Natural Variance

This page is about dunning recovery, the campaign that starts when a subscription payment fails. Cancel flow save rates vary too, but only from sample size, since every session resolves the moment it ends.

Complete cohorts. Every outcome counted.

What natural variance is

Natural variance is how much your recovery rate swings from one period to the next when nothing about your recovery process has changed. It shows most clearly month to month, where the swings are largest; rolling analysis is how you diminish them. We measure it as the gap between the highest and lowest comparable period in your complete history for a stable recovery process, not a standard deviation, not a confidence interval.

Natural Variance = Highest Period Recovery Rate - Lowest Period Recovery Rate

• Period = the unit you're comparing, a calendar month or a rolling window, held constant across the calculation
• Highest / Lowest Period Recovery Rate = the recovery rates of the best and worst comparable periods in your complete history
• Comparable periods = periods measured with the same recovery process, made only of complete cohort days

On this page the period is a rolling window, the unit Rolling Analysis builds. An account's best 30-day rolling window is 100% recovery and its worst is 66.7%. Its natural variance is 33.3 points. That is the range the business operates in.

Every comparison on this page is in percentage points. A move from 64% to 74% is a 10-point increase, not a 10% one.

The typical band

A typical account swings 20-30 points between its best and worst rolling window over a multi-year history. That is a property of subscription dunning data, not a problem with the recovery system. Smaller volumes swing more, larger volumes less.

What happens without it

You want to improve recovery, and a 10-point increase would be a big win. You look at your last completed month, see 64%, and set the target at 74%. Next month you run a new process. It's expensive, but worth it if it delivers. You hit 73% and call it a win.

Your 30-day rolling windows over the past year ran between 64% and 78%, with most between 76% and 78%. The 64% month was an outlier at the bottom of your band, and 73% is below your typical range. Without the band, attribution is a mess.

How to read the band

A bad week inside the band isn't a crisis. A good week inside the band isn't a victory. Only a move outside your range, sustained across several rolling windows, is evidence that something changed.

What drives it

Six things move a recovery rate for reasons that have nothing to do with the recovery process:

• Sample size. Fewer failed payments means a wider swing from one window to the next.
• Seasonality. Holidays, promotions, and surges in new subscribers show up in recovery months later.
• Price points. Customers on lower prices can be more likely to churn.
• Billing cycles. Annual and quarterly renewals fail and churn differently from monthly ones.
• Customer acquisition. Deep discount and unintentional signups churn more.
• Service quality. Product, support, and onboarding set how much customers want to stay.

Each one can move the band without any change to the recovery process.

Customer mix is the main engine

Each period contains a different blend of customers, and the ones who fail come from different cohorts: one is on a second monthly order, another is hitting the annual renewal from last year's promotion. Whether a customer updates a card, cancels, or ignores the emails depends on how much they want to stay, so recovery rate is partly a loyalty measurement. It moves when the mix moves, with no change to the recovery process.

Two coffee subscription companies can run an identical recovery process and post 20% and 80%. The 80% company's customers opted in knowingly for a modest discount and like the product. The 20% company's customers didn't realize they'd subscribed, or took a steep discount planning to cancel before renewal, so what looks like a payments problem is a loyalty problem.

Segmentation

Segmentation is how you narrow that variance into something actionable. A 30-point swing across the whole book often shrinks to a tight band inside a clean segment. Same data, sharper question.

Much of the noise comes from newer customers on their second or third renewal, who respond more to discounts and promotions. A promotion or viral moment two to three months ago puts a wave of first and second renewal customers into recovery now, dragging the blended rate down while nothing about the process changed.

The practical minimum is one split, customers on their first renewal against everyone else. That alone explains a surprising share of the variance. Filtering to long tenured customers, or to one subscription type, often gives a band narrow enough to act on.

Volume and window length

With low volume the band is wide and no single window is readable. The smaller the account, the longer the window it needs and the more history before a difference is distinguishable from noise.

A longer window shows less variance. 60 days is smoother than 30, and 30 is smoother than 14. But past 30-45 days the signal goes stale. Starting points and the tradeoff are on Rolling Analysis.

Don't measure it this way

• Comparing one month to the next and calling the difference performance. Two months are two points inside the band. Check the band first.
• Treating a single window as a baseline. The 64% month above was one window. A baseline is the full range of complete windows.
• Reporting a standard deviation or confidence interval as "variance." Natural variance is the high and low of your history.
• Attributing a swing to a tool or setting change before checking whether it sits inside your band.
• Ignoring a shift in customer mix. Segment before you conclude.

Check the range first, then decide whether the process changed.

---

Prerequisite: Rolling Analysis, which builds the windows this page measures the spread of.