Why does the sales rep who had a record month usually have an ordinary one right after? Most of the time the cause is regression to the mean: an unusually extreme result tends to be followed by one closer to average, because part of what made it extreme was luck, and luck does not repeat on command.

You meet this whenever you pick something out because it was extreme and then measure it again. That covers the top performer on a dashboard, the campaign that spiked, the worst store in the region, and the A/B test variant that looked like a runaway winner. It decides whether you credit a meeting, a budget bump or a new process with a change it did not cause.

How far back does an extreme result fall?

Think of any measured result as two parts added together: a stable part (real skill, real demand, real height) and a noisy part (the week's leads, the weather, a lucky draw). When you choose the single best result on a board, you are partly choosing the best skill and partly choosing the best luck. Next time the skill is still there, but the luck gets redrawn, so the result drifts back toward the average.

The drift has a size, and there is a simple rule for it. Measure how far the first result sat from the average. Multiply that distance by the correlation between the first and second measurements, a number from 0 to 1 describing how consistently the same things score high twice. The answer is how far from average you should expect the second result to land.

The rule comes from where the word came from. Francis Galton grew sweet peas from seeds of known weight and compared the offspring seeds with their parents. In his own account, "the mean Filial deviation was only one-third that of the parental one," meaning offspring kept only a third of the parents' distance from the average (Galton, Memories of My Life). His 1886 study of human height, built on 930 adult children and their parents, found children's deviation was about two-thirds of their parents'. Tall parents had tall children, just not as tall. Galton called it "regression towards mediocrity," and the name stuck to the whole of regression analysis.

Two conditions make it happen. The measurement has to contain some randomness, and the two measurements have to be less than perfectly correlated. The weaker the correlation, the bigger the pull back toward the average.

Now the sales floor. A rep posts the best month on the team, and her manager sits her down to talk about repeating it. Suppose her metric's month-to-month correlation were one-half. Then before anyone says a word, you would expect only half of her lead over the team average to survive into next month. When it shrinks, the meeting looks as if it did nothing, or did harm. Meanwhile the rep with the worst month gets a warning, improves, and the warning gets the credit. None of the sources here report a month-to-month correlation for sales figures, so the real number is something you would calculate from your own history.

Three places it has fooled people

Flight instructors. In the 1960s Daniel Kahneman told Israeli Air Force instructors that praise teaches better than punishment. A veteran objected: cadets he praised for a smooth maneuver usually did worse next time, and cadets he criticized usually did better. Kahneman saw that both groups were simply returning to their usual level after an unusual attempt. His summary, reproduced on PubMed Central: "We are statistically punished for rewarding others and rewarded for punishing them."

Speed cameras. Officials credited cameras with saving about 100 lives a year. Cameras tend to go where accidents have just peaked, and peaks recede on their own. An analysis allowing for this found roughly half of the apparent reduction would have happened anyway (Select Statistical Consultants).

Knee pain trials. People join treatment studies when symptoms are bad, and osteoarthritis pain can swing by up to 7 points on a 10-point scale from day to day. A review in Cartilage cites work finding regression to the mean can account for up to a full point of improvement on that scale. In one ankle study, treatment and control groups both improved by 10 points on functional scores, which suggests most of the gain was the statistics, not the injection.

It is not a force, and it is not a market forecast

The first misreading is that numbers "settle down" because something pushes them back. Nothing pushes. As the Wikipedia entry puts it, regression toward the mean "is not based on cause and effect, but rather on random error in a natural distribution around a mean." The rep has no tendency toward mediocrity; you picked her on a lucky month.

The second is the phrase "mean reversion" in market commentary, which sounds identical but makes a different claim.

Regression to the mean Market "mean reversion"
What it claims Results picked for being extreme will, on average, be less extreme when remeasured Prices, earnings or margins stretched above a long-term trend will fall back toward it
Why it happens Luck in the first measurement is not repeated An argument about economics and valuation, which can be right or wrong
Timing Shows up on the very next measurement In Lance Roberts's words, "not the same as a price target with a date attached"

Before you credit the fix

The question to ask about any before-and-after story is: was this chosen because it was extreme? If the campaign got its budget bump after a spike, or the team got coaching after a slump, some of the next change was coming regardless.

The checks that separate a real effect from the rebound are the ones the Cartilage review recommends for clinical trials, and they carry over to a dashboard. Compare against a similar group that got no intervention. Use several baseline periods instead of one peak or trough. Do not select on the same number you then use to judge success; pick the reps to coach by one measure and evaluate them on another. When none of that is possible, say plainly that part of the change is expected drift, and estimate how much from your own month-to-month correlation.