Question

Which sales signals predict together.

Signals that predict together are pairs whose combined presence on a deal beats the better of the two on its own. A security review opening is useful. A security review opening on a deal where legal redlines also came back is a different and usually stronger thing. Testing for it means counting closed deals that had both, measuring the win rate, and comparing it to the stronger single signal rather than to the overall win rate.

Why pairs are a separate question from weights.

A weight says how much one thing counts. It is fitted by asking how often that signal appeared on wins against losses, one signal at a time, which is what almost every deal scoring model does.

That misses anything that only matters in combination. If a pricing page visit predicts nothing on its own but predicts strongly when a second stakeholder has already joined the deal, a model fitted signal by signal weights it near zero and never recovers the combination.

The number of possible pairs across a normal signal set is small enough to test exhaustively on closed deals. The work is counting rather than modelling, which is why this is rarer than it is difficult.

Measure against the better half, not the overall rate.

The mistake that makes a combination look like a finding is comparing it to the overall win rate. A pair containing one strong signal beats the overall rate automatically, because the strong half already did.

The question worth asking is whether the pair beats the better of its two signals alone. If security review alone wins 61 percent of the time and security review with legal redlines wins 84 percent, the combination is worth 23 points. If it wins 62 percent, it has found nothing.

Two signals that always occur together are a related trap. If they never appear apart, each alone has the same record as the pair, so the combination explains nothing its halves do not.

How much evidence a combination needs.

Enough closed deals that one outcome does not move the win rate much. On a handful of deals a perfect record is a lucky quarter rather than a pattern, and the threshold is worth setting deliberately rather than discovering later.

Enough lift that the pair sits outside the noise of the signals it is made of. A combination that beats its stronger half by a point or two has found nothing worth acting on.

Whatever floors you choose are arbitrary in the sense that any number is. They are worth writing down, because a threshold somebody can argue with is better than one nobody knows about.

What to do with a combination you find.

Not apply it automatically. A combination discovered in a few dozen deals is a candidate, and a model that promotes its own candidates into live weights without anybody looking is the kind of scoring nobody can audit.

Show the evidence next to the claim: how many closed deals had both, how many were won, and the lift against the stronger half. Somebody who sells into that market can usually tell in a second whether a combination is real or an artefact of how deals get logged.

Then let them accept it or refuse it, and keep the record of which they chose.

Questions

Why do most deal scores miss signal combinations?

Because they are fitted one signal at a time. A model that asks how often each signal appeared on wins weights a signal that only matters in combination at close to zero, and nothing in that process ever looks at pairs.

How many closed deals do you need to test a combination?

Enough that one deal does not move the result much. On a handful of deals carrying both signals, a single outcome swings the win rate by ten points or more, which makes a perfect record meaningless.

Should a discovered combination change the score automatically?

A discovered combination should not change a score automatically. One found in a few dozen deals is a candidate rather than a conclusion, so the right move is to show the evidence, let somebody who knows the market accept or refuse it, and record which they chose.

Can signal combinations be shared between companies?

Only in aggregate and only with a floor. Rolling up what several companies in an industry independently accepted is useful. Doing it with one or two companies behind it exposes a single customer model as a benchmark, which is a privacy problem before it is a statistical one.

It arrives in Slack

Every morning, each rep gets one message naming the deals on their own book that moved overnight, biggest mover first, with the reason beside each number and the deals that have gone quiet underneath.

$499 a month, or $4,500 a year.. One admin and up to 15 users. No per seat charge. Connect a CRM and it runs. No implementation fee. 14 day trial. No card.

Back test my own deals, free Try it Pricing