Skip down to the research record

Dynamics 365 Sales scoring needs 2 deal-count minimums

Quota Forecast Editorial Team · Published · 12 min read

Dynamics 365 Sales predictive opportunity scoring needs at least 40 won and 40 lost closed deals, plus 40 final-stage deals for a stage-level model.

Key takeaways

  • Count closed won and closed lost deals separately, within the time frame you plan to use.
  • Confirm that each side reaches 40 before you plan setup.
  • If you want a stage-level model, count closed deals in your last stage as well.
  • Clean duplicates and bulk closures before you trust the totals.
  • Confirm your license and any existing model with your admin or vendor.

If you lead revenue at a small B2B software company, you have probably been pitched predictive opportunity scoring. The idea is simple. A model looks at the deals you have open today, compares them with deals you already finished, and tells you which ones look likely to close. For a team with a stretched sales calendar, that can help you decide where to spend time, which deals need a manager's attention, and which ones are quietly drifting.

The catch is that a predictive model can only learn from history. If your closed deals are few, the model has very little to compare against, and the scores it produces may not be worth acting on. Before you put effort into setup, check whether you have enough closed deals to train the model at all.

Microsoft publishes concrete thresholds for its predictive opportunity scoring feature in Dynamics 365 Sales. This post walks through the two deal-count minimums in that guidance, shows how to check them against your own records, and covers what to do when you come up short. The numbers belong to Microsoft's product. Use them as the bar to test against, and ask your CRM vendor what applies to the system you run.

What the model is actually learning from

Microsoft describes the feature in plain terms: predictive opportunity scoring uses a predictive machine learning model to calculate a score for open opportunities based on historical data. The important words are historical data. The score on an open deal is only as informed as the record of deals that have already finished.

It helps to picture what the model is doing. It takes the information recorded on past deals, notes how those deals ended, and looks for patterns that separate the wins from the losses. When an open deal shows a similar pattern, the model gives it a higher or lower score. You do not need to understand the internals to make a sound decision about setup. You do need to understand the limit: a model cannot learn a pattern from deals it has never seen finish.

This is why the number of closed deals matters more than many revenue leaders expect. A team can have a large, active pipeline and still have a thin history of outcomes. Open deals tell you what is happening now. Closed deals tell you what happened before. A scoring model needs the second kind of information in volume, and it needs both wins and losses.

The first minimum: 40 won and 40 lost

The first rule is the one most teams hit first. Microsoft states that your organization must have created and closed at least 40 won and 40 lost opportunities during the time frame selected in the Train with opportunities from the past field of the scoring model.

Three details in that sentence change how you should count.

First, won and lost are separate counts. A team with plenty of wins but only a handful of losses does not pass. A model that sees only successes has no contrast to learn from, so the losses carry as much weight as the wins in the rule.

Second, the count is tied to a time frame. The rule refers to the window you select in the Train with opportunities from the past field, not to your whole CRM history. A team that has been selling for years may still fall short if only a small share of its closed deals sits inside the window the model will train on.

Third, the wording is created and closed. The deals have to have been created and closed during the selected period. A deal created before the window and closed inside it may not count the way you expect, so check the exact definition in your environment rather than relying on a rough tally.

Consider a worked example. Suppose you choose the last year as your window, and your CRM shows a healthy total of closed deals. Split them by outcome, though, and the lost side comes up short of the 40 floor. The total looks comfortable, but the rule applies to each side separately. You would need to widen the window, keep closing deals, or both. Nothing about the team's selling has changed. The count is simply too small for this rule.

Forty is a floor, not a target. Meeting the minimum means you satisfy the requirement to train the model. It does not tell you that the resulting scores will be accurate. That is a separate question, and you answer it by reviewing the scores against actual outcomes once the model is running.

The second minimum: 40 closed deals in your last stage

The second rule applies when you want a stage-level model rather than a single score across the whole pipeline. Microsoft says that to set up a model for each stage, you must have at least 40 closed opportunities in the last stage of the business process in the selected time frame.

Start by identifying your last stage. In Microsoft's wording it is the last stage of the business process. Your CRM may label it differently, such as a closed stage, a won or lost status, or a final stage in a pipeline you have customized. Whatever the label, the count is of closed opportunities that sit in that final stage within the selected window. For how other tools label these stages, see Open, won, and lost stage rules in 4 sales tools.

Why would a team want a stage-level model? Ask your admin what it adds in your version, since it asks more of your data. The rule asks for enough finished deals at the end of the process, not just deals scattered across earlier stages.

Read the second rule as attached to stage-level models. Microsoft's rule is about setting one up, so confirm with your vendor or admin what an overall model requires in your version before you plan around it.

Here is the trap many small teams fall into. Your total closed count may look fine, but if most of your history ends in an earlier stage because deals are archived, parked, or closed outside the normal process, your final-stage count can be much lower than your overall count. Pull the final-stage number on its own. Do not assume it matches the headline total.

The gap between the two counts is often the most useful thing you learn from this check. A low final-stage count can point to a process problem, such as deals that never get a clean close, not just a data problem. Fixing that process may do more for your forecast than any model.

Consider a second worked example. Suppose your CRM shows a large number of closed deals in the window, which clears the first minimum comfortably. Yet only a minority of those ended in the final stage, because the rest were closed out early or left in an earlier stage. A stage-level model based on that final stage may fall short of the second minimum, so count it before you plan one.

Set the window before you count

Both rules point back to the same setting: the time frame. That makes the window a decision, not a detail, and you should make it before you tally anything.

