After getting frustrated with the inaccuracy of traditional forecasting models, we finally figured out how to forecast revenue in a way that felt real. Not just in a spreadsheet, but in the day-to-day rhythm of customers deciding whether to stay or leave. We had to create thresholds, build clustering and survival models, and set validation metrics.
But in the end, we built something the business could actually use. And it worked.
This four-part series is about how we built a forecasting tool that uses behavior-driven segmentation and risk-aware modeling to finally get accurate predictions, which then inform resource allocation and strategic decision-making.
Before we could start to rethink our model, though, we first had to face a hard truth: forecasting was not working, at least not in the way we needed it to.
For a long time, we were projecting revenue using well-intentioned assumptions, simple averages, spreadsheets, and legacy logic. These were familiar tools, but they unintentionally ignored how customers actually behave.
This meant that risk went unnoticed until it was too late, timing was treated as static, and forecasts were often disconnected from what was happening across billing, engagement, and support.
We weren’t necessarily doing anything wrong, but we were far from seeing the full picture. So we set out to change that.
In Part 1, we lay the groundwork with behavior thresholds and indicators. Not hard-coded guesses, but contextual breakpoints that tell us when something has changed with an account or when a customer starts drifting from normal.
In Parts 2 and 3, we use those triggers to group customers based on how they arrived at their risk level, not just where they landed. And in Part 4, we bring it all together, showing how behavior-driven segmentation and risk-aware modeling finally make forecasting real enough to inform resource allocation and strategic decision making.
With the new forecasting model in place, we predicted who was at risk, why, how much time before they likely drop, and the stakes if we fail to act on time.
One of the most reliable forecasting directions we identified was the amount of retained and lost revenue in the presence of pre-churn trigger, informing actionable interventions. We proved that forecasting only becomes trustworthy when it's built on customer grouping in clusters, similar behaviors, risk scores, and survival signals that move with time.
If you're creating or using financial forecasting to inform business decisions, these articles are for you. If you’ve ever been burned by clean charts that hide messy realities or inaccurate forecasting that leads you astray, you’re in the right place.
And if you have your own data frustrations to vent, I get it. You’re not alone. Reach me at [email protected] to talk about solutions that can relieve your daily data headaches.
Visit our newest service page to discover what Impact's AI Consulting Services can do for you.
The Lost Art of Context: Foundations Before Forecasts
To understand the stake, we didn’t start with forecasting. We started with a question. What exactly counts as risk? And the answer: much more than money.
Before we could model, segment, or predict, we had to make sense of our existing customer signals, spread across multiple platforms. We had over 95,000 billing records.
We had customer engagement logs, CRM trails, and buyer intent scores. But no one agreed on what a drop in these metrics meant in terms of client retention.
When did a CRM silence go from normal to worrying? When did slow payments signal financial risk? These were the sorts of questions we were aiming to answer.
In fact, forecasting couldn’t work until we answered those questions. That’s why our process began with thresholds rather than models.
When Everything Looks Normal, Nothing Stands Out
From the start, we pulled from existing data. We had contract age, billing volume, engagement frequency, and had calculated metrics like days since last CRM touch and MRR trends. On top of that were the external signals like buyer intent.
But none of it was meaningful without context.
At first, we tried generic rules, flagging things like intent scores below fifty, contracts older than six months, tenures under one year. These felt like smart defaults, but they failed because what looked risky in one service line was totally normal in another.
Risk wasn’t absolute. It was contextual.
At that point, exploratory data analysis seemed worth a try. We segmented by department, by contract type, by engagement level. Then we started flagging behavioral breakpoints that showed real friction.
For example:
- Intent scores below thirty were rare and aligned with churn.
- CRM gaps longer than seven days revealed silence in fast-paced teams.
- Customers stuck in early funnel stages weren’t progressing.
Strategic value scores below or equal to one often point to customers who were never a good fit to begin with.
These breakpoints came from the shape of the data. Each one became part of our trigger system, a set of six behavioral thresholds that covered engagement, financial health, lifecycle timing, and strategic fit. These triggers didn’t diagnose churn but signaled pre-churn troubles, the kind that often show up weeks or months before a decision is made.
These triggers then became inputs to two critical systems. The first was our churn model, which we called Version 6.3. A churn model, at its core, is built to answer a binary question. Will this customer churn, yes or no? It’s a classification model often used to plan retention campaigns, guide renewal teams, and flag silent accounts that are drifting away.
A model that uses historical patterns to assign a probability. That probability acts as a signal to sales, marketing, and customer care teams on who to target, when, and why.
We also used the same triggers in our survival model, Version 7. Survival modeling works differently. It goes beyond asking if a customer will churn and predicts when. Think of a patient's diagnosis. A churn model might tell you who’s likely to get sick, while a survival model estimates how long until symptoms show up. In our case, it helped us understand urgency and time-to-act windows. This allowed us to identify contracts that were likely to churn soon versus those that had more time.
By feeding the same triggers into both models, we got two views of reality. The churn model told us who to worry about, and the survival model told us when to act. So when both pointed to the same contract, it validated the strength of the signal. That alignment gave us confidence that our thresholds were doing their job.
Most importantly, those triggers became the raw material for our clustering method. Each contract had its own sequence of fired triggers and its own behavioral fingerprint. Those fingerprints became our map.
That’s when things got clearer. Customers shared risk scores and much more. They shared paths to risk. And those paths told us when to act. Because thresholds were our smoke detectors. They surfaced time-to-act windows and gave us the signal that the churn process had started. But there was still time.
Test, Validate, Proceed
Before moving to the next phase of the process, we validated all of the risk scores. We tested everything: we used Brier scores to measure how accurate our probability estimates were, we used area under the curve (AUC) to see how well our model separated churners from stayers, and we used Recall to understand how many risky customers we could correctly catch. Finally, we tested our churn and survival scores as covariates, which means we treated them as inputs to other models to see if they held up under pressure.
And once we had it working, the hard part was mostly behind us. We made it simple enough for a dashboard with clear scores, meaningful flags, and stories behind every spike. Stories our sales and CS teams could act on.
Forecasting doesn’t begin with a line on a chart. It begins with a customer who stops showing up the way they used to. And if we can catch that moment early, we can change the story.
If you want your forecasting to work, start with the thresholds. Start where behavior begins to shift. And listen closely. Because the data isn’t just flowing. It’s pulsing. It’s alive.
Discover what Impact's AI Consulting Services can do for your business.
