To get where we are today with our forecasting model, we first built the foundation: context-aware thresholds that turn raw customer data into behavioral signals. These thresholds, derived from exploratory data and contract lifecycle awareness, gave structure and shape to the chaos. This was the pulse of our data. Then, we took that pulse and found patterns. We clustered customers not by score alone, but by how risk had accumulated across time, triggers, and context. We used a rarely applied algorithm, K-Prototypes, to group contracts based on both numerical and categorical behavioral signals.
Each cluster told a different story, and those stories began to unlock new ways to plan outreach, align teams, and ultimately, retain revenue.
Now, everything converges. Thresholds, triggers, survival scores, and clusters transform into a functional forecasting framework. One that’s interpretable, time-aware, customer-specific, and revenue-relevant. One that addresses a fundamental problem nearly every business faces.
Find out what AI can do for you in your organization by exploring Impact's AI Consulting Services.
The Forecasting Woes We Set Out to Solve
Forecasting in many organizations, especially in customer success, sales, and revenue operations, still begins in Excel. Take last year’s revenue, apply a percentage, adjust for churn if someone left, maybe bump it for growth targets, and call it a projection.
It’s neat, it’s familiar, and it’s dangerously blind to risk. Traditional forecasts treat all customers as equal. They don’t differentiate between stable renewals and accounts quietly drifting away. They don’t account for the timing of risk, or how it evolves across contract lifecycles and departments.
We knew this approach was failing us. And we knew our models had more to give. So we set out to build something better: a forecasting pipeline that reflects how customers behave, when they are likely to leave, and how much value is actually associated with that risk.
From Predicting Risk to Projecting Revenue
The shift started when we stopped treating forecasts as outputs and started treating them as simulations. We didn’t want to just know how much revenue we might keep. We were after something deeper. We wanted to know which clusters were at risk, how fast they were declining, and what specific actions might change the trajectory.
Here’s How We Did It
Each contract had a churn risk score from our classification model (V6.3), and a survival probability from our time-to-churn model (V7). These two layers gave us insight into which contracts were likely to churn, and when. Combined with real contract values, these scores let us compute expected retained revenue at the contract level. This was not just a prediction anymore; it was a financial signal.
We aggregated these values by behavioral cluster, each with its own decay pattern:
- Loyal customers held steady
- Silent accounts declined slowly
- Billing-risk clusters dropped off quickly if left alone.
We applied time-series models like Prophet to each cluster to both forecast growth and simulate what would happen in different scenarios.
This layered structure gave us something we never had before: forecasts that were built from the bottom up, responded to behavior, and created urgency where it was due, and calm where it was earned.
Why This Framework Is Different
Let’s be clear, we didn’t invent classification, or survival analysis, or clustering. But the way we sequenced and combined these models and the purpose we assigned to each is what made this approach novel.
We began with interpretation, looking at behavioral thresholds, not raw features. Then we layered in precision by including survival scores for timing, and churn scores for likelihood. Next came grouping: K-Prototypes clustering let us group by the path customers took to risk, not just when they were. Then, finally, forecasting: a projection engine that translates risk into revenue and supports strategic planning.
Most published frameworks focus on accuracy while only a few integrate behavior, timing, revenue, and interpretability. Fewer still use these models to enable department-specific outreach, internal alignment, or specific strategies based on cluster grouping.
That’s what we built, and that’s why this matters.
This forecasting pipeline doesn’t just respond to customer churn. It creates a system for understanding, planning, and acting across functions.
Market-Aligned, Research-Backed, Business-Ready
As we’ve studied the literature, from telecom churn models to modern frameworks like RetenNet, we’ve seen the gaps. Most pipelines stop at churn prediction.
Our approach brings together what most traditional pipelines separate by being:
- Interpretable: thresholds, triggers, clusters with names
- Temporal: survival modeling for urgency tiers
- Financial: revenue forecasting by customer and group
- Operational: outreach strategies aligned to behavioral clusters
Importantly, this approach is also modular. Each piece can be improved, replaced, or extended. We are already preparing this framework for connection to live data, which will enable rolling forecasts that evolve with real-time behavior.
A Framework Ready for Peer Review
This isn’t just a pipeline. It’s a publishable framework. Our next step is to integrity test the entire pipeline across:
- Backtested revenue vs. actual retention
- Model calibration metrics (Brier score, AUC, recall)
- Department-specific intervention outcomes
We’ll benchmark our performance against baseline methods, average-based forecasting, raw churn flagging, and static segmentation, and document the improvement in precision, interpretability, and operational usefulness.
We’ll measure impact in model terms, revenue protected, and customers retained.
Final Thought: Forecasting With Purpose
This project was born out of forecasting frustration and ended as a new way of seeing customers as evolving journeys.
We didn’t set out to innovate for the sake of innovation. We built this because our teams need answers that can be trusted, stories that can be acted on, and forecasts that are reliable enough to guide decisions.
If your organization is forecasting with averages and hoping for the best, we understand. We’ve been there. But we’re evolved.
We built this forecasting framework with intention, and we’re just getting started. Because real forecasting doesn’t start in Excel. It starts with behavior. And it ends with purpose.
Discover the potential of AI in your organization by visiting Impact's AI Consulting Services.