Use these steps to set it:

  • Decide what period reflects how you sell today. A shorter window keeps the history fresh but may hold too few finished deals. A longer window holds more deals but may include old pricing, retired products, or a sales motion you have since changed.
  • Test more than one window. Count closed won and closed lost deals for each window you are considering, and write the numbers down. Seeing how the counts change as the window grows is more useful than a single snapshot.
  • Count from a report, not from memory. Build the count in your CRM so the numbers reflect your actual records rather than your recollection of them. Hand counts drift.
  • Note the date you ran the count. Your totals change as deals close, so a count from last month may already be stale.

A useful habit is to keep a short log: the window you tested, the won count, the lost count, the final-stage count, and the date. That log turns a one-time check into a record you can revisit when you decide whether the model is ready.

Clean your counts before you trust them

A count can pass the minimum and still describe the wrong history. Before you rely on the numbers, look for the usual distortions.

Duplicate deals are the most common. When the same opportunity is entered twice, it inflates the total and gives the model two copies of one outcome. Search for near-identical names, accounts, and amounts within the window, and merge or remove the copies according to your own policy.

Bulk closures are the second. If someone marks a batch of stale deals as lost at the end of a quarter, those losses may reflect a cleanup decision rather than a buyer's decision. They still count toward a total, but they teach a model something other than how buyers behave. Note how many such closures happened and when. For checks on bulk changes, see Bulk-edit sales records: 4 checks before saving changes.

Skipped stages are the third. Reps who jump a deal from the first stage to the final stage leave a thin trail of stage history. That matters most for a stage-level model, because a stage with little history gives the model little to work with. Look at how often deals move through each stage in a normal way.

Mixed pipelines are the fourth. If you sell two different products through two different sales processes, decide whether they belong in the same history. Combining them can help the counts pass while blending two different buying patterns into one model.

None of this requires special tools. It requires an afternoon with a report, a written policy for how you treat duplicates and bulk closures, and a willingness to admit that some of your history is not as clean as you assumed. Write down each decision you make, because the next person to review the model will need to know why the counts look the way they do.

If you fall short, you have options

Failing a minimum is useful information. It tells you which lever to pull. Here are the common choices.

Wait and keep closing. If you are close to the floor, the simplest path is time. Keep working your pipeline, record outcomes accurately, and recount on a set schedule. Put the recount date on a calendar so it does not depend on someone remembering.

Widen the window, with eyes open. A longer window can add won and lost deals quickly. The tradeoff is that older deals may reflect conditions you no longer have. If you widen the window, compare the older deals with recent ones before you accept the longer history as representative.

Ask about an overall model instead. If you meet the first minimum but not the second, ask your admin whether an overall score is available to you for now. A stage-level model is more specific, but it asks for more finished deals at the final stage. Starting simpler is a reasonable choice for a small team.

Build a non-predictive view in the meantime. You do not have to wait for a model to get a clearer forecast. Many teams start with simple measures from their own history, such as how often deals in each stage have actually closed, and use those to sanity-check the pipeline. These measures are not predictive scoring, but they are honest and easy to explain to your team.

Whichever option you choose, write down the reason. A future reviewer, possibly you, will want to know why the model was not turned on earlier or why the window was set the way it was.

Check your license and existing setup

Two administrative checks come before any counting project.

First, licensing. Microsoft says that if you have a Dynamics 365 Sales Enterprise license, you can enable predictive opportunity scoring in lead and opportunity scoring quick setup. The quick setup route is tied to that license in the documentation, so confirm your license level with your administrator or vendor before you plan around it. Do not assume a lower tier includes the feature. Ask.

Second, existing models. If your organization already built a model, its age matters. Microsoft says that if you're using a model that you created in a version of Dynamics 365 earlier than 2020 release wave 2, delete the model before you create a new one. Deletion is not something to do on a guess. Find out who owns the old model, what it currently feeds, and whether anyone reads its scores, then decide with them.

The same page also notes a storage change: as of 2020 release wave 2, the application writes the opportunity scoring data to the Predictive Score (msdyn_predictivescore) table and no longer writes to the opportunity entity. If your reports or integrations read scores from the opportunity record, that dependency may need a review. This is the kind of detail a revenue leader should hand to the admin team with a clear question, rather than solve alone.

Who should own the decision

Revenue leaders often treat scoring as a tool problem, something the admin team switches on. It is also a process decision. Someone has to decide which window represents the business, how duplicates are handled, and what counts as a closed deal for scoring. Those choices affect the forecast more than the setup screens do.

A workable split looks like this. The revenue leader owns the question: are we ready, and what will we do with the scores? The sales operations or CRM admin owns the counts and the data cleanup. Sales managers own the review of whether the scores match what they see in live deals. Write those roles down, even if the team is three people. When the recount shows a pass or a fail, everyone knows who acts next.

Be clear about what you want from the model before you turn it on. If the goal is to focus reps on the deals most likely to close this month, say so. If the goal is to flag stalled deals earlier, that is a different use, and it may depend on different parts of your history. A model that is ready for one purpose is not automatically ready for another.

Take one step today

Run a closed-deal report for your chosen window today, and write down the won, lost and final-stage counts with the date. Any count below 40 tells you which rule to work on first.

Frequently asked questions

Do I need both minimums? Start with the 40 won and 40 lost rule; the 40 final-stage rule applies only if you want a stage-level model.

Does meeting 40 mean the scores are accurate? No. Meeting the minimum lets the model train, but you still need to compare its scores with actual outcomes.

Sources