Sample Size and Regression to the Mean in Baseball Statistics

By 4Baseball Staff · November 5, 2026 · 10 min read

Why a hot month proves little, how many plate appearances it takes for strikeout rate, walk rate and batting average to mean something, and how to regress a number toward the mean, with real historical data.

A hitter who goes 20-for-50 has a .400 batting average, and no one who understands baseball believes he is a .400 hitter. The harder question is how much of that .400 is real. The answer is a small amount, and the amount depends on which statistic you are looking at: fifty plate appearances tell you something about a hitter's strikeout rate, almost nothing about his batting average, and a great deal less about his power than a few big swings suggest.

Two ideas explain all of this. Sample size says that a statistic computed from few chances is mostly noise. Regression to the mean says that an extreme result, especially from a small sample, is likely to be followed by a less extreme one, because part of what made it extreme was luck. This article explains both with plain arithmetic, shows how fast different statistics become reliable, and tests the idea against real data from the historical record.

Why small samples lie

Start with the simplest case, a coin. Flip it ten times and you will often see seven heads, which is a 70 percent rate from a coin whose true rate is 50. Flip it a thousand times and seven hundred heads would be astonishing. The coin has not changed. The sample got bigger, so the luck averaged out.

A batting average works the same way, with a twist: the hitter's true rate is not .500 but something nearer .250 to .300, and each at-bat is a weighted coin. The standard error of an observed average, the typical distance between what you see and the truth, is the square root of p times (1 minus p) divided by the number of at-bats, where p is the true rate. Here are the figures for a hitter whose true average is .300.

The first row is the point. A true .300 hitter has a perfectly ordinary chance of hitting .200 or .400 over fifty at-bats. Anyone who reads those numbers as signals will be wrong most of the time. Even a full season, with 500 at-bats, leaves a band of about 80 points of batting average. That band is the reason that a .280 hitter and a .320 hitter cannot be told apart from one season of results, and the reason the batting title is partly a lottery.

This is only the sampling error. Real hitters also change over time, play in different parks and face different pitchers, so the true rate itself moves. The table gives a floor on the uncertainty, not a ceiling.

What it means for a statistic to stabilize

Analysts describe the speed at which a statistic becomes reliable with a number called its stabilization point. The usual definition is the sample size at which half of the variation you see among players is real skill and half is luck. At that point, the statistic has a reliability of 0.5. Below that size, more than half of what you see is noise. Above it, more than half is signal.

The key formula follows from that definition. If k is the stabilization point, the reliability at n chances is n divided by (n plus k). A hitter who has had exactly k chances is half signal and half noise. One who has had three times k chances is 75 percent signal. The same ratio gives you the weight to put on a player's own record when you estimate his true talent:

estimated talent = (observed total + league average rate times k) divided by (chances + k)

In words, you add k imaginary chances at the league average to the player's real record. That is the entire method. Everything else is a question of choosing k.

The values of k below are rounded approximations, taken from published research by Russell Carleton and others that has become standard in baseball analysis. Different studies produce somewhat different numbers, and the figures are for hitters in plate appearances unless marked at-bats. Treat them as a guide to the order of magnitude, not a constant of nature.

For pitchers, the same research puts strikeout rate near 70 batters faced, walk rate near 170 batters faced and batting average on balls in play near 2,000 balls in play. The last is the key number in modern pitching analysis, because it says that a pitcher's results on balls in play are almost entirely noise over any plausible span.

Why the numbers differ so much

Strikeout rate becomes reliable in a few weeks because it depends almost entirely on the two players involved: the hitter's contact skill and the pitcher's stuff. There is no fielder, no park and little luck between the pitch and the strikeout.

Batting average is the opposite. It includes strikeouts, but it also depends on what happens to every ball that is put in play, and that depends on where the fielders stand, how hard and at what angle the ball was hit, the park, the weather and sheer chance. A ball hit well can be caught, and a ball hit weakly can drop. Over a season, those effects wash out only partially, which is why 910 at-bats, more than a full-time player gets in a year and a half, are needed for the average to be half signal.

Put those figures against a normal season. A regular might have 550 at-bats and 600 plate appearances. His strikeout rate in 600 plate appearances has a reliability of 600 divided by 660, about 91 percent, so a year of strikeout rate is almost entirely skill. His walk rate is about 83 percent reliable. His on-base percentage is about 57 percent. His batting average, at 550 divided by 1,460, is about 38 percent. A hitter's batting average for one season is therefore more than 60 percent luck and circumstance.

Regressing a number: worked examples

Here are four worked examples. The players are hypothetical, and the league rates used are simply illustrative.

A hot start. A hitter is 34-for-100 (.340). Assume a league average of .250 and k = 910 at-bats. The estimate of his true average is (34 + .250 times 910) divided by (100 + 910), or (34 + 227.5) divided by 1,010, which is .259. A hundred at-bats of .340 move the estimate by nine points.

A good full season. A different hitter is 150-for-500 (.300). With the same league mean, the estimate is (150 + 227.5) divided by 1,410, or .268. A .300 season is evidence of a better-than-average hitter, but the best guess is a .268 hitter.

A veteran's hot start. The correct prior for a veteran is not the league average but his own history. Suppose he has a career average of .290 over 4,000 at-bats, and he starts 34-for-100. Using .290 as the prior mean, the estimate is (34 + .290 times 910) divided by 1,010, or .295. The history dominates, as it should.

