There is a belief that sounds like common sense: if you have a conversion rate for each branch, the company’s conversion rate is the average of those rates. It is taught in school as “find the mean” and it is wrong often enough to cause real damage in reports.

Three branches, one wrong answer

A loan company tracks how many enquiries turn into sanctioned loans. In one month, with invented figures:

BranchEnquiriesSanctionedRate
Coimbatore10220%
Chennai1,000909%
Pune5005010%

Average the three rates: (20 + 9 + 10) divided by 3 gives 13%. Now do it the other way, adding up the sanctioned loans (142) and dividing by total enquiries (1,510). That gives 9.4%. I checked both. The gap is almost four percentage points, and the second figure is the one that describes what happened to the company’s enquiries.

The reason is simple. The first calculation gave Coimbatore, with ten enquiries, the same vote as Chennai, with a thousand. Two lucky conversions in a tiny branch pulled the headline up by more than three points.

The rule underneath

A rate is a fraction. Fractions don’t add the way whole numbers do. To combine them you go back to the parts: sum the numerators, sum the denominators, divide once. The result is the same as a weighted average of the rates, where each rate is weighted by its own denominator. Chennai carries a weight of 1,000, Coimbatore of 10.

This applies to any ratio: conversion rate, defect rate, cancellation rate, average order value, cost per lead, attrition percentage, on-time delivery. “Average of averages” is the same trap wearing different clothes. Averaging each store’s average basket and calling it the company’s average basket will mislead whenever stores differ in size.

But sometimes the simple average is right

Here is where the myth-busting needs a counter-myth, because “never average percentages” is also too strong. The right method depends on the question. If the question is “what happened to the company’s enquiries?”, pool them. If the question is “how does a typical branch perform?”, an unweighted average of branch rates answers it, since each branch is the unit being studied. A regional head ranking branches on their process quality might prefer that view, ideally with a note that tiny branches are noisy.

So there are two legitimate numbers. The mistake is not picking one. The mistake is picking one by accident and labelling it “overall conversion”.

Where this goes wrong in real tools

  • A pivot table or BI visual with a rate column: if you ask it to “average” a pre-calculated rate field, you get the unweighted mean. Summing the numerator and denominator fields and dividing, in a calculated measure, gives the pooled rate. Many BI tools do this correctly only when the measure is defined as a ratio rather than as a stored percentage column.
  • Roll-ups. A monthly rate, averaged across twelve months, treats a festival month and a quiet month alike. Add up the year’s numerators and denominators instead.
  • Stored percentages in the warehouse. If a table keeps only the rate, you cannot rebuild it up the hierarchy. Store the counts and compute the rate at the end.

A related surprise

Pooling can also reverse a story. A product can have a better rate than a rival in every individual segment and a worse rate overall, if the mix of segments differs. This is Simpson’s paradox, and it is the same root cause: the denominators are not equal. When pooled and segmented figures disagree, look at the mix before looking for a bug.

Small denominators deserve one more warning. Coimbatore’s 20% rests on ten enquiries. Another month, with three conversions or none, it could read 30% or 0%. If you rank branches by rate, set a minimum volume before a branch is allowed on the leaderboard, or show the counts beside every percentage.

A habit that costs nothing: whenever you see a percentage in a report, ask “out of how many?” If the report cannot answer, the percentage is not ready to be averaged, compared or quoted.

Figures and names in the examples are illustrative.

Leave a Reply

Trending