Every B2B sales team that has been operating for a year or two has something valuable sitting in their CRM that most teams never use systematically: a record of exactly what a buying account looked like before it became a customer. Company size, industry, technology stack, what signals were visible in the 60 days before close, how long the sales cycle was, what the first touchpoint was.
That data is a fingerprint of your buyer. Training a scoring model on it means you are scoring your open pipeline against a pattern derived from your own wins, not against a generic ICP definition that a vendor constructed from aggregate industry data. The difference in predictive value is substantial.
This article is about what to include in that training set and how to think about feature selection, not about machine learning infrastructure. The engineering choices that power a production scoring model are a separate topic. What matters here is understanding what inputs produce a model that ranks accounts correctly.
The two types of features a closed-won model needs
Scoring model features fall into two categories: firmographic (static properties of the account) and behavioral (dynamic signals that the account produced before or during the sales cycle). Both are necessary. Firmographic features alone produce a model that tells you which accounts fit your ICP profile. Adding behavioral features produces a model that tells you which accounts are currently acting like accounts that have historically bought from you.
Firmographic features to include:
Company size (by employee count or revenue band, depending on how your ICP is defined) is the most consistently predictive firmographic variable. The key is consistency: if you measure company size by employee count, use the headcount that existed at the time of close, not the current headcount. Accounts grow, and a model trained on current headcount when the deal was closed at a smaller stage will produce misleading correlations.
Industry classification matters, but requires careful handling. The standard industry codes (SIC, NAICS, or vendor-proprietary codes) lump together companies that behave very differently. If your ICP is more specific than a top-level industry code can capture, you may need to create a custom classification that matches how your buyers actually differ from each other. A model trained on "software" as a category will not distinguish between companies that are relevant buyers and companies in the same category that have never bought from you.
Technology stack, where available, is often underutilized. Certain tooling combinations (CRM platform, data infrastructure, specific integration categories) are strong predictors for B2B software products that sit alongside existing workflows. Including current technology stack in your firmographic feature set, even as a binary presence/absence variable for key tools, can substantially improve model precision.
Behavioral features: what the account was doing before it closed
Behavioral features are where the closed-won model gains most of its advantage over a static ICP filter. The question you are asking is: in the 60 to 90 days before a deal closed, what observable signals was the account producing?
The challenge is that most CRMs do not systematically capture this data unless you built the processes to record it. Common behavioral features that can be reconstructed from CRM history include:
Job posting activity at time of close: Was the account actively hiring in roles that correlate with your product's use case? Engineering hiring spikes, new operations or analytics roles, and VP-level additions in relevant departments are strong indicators of organizational buying activity. This data requires an external source (job board monitoring) to capture reliably, but it can often be reconstructed historically by comparing the hiring activity that was visible at close date against current company size.
News events in the 90-day window: Funding announcements, expansions, acquisitions, and executive changes that occurred in the period preceding close. These are extractable from news databases if you have the company names and close dates from your CRM. Even a rough annotation (did a significant news event occur in the 90 days before close?) can be a useful binary feature.
Sales cycle length and touchpoint patterns: How many touches preceded the close? What was the gap between first touch and first meaningful response? These patterns, while not directly applicable to scoring open accounts, are useful for calibrating where accounts are in a buying cycle and which ones are likely to convert quickly versus require a longer nurture.
What to exclude and why
Some features that seem predictive are actually measuring your sales team's behavior rather than the buyer's. Deal size is a good example. If you include deal size as a feature in your closed-won model, the model will learn to score accounts that look like they could produce large deals highly. But deal size is partly a function of how your reps negotiated, what discounts were applied, and how the scope of the deal evolved during the sales cycle. It is not a property of the account that exists before the sales cycle starts. Including it can produce a model that scores accounts by their projected revenue rather than their actual buying readiness.
Similarly, features that are only knowable after significant sales engagement (number of stakeholders involved, champion strength, procurement involvement) should not be in the training set for a cold-outbound scoring model. Those features are useful for forecasting once an opportunity is in your pipeline, but they are not observable at the point when you are deciding which cold accounts to prioritize.
The training set size question
How many closed-won deals do you need to train a model that produces reliable scores? The honest answer depends on the number of features, the diversity of your buyer base, and how predictively distinct your won deals are from your lost or churned ones.
A rough practical threshold: fewer than 30 deals produces a model that is fitting to noise rather than signal. At 30 to 80 deals, the model can identify broad patterns but should be treated as one input rather than the primary ranking signal. Above 100 deals, the model is robust enough to serve as a primary prioritization signal, particularly when combined with real-time behavioral monitoring.
For early-stage teams with fewer than 30 deals, the right approach is to use firmographic matching and signal-based heuristics rather than a supervised model. As the deal history grows, those heuristics can be progressively replaced by learned weights from the actual won pattern.
How signal features connect to the training data
The full picture of a closed-won model is not just what the account looked like at close, but what it looked like in the period before first touch. This is the information that makes the scoring actionable for outbound, not just for pipeline forecasting. If your model learns that accounts with a specific hiring pattern in the 30 to 60 days before first touch had a higher conversion rate to opportunity, that becomes an actionable signal: when you see that pattern in an open account today, that account should surface as high-priority for outbound.
Building this connection between your historical patterns and your real-time signal monitoring is what closes the loop between the scoring model and the daily prioritization question. The CRM data tells you what to look for. The signal monitoring tells you which open accounts are currently showing it.
