Corporate life · Distributions
Does Forced Ranking Work? The Bell Curve in a Team of Ten
The short version: a bell curve describes a large group, not every small piece of it. If a company's performance really is 20% top, 70% middle and 10% bottom, a randomly formed team of ten has exactly that mix only about 12% of the time. Forcing every team into the quota mislabels 1.7 people per team on average, even when the manager knows everyone's true performance. In 35% of teams, nobody is in the company's bottom 10%, and somebody is ranked there anyway.
It is review season. Cousin Kevin's company uses a calibrated performance distribution. The rules are on the intranet, behind a login that expires every 30 days. They say each team of ten must have exactly two people rated Exceeds, seven rated Meets and one rated Below.
Kevin's team had a good year. Every one of them hit their targets. This information has been noted and has no bearing on the process. One of them will be Below. The rules do not say which one, only that it must be one, and that the manager's decision is final and will be reviewed.
The curve is about the crowd
The idea behind forced ranking comes from the normal distribution: in a big enough group, most people are near the middle and a few are at each end. For a company of 10,000, saying "about 10% are our weakest performers" may be roughly true.
The mistake is to assume every team of ten is a tiny copy of the whole company. It isn't. A small sample bounces around. Some teams happen to have three of the company's best people, some have none, some have two of the weakest. The quota pretends this doesn't happen.
How often does a team of ten fit the quota?
Let's make it as fair as possible. Suppose the company really is 20% top, 70% middle and 10% bottom, teams are formed at random, and every manager knows exactly how good each person is and simply sorts them. No bias, no favourites, no politics. Then:
| A random team of ten, perfect manager | Result |
|---|---|
| Team's real mix matches the 2 / 7 / 1 quota | 11.9% of teams |
| At least one person gets the wrong label | 88.2% of teams |
| Average number of people mislabelled per team | 1.7 |
| Nobody in the team is in the company's bottom 10%, but someone gets Below | 34.9% of teams |
The 34.9% is easy to check: each person has a 90% chance of not being in the bottom 10%, so all ten escape with probability 0.9 to the power of 10, which is about 0.35. In a team of five it gets worse: 59% of teams contain nobody from the company's bottom 10%, and one of them is labelled Below anyway.
And this is with a perfect manager. Real ratings also contain noise, mood and memory, which add errors on top of the ones the quota creates by itself.
Random teams of ten forced into a 20 / 70 / 10 quota, perfect manager
The rule that can't be satisfied
So every year, in about one team in three, a manager has to find a weakest performer who doesn't exist. They do find one, because the form doesn't accept "nobody". The person might have been in the middle of the company, or near the top. The label goes in their file. It looks exactly like a label that was earned.
Forced ranking doesn't measure the distribution. It imposes it. Then next year the company looks at its ratings, sees a perfect bell curve, and concludes that the curve was right all along.
Statistics for noobs: big-group facts, small-group lives
A percentage that is true for a large population is a statement about the average of many small groups, not about each one. "10% of people are left-handed" doesn't mean every group of ten contains exactly one left-hander; about a third contain none. The smaller the group, the more it wanders from the population figure.
If you must rank, rank across a group big enough for the percentages to mean something, or rate people against a standard rather than against each other.
Where it costs you money
Fund rankings work the same way. "Bottom 10% of funds this year" sounds damning, but in a small category with similar funds, somebody has to be last even if the difference is luck. Look at how far behind they are and for how long, not only at the position.
The moral
Kevin's manager chose the Below by a process the rules describe as holistic. It was Dave, who had been on holiday during the calibration meeting. Dave has appealed. The appeal will be calibrated against other appeals, of which 10% must be rejected. There was only one appeal.
What this model can't tell you
- It's a model. Real teams aren't random: hiring and transfers can cluster strong or weak people, which makes the mismatch bigger or smaller.
- Performance isn't one number. We assumed a single true score per person; real work is messier.
- Large groups are different. Across hundreds of people, a quota fits the real mix much better; the problem is applying it to every small team.
Kevin's company, Dave and the quota are an illustration. The figures come from our own model: performance normally distributed across the company, a 20% / 70% / 10% quota, teams of ten formed at random and a manager who sorts them perfectly. The exact match is a multinomial probability (11.9%), the rest is a simulation of 400,000 teams. Educational content, not legal, career or financial advice.