Predictive deal scoring is deal scoring done by a model: it estimates each open deal's probability of closing, and often when, by comparing the deal's recent behavior with how similar deals closed in your history. The strongest inputs are changes in stage, amount and close date, not static fields. A good score is one tested on a held-out quarter.
Predictive deal scoring is deal scoring done by a model: it estimates each open deal's probability of closing, and often when, by comparing the deal's recent behavior with how similar deals closed in your history. The strongest inputs are changes in stage, amount and close date, not static fields. A good score is one tested on a held-out quarter. Updated November 2026: added the rule-based vs predictive formula, data minimums from ORM's COO, and a worked calibration check.
This page covers the model: its inputs, how much history it needs, and how to prove it works. For the basics of what a deal score is and why it beats a rep-set close date, see what is deal scoring.
What Predictive Deal Scoring Changes
Predictive deal scoring replaces "the rep thinks this will close" with "deals like this one closed at this rate." It does not remove rep judgment. It gives the forecast a baseline, so rep input is spent on the deals where the rep knows something the data does not.How is predictive deal scoring different from rule-based deal scoring?
Rule-based deal scoring adds up points someone chose. Predictive deal scoring learns the weights from your won and lost deals and returns a probability. Both rank deals. Only the predictive version produces a number you can check against outcomes.A rule-based score, the kind you build in a spreadsheet or as a custom CRM property, is a weighted sum:
`Deal score = base + sum of (weight x signal present)`, capped between 0 and 100.
An illustrative example: start at 50, add 20 if the stage moved in the last 14 days, add 15 if the economic buyer is engaged, add 10 if the amount went up this quarter, subtract 25 if the close date was pushed this quarter, and subtract 20 if nothing meaningful changed in 30 days. A deal that moved stage, has the economic buyer engaged and had its close date pushed scores 50 + 20 + 15 - 25 = 60. In Excel that is one SUMPRODUCT of a weights row and a row of 0 or 1 flags.
The weakness is the weights. Someone picked them, and a 60 is not a 60% chance of anything. Qualification frameworks such as a deal scorecard or a MEDDIC scoring model have the same limit: they grade a deal against criteria, not against history.
A predictive model fits the weights to what actually closed. Vendors now ship both kinds. HubSpot says its deal scores "predict the probability of winning the open deals in your pipelines" on a 0 to 100 scale, and adds that they "shouldn't be used as the only factor in making business decisions." Aurea says its learned model "is calibrated so that a score of 70 means approximately a 70% chance of winning." That is the right claim to make. It is also the one to test.
What inputs does a predictive deal score use?
The strongest inputs describe how a deal is changing, not how it looks today. Deal stage, size and industry tell you what a deal looks like. Movement tells you what it is doing."The most interesting data in a CRM for forecasting is often not the data itself. It is the change in the data," says Pete Furseth, ORM's COO. ORM snapshots pipeline every day, so the model sees what changed and when. A deal in Proposal is useful. A deal that moved into Proposal yesterday is more useful.
| Input | Static version (weaker) | Change version the model should see |
|---|---|---|
| Stage | Current stage | Days since the last stage move, and the direction of the move |
| Amount | Current amount | Whether it changed, up or down, by how much, and how late in the cycle |
| Close date | Current close date | Number of pushes, and whether a push crossed a quarter boundary |
| Activity | Count of logged calls and emails | Days since the last meaningful change (stage, close date, or amount) |
| Multi-threading | Contacts on the record | Contacts engaged recently, and their seniority mix |
| Time-in-stage | Days in the current stage | Days in stage relative to the deal's group |
| Deal attributes | Size, industry, source, motion | Used to group deals, since a stage carries different odds in each motion |
- Meaningful activity is a change in stage, close date, or amount. A logged call is not. A score built on activity counts rewards busy reps, not moving deals. - The earliest warning is the absence of any signal. No stage change, no date change, no new notes. A score built on activity counts will never catch that, because nothing happened to count. - New business and renewals need different inputs. For new business, expected timing is a function of when the opportunity last changed stage. For renewals, it is a function of the renewal date, with closes usually spread across the 30 days before and after it. One model across both will misread both.
