Deal Health Scoring from First Principles

Deal health scores exist on a spectrum from useless to untrustworthy. On one end: simple heuristics like "has a next step" or "activity in the last week." On the other: black-box ML models that output a number no one can explain.

Neither extreme works. Simple heuristics don't predict outcomes. Complex models don't earn trust. The useful middle ground is a scoring system built from first principles—signals that have a clear mechanism for why they matter, weighted by empirical relationship to outcomes.

Start with Mechanism, Not Correlation

The first mistake teams make is building scores purely from data. They run regression analysis, find variables that correlate with closed-won, and add them to the model. This produces scores that overfit to noise and don't generalize.

The better approach: start with hypotheses about why something would predict close probability. Then test those hypotheses empirically.

For example, "deals with multiple stakeholder engagement close at higher rates" isn't just a correlation—there's a mechanism. Multi-threaded deals have multiple internal advocates, reducing single-point-of-failure risk. If your champion leaves, the deal can still progress through other contacts. If you only know this from data, you don't know if it's a spurious correlation. If you start from the mechanism, you can trust the signal.

The Four Signal Categories

Deal health signals fall into four categories, each capturing a different dimension of deal risk:

1. Engagement signals: Is the buyer actively engaged?

• Days since last buyer-initiated activity
• Meeting frequency trend (increasing, stable, decreasing)
• Response time to rep outreach
• Email/call ratio to baseline

The mechanism: engaged buyers close. Disengaged buyers ghost. Engagement decline predicts deal death before stage or close date change.

2. Stakeholder signals: Is the buying committee forming?

• Count of distinct stakeholders engaged
• Economic buyer engagement (any/none)
• Stakeholder role diversity (technical, business, executive)
• New stakeholder additions in the last 30 days

The mechanism: complex purchases require consensus. Deals that stay single-threaded lack the internal support to close. Stakeholder growth indicates organizational commitment.

3. Velocity signals: Is the deal progressing at a healthy rate?

• Deal age relative to segment average
• Current stage age relative to historical average
• Stage progression rate (accelerating, normal, stalled)
• Time since last stage change

The mechanism: deals have a natural rhythm. Deviations from that rhythm indicate friction—either buyer-side hesitation or missed qualification.

4. Qualification signals: Does this deal meet the criteria for closing?

• Budget confirmed (yes/no/unknown)
• Timeline stated (within quarter, next quarter, undefined)
• Decision criteria documented
• Champion identified and validated

The mechanism: deals close when the conditions exist. Missing conditions create risk that stage doesn't capture.

Weighting Signals Empirically

Once you have signals with clear mechanisms, weight them based on your historical data. The process:

1. Pull all closed deals from the last 8-12 months (closed-won and closed-lost)
2. For each deal, calculate each signal at multiple points in time (30 days out, 15 days out, close date)
3. Run logistic regression with closed-won as the outcome
4. Use the coefficients to weight each signal in your score

This approach differs from pure ML because you're constraining the model to signals you understand. The regression validates which signals actually predict outcomes in your environment and by how much.

Key finding: weights vary by company. Stakeholder count might matter more in enterprise sales than SMB. Engagement recency might matter more in fast-cycle sales. Don't copy someone else's weights—derive your own.

From Score to Action

A score that just says "72" isn't actionable. A score that says "72, driven by declining engagement and single-threaded stakeholder map" is actionable.

Build the score to surface the drivers. For any deal, a rep or manager should be able to see:

• Overall health score (0-100)
• Contribution from each signal category
• Specific signals that are below threshold
• Trend direction (improving, stable, declining)

The score becomes useful when it directs attention to specific actions: "This deal's health is dropping because engagement has declined—when is your next meeting?" That's a coaching conversation, not a number.

Calibrating to Reality

Any scoring system needs calibration. A score of "80" should mean something consistent—ideally, that deals scoring 80 close roughly 80% of the time.

Run monthly calibration checks:

1. Bucket deals by score range (0-20, 21-40, 41-60, 61-80, 81-100)
2. Calculate actual close rate for each bucket over the trailing quarter
3. If actual rates diverge significantly from implied rates, adjust weights

Common calibration issues:

• Score compression: All deals cluster between 50-70. Solution: increase weight on differentiating signals.
• Stale signals:Score doesn't change even as deals go dark. Solution: increase weight on recency signals.
• False positives:High-scoring deals don't close. Solution: add qualification gates that cap scores when criteria aren't met.

The Score Is a Starting Point

A deal health score is a hypothesis, not a verdict. It says: based on observable signals, this deal looks like it has an X% probability of closing.

The score can be wrong. The rep might know context the data doesn't capture. The buyer might have told the rep something that hasn't hit the CRM yet. These exceptions are fine—they're opportunities for the rep to add color to the quantitative signal.

The dangerous behavior is ignoring discrepancies. If the score says 40 and the rep says 90, that's a conversation: What does the rep know that the data doesn't show? Is the rep being optimistic, or is there legitimately information missing? Resolution either updates the data or updates the forecast—either way, alignment improves.

What to Do This Week

Start simple. Pick one signal from each category and track it for your top 20 deals:

• Engagement: Days since last buyer-initiated activity
• Stakeholder: Count of distinct contacts engaged
• Velocity: Days in current stage vs. historical average
• Qualification: Is the economic buyer engaged (yes/no)?

Track these weekly for a month. At month end, see which deals closed and which didn't. You'll start to see which signals differentiated outcomes—and you'll have the beginnings of a principled deal health model.