Come, Look at This Column: Forecast by Cohort, Not by Hope

The Illusion of the Flat Line

Come, look at this aggregate chart of our active customer count. Observe the grand, sweeping line of the total user base. It is wonderfully flat. It holds steady, month over month, varying by only a fraction of a percent. It looks like stability. It looks safe. The board of directors will look at this line, nod in approval, and assume the business is built on bedrock.

But I want you to take the spreadsheet, grab the corner of the data, and crack it open. Because the average hides the shape, and this chart is hiding a violent, churning undercurrent. If we build our strategy on this blended average, we will starve. We are not experiencing stability; we are experiencing two massive, opposing forces that happen to cancel each other out in the grand total. We must pull the layers apart. We must clean up the noise.

The Noise: Blending everyone together turns the forecast into a polite lie. When you average the behavior of a Fortune 500 company that signed a three-year contract with the behavior of a panicked freelancer on a month-to-month plan, you create a fictional metric. You create an “average customer” who does not physically exist in the real world. This noise obscures the structural integrity of your cash flow. You cannot predict the future of a blended mass, because a blended mass has no distinct physics. It simply drifts.

The Pattern: We do not guess; we extend the line. But we must draw a separate line for each distinct organism in our ecosystem. We must look at segmentation not as a marketing exercise, but as a survival tactic. Who pays the rent, and who just makes noise? We must separate the Whales from the Barnacles.

Let us look at the raw export and rebuild the forecast by cohort. First, we pull out the ‘Sleepy’ accounts. These are the legacy clients, the ones who have been with us for three years and haven’t logged in for six months. They are generating recurring revenue through sheer inertia. Their churn rate is nearly zero.

Next, we isolate the ‘New’ accounts, the cohort that signed up in the last ninety days. Come, look at their column. Do you see the void? Do you see the massive drop-off at day thirty? Their churn rate is an aggressive, terrifying forty percent.

Keep the starting cohorts fixed while following them through the same period. Later sign-ups belong in a separate layer; they must not quietly refill the original new-customer cohort.

This illustration applies the article’s 2% legacy churn and 40% new-cohort churn to the same 30-day window. Let L and N be the starting customer counts. The rates are scenario assumptions, not a measured customer history:

CUSTOMER COUNTS / ONE 30-DAY WINDOW

START                   END: RETAINED CUSTOMERS
Legacy: L ------------> 0.98L
New:    N ------------> 0.60N

Lost: 0.02L + 0.40N
Later arrivals still active at end for a flat total:
0.02L + 0.40N

START TOTAL = END TOTAL, only with those replacements:
L + N = 0.98L + 0.60N + (0.02L + 0.40N)

A large legacy layer can make new-cohort losses look small in the total, but it cannot cancel them by staying still. A genuinely flat total needs replacement customers or growth elsewhere. The replacement count above means later arrivals still active at the end; gross sign-ups may need to be higher if some leave during the window. Count customers here; a revenue chart would also need each cohort’s revenue per customer.

When we view the data through cohorts, the polite lie of the flat line shatters. We realize that our ‘stable’ total can hide poor retention in the new cohort. The legacy base and new arrivals who stay mask those losses; now we know where to investigate onboarding.

If we forecast using the blended average, we will project steady growth, assuming new clients will behave like old clients. But they will not. By forecasting separately for new customers, repeat customers, and legacy accounts, the curve becomes sharper. The pattern reveals itself. We apply a 40% churn assumption to the new row, and a 2% churn assumption to the legacy row.

Remove the gridlines. Delete the background. What remains is the truth: a detailed map of exactly where we are bleeding, and exactly where our fortress holds strong. When we forecast by cohort, we stop hoping for a good quarter, and we start engineering one.

FAQs

What is wrong with a blended average?

A blended average takes wildly different behaviors—a loyal enterprise client and a flighty startup—and merges them into a fake, non-existent middle ground.

How do I define my cohorts?

Start with time and size. Group customers by the month they joined, and separate your top 10% highest-paying accounts from the bottom 90%.

Will forecasting by cohort take more time in Excel?

Initially, yes. You have to build the rows. But once built, it saves you hours of arguing over why the 'average' forecast keeps failing.