The same goes for sales motions. Pete's point about stage weights applies to scores too: an enterprise motion and a commercial or SMB motion should not share one set of weights, and neither should different products, but in practice they usually do.
How much data does a predictive deal scoring model need?
Roughly a year of history with enough wins and losses in it. Pete's practical example is a company winning about 10 deals a month with a year of data, provided there are enough closed-lost deals as well. Vendor thresholds vary: Aurea's learned model waits for "at least 150 won and 150 lost opportunities within 36 months."Clean data is not the requirement. Consistent data is. "Everyone has 'bad data' and it doesn't matter," Pete says. If your team records pipeline in a consistently imperfect way, the model learns that pattern. What breaks a model is a field that means different things to different teams, or a closed-won amount nobody can define. Below the volume threshold, run a rule-based score and move to a model as history builds.
Start with the data plumbing, not the algorithm. The model needs stage, amount and close-date history on every opportunity, plus activity linked to the right opportunity record. A daily snapshot or opportunity field history is what turns a static CRM record into the change data in the table above.Scoring a deal versus grouping it
Most deal scoring assigns a probability to an individual opportunity. That is a hard estimate to make well, because a single deal contains one outcome and it has not happened yet.
Grouping solves the sample problem. At ORM each opportunity is grouped by a machine learning model, and for each group a curve is predicted describing how long deals in that group take to close. Those curves run from 1 to 80 weeks, with most of the expectation resolving before week 12 and very few groups carrying meaningful expectation past 52 weeks. An individual deal then inherits the curve of its group.
| Approach | Unit of estimation | Weakness |
|---|---|---|
| Rep-entered probability | One deal | Optimistic, revised late |
| Stage weight | One stage, all deals | Same weight across unlike motions |
| Rule-based score | One deal, hand-set weights | Points, not a probability you can test |
| Learned deal groups | A behavioral population | Needs history, which most teams have |
Why a curve beats a score
A score answers how likely, a curve answers how likely and when. The second is what makes weekly forecasting possible, because it expresses what proportion of a pipeline should close in each remaining week rather than assigning a single number to a single date.
It also changes what counts as a warning. A deal sitting well beyond its group's curve is not merely slow, it is behaving unlike the population it was assigned to, which is a signal rather than a delay.
How do you validate a predictive deal score?
Test it on a quarter it never saw, then check that each score band closed at the rate it predicted. Pete's rule for buyers: during implementation, the vendor should hold out a quarter of your own history and measure the model's accuracy on it, not show you a demo set. The method is backtesting, and the check that matters most for scores is calibration.Three checks, in order:
1. Calibration by band. Group held-out deals by score. For each band, compare the average predicted probability with the actual close rate. Calibration gap = actual close rate minus average predicted probability. 2. Expected vs actual wins. Add up every deal's probability to get expected wins, then compare with real wins. This is the number that rolls into the forecast. 3. Brier score. The average of (predicted probability minus outcome) squared, where a win is 1 and a loss is 0. Lower is better. Three deals scored 0.8 (won), 0.3 (lost) and 0.6 (lost) give (0.04 + 0.09 + 0.36) / 3 = 0.163. Scoring every deal at 0.5 gives 0.25, so a useful model has to beat that.
Worked example (illustrative). A model scores 320 deals at the start of a held-out quarter:| Score band | Deals | Avg predicted | Expected wins | Actual wins | Actual close rate | Gap |
|---|---|---|---|---|---|---|
| 0 to 20% | 120 | 10% | 12.0 | 14 | 11.7% | +1.7 pts |
| 20 to 50% | 80 | 35% | 28.0 | 26 | 32.5% | -2.5 pts |
| 50 to 70% | 50 | 60% | 30.0 | 27 | 54.0% | -6.0 pts |
| 70 to 80% | 40 | 75% | 30.0 | 22 | 55.0% | -20.0 pts |
| 80% and up | 30 | 88% | 26.4 | 25 | 83.3% | -4.7 pts |
| Total | 320 | 126.4 | 114 | 35.6% |
Before trusting any score, ask the question Pete puts to every AI vendor: is there traceability back to the raw data? A score you cannot explain deal by deal will not survive a forecast call.
