation. It is about what deals like this one have done before.

At ORM every opportunity is grouped by a machine learning model, and for each group we predict a curve for how long it will take to close. Grouping is what makes the approach work with imperfect data. A single record with a sloppy amount and a guessed date still lands in the right group based on its other attributes, and the group's history carries the prediction.

What is a close curve?

A close curve is the distribution of closing probability across weeks for a group, rather than a single predicted date. It answers what share of this group closes by week four, by week twelve, by week thirty.

Those curves run from 1 to 80 weeks, with most of the expectation happening before week 12 and very few groups carrying expectation past 52 weeks. Reading the shape matters more than reading the peak. A group whose curve is tall and narrow at week six is predictable, and pipeline in it can be planned against. A group with a long flat tail out past week 40 is a group where quarter assignment is close to arbitrary, and treating those deals as in-quarter revenue is how a forecast breaks.

Why is the rep's close date a weak input on its own?

Because it records intent, and intent gets revised. The valuable information is in the revision, not the value.

The best signal for slippage is a sales rep changing the close date. Once a deal slips from one quarter to the next, it is less likely to close at all, even when it still sits in commit. That is a behavioral fact a model can learn and a spreadsheet cannot see, because a stage-weighted model reads the current date and knows nothing about the two dates before it.

The last row is where most pipeline reviews spend their time. Activity counts feel like evidence, and reps can generate them on a deal that has stopped moving.

What counts as meaningful activity?

A change in stage, close date, or amount. Nothing else qualifies as movement on the record.

This definition also gives you the earliest warning available. The earliest signal on a deal is the lack of a signal: no activity, no data changing, no notes. From the seller's side the same pattern reads plainly. If a buyer is not returning email, not picking up calls, and not answering texts, the deal is in trouble regardless of what the stage says.

Build the weekly report around absence rather than presence. Opportunities with a close date inside the quarter and no meaningful change in 30 days are the list worth reviewing. See deal slippage for how the pattern develops over a quarter.

How long should a deal stay in the pipeline?

Twelve months without a meaningful change, and then it is inventory. ORM applies a twelve month rule for most customers.

More than 10 percent of a typical pipeline fails that test. The cost is not the storage. It is that every ratio calculated from the pipeline total is wrong by whatever share the dead records represent. A coverage number that includes a year of untouched opportunities describes a pipeline that does not exist. That is one reason coverage should be treated as an input rather than a conclusion, an argument we make in full in why the 3x pipeline coverage rule is wrong, with the calculation mechanics in pipeline coverage.

What does timing prediction change about the quarter?

It separates the revenue you can see from the revenue you have to create. A quarter has three sources: carry-over deals already in pipeline on day one that are expected to close, in-quarter deals that do not exist yet but will be created and closed inside the period, and pull-forward deals from future quarters that may close early.

Close curves make the first source computable and expose how thin it usually is. Of the pipeline carrying close dates inside the quarter on day one, roughly 20 percent typically closes in that quarter. The other 80 percent of that value does not land in the period it was assigned to. A team that plans as though its day-one in-quarter pipeline is the quarter is planning against a number that has historically been wrong by a factor of five.

That reframes the weekly forecast call. The useful questions become how much has to be created and closed inside the quarter, which groups can plausibly produce it given their curves, and what pulling deals forward costs the next period. Pull-forward is real revenue and it is borrowed, usually with a discount attached and a hole left behind.

How accurate can timing prediction be?

Accurate at the group level, approximate at the deal level, and that is the correct expectation. No model tells you that a specific opportunity closes on the 14th.

What it does reliably is tell you that a group of 60 similar deals produces a predictable share of closes across a known window, which is what a forecast needs. Across the market, forecast accuracy on new and expansion business usually sits around 90 percent, reached through manual effort that degrades as conditions shift. ORM targets 95 percent without manual adjustments and holds it from day one through day ninety, updating as the quarter progresses.

Use the deal-level prediction for coaching and the group-level prediction for planning. Teams that invert those two end up arguing with the model about individual deals, which is a debate nobody wins and which misses what the curves are actually for. Cycle-length context sits in our sales velocity guide.