ABCABC CompanyRevScoreboard · the ten things
ABC Company is fictional. Figures come from a real sales organization; every name has been changed.
07 of 10 · The Diagnostic

The Win-Signal Regression

What actually predicts a win here, tested rather than argued.

1The question

Which of your beliefs about winning hold up?

2Why it matters

Everyone in the building has a theory of why deals close: the demo, the referral, the price point. They can't all be right, and until now the loudest one won.

A logistic regression tests them all at once on your own record and reports which survive, with the evidence disclosed: how many deals, how many wins, how much weight each finding can carry.

The predictors you believe in that have no field today aren't a failure. They are the first things worth building.

What the research says

Mechanical (rule-based) prediction was about 10% more accurate than expert judgment on average. It was substantially better in 33–47% of studies; experts were substantially better in only 6–16%.

Grove, Zald, Lebow, Snitz & Nelson (2000), Psychological Assessment 12(1), 19–30 · doi.org/10.1037/1040-3590.12.1.19

Research comparing clinical and actuarial judgment shows the actuarial method to be superior.

Dawes, Faust & Meehl (1989), Science 243(4899), 1668–1674 · doi.org/10.1126/science.2648573

With fewer than about 10 outcome events per predictor, logistic regression coefficients were biased. Below the floor, the answer is “not computable.”

Peduzzi, Concato, Kemper, Holford & Feinstein (1996), Journal of Clinical Epidemiology 49(12), 1373–1379 · doi.org/10.1016/S0895-4356(96)00236-3

Confirmation bias: seeking or interpreting evidence in ways partial to the beliefs we already hold. Write your three predictors down before the model runs.

Nickerson (1998), Review of General Psychology 2(2), 175–220 · doi.org/10.1037/1089-2680.2.2.175
AEvidence grade A. Proven across many studies: a meta-analysis or a finding replicated again and again. The grade rests on the method and where the candidate predictors come from. The findings are ABC's own, graded by their intervals.
What it doesn't settle. A regression on one company's deals finds associations, not causes. And the 10-events rule is a heuristic that later work disputes; the event count is shown with every model.

3What to collect, every time

Input A number the CRM can't give you: the plan, quotas, working weeks, spend.Field Does the field exist, is it required, and who fills it in?Definition Your own definition of a thing we then compute against.Prediction Your answer now, in writing, checked against the record later.
Input
How many deals have closed, won or lost, in three yearsThe sample floor.
Definition
Anything that broke comparability mid-historyPricing, product, a CRM rebuild. The model is fit around the break.
Prediction
Your top three predictors of a winPlaced next to the model's list.
Field
A recorded reason on every lost dealWho records it, and from what list.

4How it's computed

  1. Take every closed deal. Outcome = won (1) or lost (0).
  2. Choose predictors known when the deal was created. Leave out anything recorded because the deal won.
  3. Fit a logistic regression with created-year held constant.
  4. Report each predictor as an odds ratio with a 95% interval and a p-value, plus the events per predictor.

5What ABC's record shows

What predicts a win at ABC

Logistic regression · 1,926 closed deals · 293 wins · 24.4 wins per predictor
0.5×1×2×4×8×no effectSource: Client referral: 8.7× (95% CI 4.93–15.36), p = 0Client referral8.7×p<.05Source: Other referrals / re-engaged: 4.17× (95% CI 2.57–6.76), p = 0Other referrals / re-engaged4.17×p<.05Source: Inbound (web / call-in): 3.12× (95% CI 1.85–5.27), p = 0Inbound (web / call-in)3.12×p<.05Source: Partner / broker: 2.21× (95% CI 1.39–3.51), p = 0.001Partner / broker2.21×p<.05Source: Other / blank: 2.16× (95% CI 1.02–4.58), p = 0.045Other / blank2.16×p<.05Company size: 1–19 employees: 2.07× (95% CI 1.54–2.79), p = 0Size: 1–19 employees2.07×p<.05Region: Region 2: 1.56× (95% CI 1.06–2.27), p = 0.022Region: Region 21.56×p<.05Region: (blank): 0.8× (95% CI 0.49–1.31), p = 0.372Region: (blank)0.8×n.s.Region: Region 3: 0.74× (95% CI 0.49–1.1), p = 0.139Region: Region 30.74×n.s.Region: Other regions: 0.6× (95% CI 0.37–0.96), p = 0.034Region: Other regions0.6×p<.05

Odds of winning compared with the reference: Outbound source (inside sales / cold call / networking), 20+ employees, Region 1, created 2024. Created year (2024–2026) held constant; not reported. Bars are 95% intervals. Pseudo-R² 0.102.

A client referral carries 8.7× the odds of an outbound deal, and a 1–19 employee company about 2.07× the odds of a larger one. Region matters too (all regions tested together: p = 0.004).
Size is in the model only where it was recorded. Employee count is filled on every win, so a blank size would “predict” losing. The 325 closed deals with no size are left out of this model; the odds ratios describe the 1,926 that have one.
Why amount isn't in the model. Amount is filled on 100% of wins and 53% of losses. It looks like the strongest predictor there is, but it gets filled in because the deal won. A field recorded after the outcome can't predict it.

6What a sales leader does Monday

  • Put the top predictor into the qualification conversation this week.
  • Stop arguing about the predictors that weren't significant.
  • Build fields for the beliefs that couldn't be tested.

7The founder version

Building your first CRM? Do this now, so you can answer the question in a year.

  • Record lost deals and why, from the first one. A model needs losses.
  • Keep a short list of what you think wins deals, dated. Test it at 100 closed deals.