Using Scores in Forecast Reviews
Predictive scores are most valuable when they disagree with rep assessments. If the model scores a deal at 75% and the rep calls it a commit, alignment is good. If the model scores a deal at 30% and the rep calls it a commit, that is the conversation that prevents a forecast miss. The structured disagreement between AI score and rep judgment is where deal slippage gets caught before it hits the forecast.Build the score into your weekly pipeline review process. Flag any deal where the model score and the rep forecast category differ by more than two tiers. Require the rep to explain the gap with specific evidence. If the evidence is compelling, trust the rep. If the evidence is vague, trust the model.
Keep both views on purpose. Some ORM customers run a rep and manager roll-up in Clari alongside ORM's statistical forecast. That gives leadership two perspectives: what the field believes will happen, and what the data says is likely.
How often should you retrain a deal scoring model?
Retrain often enough to catch a change in buying behavior, and recheck calibration every quarter. ORM retrains its models monthly. Markets move: in 2026, Pete says, cycle length has been the biggest culprit as buyers delay decisions, and a model trained on faster quarters will overrate today's pipeline. Rerun the calibration check above each quarter. If deals scored 70 to 80% are closing well below that, the model is stale.Expansion deals carry their own signals, which is the job of an expansion propensity model, not a new-business deal score.
At ORM, models are trained on your own historical sales performance in 4 to 6 weeks, and the forecast they feed targets 95% accuracy from day 1 to day 90 of the quarter.
See how long deals take to close by group, what counts as meaningful deal activity, and the sales cycle length guide.
Frequently Asked Questions
What is deal scoring?
Deal scoring gives each open opportunity a number that says how likely it is to close. Rule-based deal scoring adds up points that an admin assigns to signals. Predictive deal scoring lets a model learn those weights from your closed-won and closed-lost history and returns a probability. The predictive version can be checked: deals scored at 70% should close about 70% of the time.
What is the deal scoring formula?
A rule-based deal score is a weighted sum: a base score plus each signal's weight times whether the signal is present, capped at 0 and 100. In a spreadsheet that is one SUMPRODUCT of a weights row and a row of 0 or 1 flags. A predictive score has no hand-set formula. The model fits the weights to past outcomes, so the output is a probability you can test.
What signals do predictive deal scoring models use?
The best inputs describe change, not state: a recent stage move, an amount change and its direction, close-date pushes, days since the last meaningful change, and how many buyer contacts are engaged. ORM counts meaningful activity as a change in stage, close date, or amount. Static fields such as deal size, industry and source still help, but they say what a deal looks like, not what it is doing.
How much data does deal scoring need?
Less than most teams think. Pete Furseth, ORM's COO, uses a company winning about 10 deals a month with a year of history as a practical starting point, provided there are enough closed-lost deals too. The data does not have to be clean. It has to be consistent, because the model learns how your team records pipeline. Below that volume, start with a rule-based score.
How do you know if a predictive deal score is accurate?
Hold out a quarter the model never trained on, score those deals as they stood at the start of it, and compare with what happened. Group deals by score band and check that each band closed near its predicted rate. A model can rank deals well and still run hot, so check calibration, not only ranking. Ask the vendor to show this on your data.
What does a close curve add over a probability score?
Timing. A score says how likely a deal is to close. A curve says how likely and when, by spreading that probability across the remaining weeks. That is what makes a weekly forecast possible, rather than a single number attached to a single date. It also turns delay into a signal: a deal well past its group's curve is behaving unlike the deals it was grouped with.
Put these metrics to work
ORM builds custom revenue forecast models that turn concepts like predictive deal scoring into prescriptive action for your team.
Schedule a Demo