How do I validate a scoring model before trusting it?

Validate a scoring model by backtesting it against your real wins and losses, reviewing a sample of scored accounts by hand, and checking that no single signal is biasing the result, before you ask a single rep to work from it. A score the team does not trust is a score they ignore.

This guide covers proving a score. For handling the data gaps validation exposes, see how do I score with incomplete data.

How do I backtest a score?

Apply the model to accounts as they looked before they closed, and check whether your actual wins scored high and your losses scored low. If your best customers do not land near the top, the weights are wrong.

Why review a sample by hand?

Numbers hide surprises. Reading twenty high-scoring and twenty low-scoring accounts with a rep almost always catches something, a data gap inflating scores, or a great account scored low because of a missing field. Pay particular attention to consumer-vertical accounts, aesthetics, dental, behavioral, where registries are thin, and check whether their high and low scores hold up once web, social, and review signal is the real driver. This review is standard before any SignalRx model goes live. What makes those scores dependable once live is that they combine external public context with your own internal signals, your revenue, support, and product-usage data tracked over time, with inference deciding which source to trust when they disagree, and where a record comes in sparse SignalRx fills the gaps from public sources such as NPPES taxonomy and CCN-keyed CMS data so it still becomes scoreable, writing the resulting scores and Tags into Salesforce, your system of record, as Revenue Plays and lists.

What biases should I watch for?

What does this look like on one account?

Illustrative, based on how SignalRx customers work. During a hand review, Harborview Radiology Group, a physician-owned group, 5 locations, specialty radiology, showed as Low Fit at 40 with a health tier of Maintain. The reviewer found a blank location-count field was scored as zero. Once treated as unknown and enriched, the account recomputed to High Fit at 79, which is the kind of surprise a backtest alone can miss.

Key takeaways

• Backtest against real wins and losses before trusting a score.
• Review a sample by hand with a rep to catch surprises.
• Watch for dominant signals and data gaps that distort the rank.

See it in action: Close Better Leads with SignalRx.

Related: How do I build a Fit Score? and How do I explain a Fit Score to sales?