Signal vs. Noise: Filtering the Data That Actually Predicts Deals

Modern revenue teams can capture almost everything. Every page view, email open, feature click, meeting, and support ticket. The instinct is that more data means better prediction, and in practice the opposite usually happens. The handful of behaviors that genuinely predict a deal get buried among thousands that merely correlate with being alive on the internet. You end up with scoring models that look sophisticated and predict nothing useful.

The teams whose scores actually get used are the ones willing to throw most of the data away. That's harder than it sounds, because noise is seductive. It arrives in volume, it's easy to measure, and it produces satisfying charts. Learning to ignore most of it is the skill.

More data usually makes prediction worse

Adding weak or spurious variables to a predictive model actively hurts, in a few specific ways.

With enough variables, a model will find patterns in past deals that are pure coincidence. It looks brilliant on historical data and falls apart on new deals, because it memorized noise instead of learning structure. That's overfitting, and it's the default outcome when you feed a model everything.

Real predictors also get diluted. Their weight gets spread thin across dozens of near-useless inputs, and their influence on the score fades.

And a model fed everything produces a precise-looking number that stakeholders trust more than they should. Precision reads as accuracy even when it isn't. The CRO sees a tidy two-digit score and doesn't ask what went into it.

Decades of applied statistics point the same way: a small set of strong, well-understood predictors beats a large set of weak ones almost every time. The work is subtraction.

Three tests before a behavior gets scored

Most captured events fail at least one of these.

Does it discriminate? A real signal behaves differently in accounts that convert versus accounts that don't. If a behavior shows up equally in winners and losers, like opening the monthly newsletter, it carries no predictive information no matter how easy it is to measure.

Is it causal, or just a mirror? Many behaviors correlate with buying because they're caused by intent that already exists, not because they create it. Visiting the pricing page correlates with buying, but the visit reflects a decision already forming. It doesn't move a lukewarm account. Separating leading indicators from reflections of intent is subtle and it matters.

Is it stable? A signal that predicted deals last year but not this year, or works in one segment and not another, is probably noise that got lucky. Real signals hold up across time and cohorts.

A behavior that passes all three is worth scoring. Everything else is telemetry you can store but shouldn't weight.

Context flips the meaning

The same raw event can be pure noise in one frame and a strong signal in another, which is why raw thresholds and simple alerts perform so badly.

Take a 40% drop in product usage. For a brand-new trial account still onboarding, it's noise. Usage is erratic at that stage by nature. For a mature account 45 days from renewal that's been steady for two years, it's a loud churn signal. Same event, opposite meaning.

Good signals normalize against each account's own baseline, lifecycle stage, and segment instead of applying one global rule. Alerts trained on raw thresholds generate so much fatigue precisely because they strip out the context that gives an event meaning, a distinction explored in what a revenue signal actually is.

Building the filter

Start from outcomes. Define the thing you want to predict (conversion, expansion, churn) precisely, then work backward to the behaviors that discriminate for it. Don't start by cataloguing every event you happen to collect. We've watched teams spend a quarter on the catalogue and never get to the model.

Validate against holdouts. Test candidate signals on data the model has never seen. A signal that only works in-sample is overfitting wearing a disguise.

Keep it interpretable early. A simple model whose logic a RevOps leader can explain in a pipeline review builds trust and exposes bad inputs faster than a black box. Sophistication can come later, once the basics hold up.

Prune. Periodically remove inputs that have stopped discriminating. Signal decays. A model is a garden, not a monument.

The filtered set of validated signals is the raw material for action. Clean signals feed the ranking and routing described in turning scattered data into next best actions, and they depend on a unified data foundation to exist at all: the single source of truth that joins events across systems. Orchestration platforms such as Revnewo handle the joining and the continuous scoring so teams can spend their time validating which signals actually predict, rather than wrangling pipelines.

Key takeaways

• Feeding a model everything makes it worse. Weak inputs cause overfitting, dilute the real predictors, and produce scores people trust for no reason.
• A behavior earns a place in scoring if it separates winners from losers, plausibly causes the outcome rather than just reflecting it, and keeps working across time and segments.
• Context decides. A usage drop means nothing in week one and everything at day 45 before renewal. Raw-threshold alerts miss that and burn people out.
• Start from the outcome, validate on holdout data, keep the model explainable, and prune what stops working.

The goal isn't to see everything. It's to see the few things that matter, clearly. If you'd rather spend less time wrangling telemetry and more time acting on what predicts deals, looking at how revenue orchestration surfaces validated signals is a sensible next step.