A strikeout rate. A hitter strikes out 25 times in his first 100 plate appearances (25 percent). The league rate is 22 percent and k = 60. The estimate is (25 + .22 times 60) divided by 160, or 23.9 percent. Because strikeout rate stabilizes quickly, a small sample moves the estimate much further than it did for batting average: here, the player's own record gets 62.5 percent of the weight (100 divided by 160), against 9.9 percent in the first example.

The method has limits. It assumes the league average is the right place to regress toward, which is a better assumption for a rookie than for a player with a track record. It also assumes true talent is constant over the sample, which is shaky for young players who are improving and old players who are declining.

What real data shows

The formulas are theory. The historical record lets us test them. For the seasons 1990 through 2019, take every pair of consecutive seasons in which the same player had at least 400 at-bats in both, which gives 3,451 pairs, and compare the first year with the second. This is a rough check, and it only includes players who kept their job, but the pattern is clear.

A correlation near 1 means that players keep their ranking from year to year, which is a sign of skill. A low one means the ranking reshuffles, which is a sign of luck. Strikeout rate is the most stable of the four, and batting average is the least, exactly the order the stabilization points predict.

The regression itself can be seen by sorting. Take the 10 percent of those hitters with the highest batting average in the first year. As a group, they hit .330 in year 1. In year 2, the same players hit .303, while the whole group's average barely moved, from .280 to .278. The top decile gave back about half of its distance above the group mean. The bottom 10 percent went the other way, from .233 to .257.

Compare strikeout rate. The 10 percent of hitters with the highest strikeout rate struck out in 26.0 percent of their plate appearances in year 1, against 15.9 percent for the whole group, and 24.7 percent in year 2. They regressed only about 13 percent of the way back. A high strikeout rate is a persistent trait; a high batting average is substantially luck.

One caution. The comparison includes only players who got 400 at-bats in two consecutive years, which tends to exclude hitters who fell off, so the real regression is, if anything, a little larger than shown.

What regression is not

Regression to the mean is often misunderstood in three ways.

1. It is not a force. Nothing pulls a hot hitter back. The numbers fall because the first sample was a mixture of skill and luck, and the luck does not repeat. A player who hit .340 over a hundred at-bats and is a true .259 hitter is not "due" to go cold.
2. It does not predict a slump. The best forecast for the next hundred at-bats is the regressed estimate, not something lower. A player with a true talent of .259 will hit about .259 from here on. The hot start is simply gone from the forecast, not reversed.
3. It does not mean skill is an illusion. The hitters at the top of the batting-average list really are better than those at the bottom. That is why the top decile is still at .303 in year 2, well above the group's .278. Regression removes the luck and leaves the skill.

The same error runs in the other direction, which is the gambler's fallacy: believing that a streak must be balanced by an opposite streak. Coins have no memory, and neither does regression. Each plate appearance has its own probability, and the past matters only because it is evidence about that probability.

Teams and records

The same arithmetic applies to teams. A team that is truly a .500 club has a binomial standard deviation over 162 games of about the square root of 162 times 0.25, which is 6.4 wins. About 95 percent of the time such a club will finish between roughly 68 and 94 wins, purely by chance. A 94-win team may be a .500 team that was lucky, and a 68-win team may be a .500 team that was unlucky.

That is why analysts are cautious about a fast start. A 20-10 record is 67 percent after 30 games. The standard error at that point is about 9 percentage points for a .500 team. It is a good sign and not much more. It is also why projections for the following season regress a team's record toward .500, and why run differential, which is less affected by close-game luck, is often preferred to wins as a guide. For how park and league context change the raw numbers, see park factors explained.

Using better information

The practical lesson is to prefer statistics that stabilize quickly, and to look under the surface of those that do not.

• Look at strikeouts and walks first. A hitter's early-season plate discipline is meaningful within weeks.
• Check the quality of contact. Measurements of how hard and at what angle a ball is hit stabilize much faster than batting average, because they remove the fielders. See exit velocity, launch angle and barrels and expected statistics such as xBA, xSLG and xwOBA, which apply exactly that idea.
• Be skeptical of batting average on balls in play. For pitchers, it needs about 2,000 balls in play to stabilize, so a low figure usually means good luck.
• Remember age. A young player may be improving and an older one declining, so his true talent is a moving target; see aging curves.
• Use a long record where one exists. A veteran with thousands of plate appearances has told you most of what you will learn. A pitcher with a long, steady career, such as Juan Marichal, shows how a big sample looks: the early-season noise washes out of the career line.

Common questions

How many at-bats do I need before I trust a batting average? More than a full season. On the commonly published estimate, about 910 at-bats are needed before the average is half skill and half luck. For a quick read on a hitter, use strikeout and walk rates instead.

Is a .400 start meaningful? It is meaningful for the strikeout and walk rates behind it, and slightly meaningful for the average itself. After 100 at-bats, a .400 hitter's regressed true-talent estimate is under .270 if the league average is .250.

Does regression apply to pitchers' earned run averages? Yes. ERA depends on balls in play, sequencing and defense, so it regresses heavily. Pitcher strikeout and walk rates regress much less.

What does a box score tell me, then? A single game is a sample of three or four plate appearances for each hitter, which tells you almost nothing about talent, though it is the right record of what happened. See how to read a box score.

Why do projection systems predict a decline for a player who had a career year? Because part of a career year is luck, and projection systems regress it. The player is not expected to get worse; the luck is simply not expected to repeat.

Are these stabilization numbers exact? No. They are estimates from particular data sets, and different methods give different values. Use them to rank statistics by speed, not as thresholds.

Where to go next

To put these ideas to use, look up players on the players pages, compare rates on the leaders page and run your own numbers in the stat calculator. The stats and analytics topic page collects further reading on the numbers behind the game.