Same ERA, Different Game-to-Game Consistency
A worked explanation of same era, different game-to-game consistency, including the key limitation and a checkable example.
Same ERA, Different Game-to-Game Consistency
Two pitchers can finish a season with identical earned run averages yet produce dramatically different game-to-game results because ERA measures the rate per nine innings while variance measures the spread around that rate. The same era, different game-to-game consistency question answers itself once you separate the average from the dispersion. Pitcher A allows 0, 0, 0, and 12 earned runs over four games of six innings each. Pitcher B allows 3 earned runs in every one of those four games. Both finish with an ERA of 4.50. The difference is in the variance: Pitcher A's game-level runs have a variance of 27 runs²; Pitcher B's variance is 0 runs². One pitcher is predictable; the other is a boom-or-bust risk.
A Worked Example
The example above uses real inputs from the MLB definition of earned run average: (Earned Runs Allowed / Innings Pitched) × 9. For Pitcher A, total earned runs are 12, total innings are 24, so ERA = (12/24)×9 = 4.50. For Pitcher B, the same calculation yields 4.50. The game-level runs for Pitcher A are [0, 0, 0, 12]. The mean is 3 runs per game. Squared deviations: (0-3)² = 9, repeated three times, plus (12-3)² = 81. Sum of squared deviations is 9+9+9+81 = 108. Dividing by the number of games (4) gives a population variance of 27 runs². For Pitcher B, all game-level runs are 3, so every squared deviation is 0, and the variance is 0 runs². The ERA is identical; the variance is not.
Calculation Steps
To compute the variance yourself, follow the population variance formula: σ² = Σ(x-μ)² / N. This is the average of the squared deviations from the mean. The honest version of variance is that it always produces squared units, runs² in this case, which is why the square root, standard deviation, exists. The standard deviation for Pitcher A is √27 ≈ 5.20 runs. For Pitcher B it is 0 runs. That is the interpretable figure: one pitcher's earned runs deviate about 5 runs from the mean per game; the other pitcher's never deviate at all.
If you are using a sample of games rather than a full season, the sample variance formula uses n-1 in the denominator (Bessel's correction): s² = Σ(xᵢ-x̄)² / (n-1). For the same four games treated as a sample, Pitcher A's sample variance would be 108/3 = 36 runs². The sample standard deviation would be √36 = 6 runs. This correction exists because a sample tends to underestimate the true population spread. The ERA remains 4.50 regardless of which variance formula you choose.
At the point where you need to verify the ERA itself, an era calculator supplies the average-rate context. The variance calculation supplies the consistency measure. Together they answer how two pitchers can share the same season rate but differ in earned runs variability. The ERA gives you the center; the variance gives you the spread.
Why the Method Works
The explanation lies in what each metric measures. ERA is a rate, total earned runs per nine innings, and it collapses all games into one number. Variance measures how far each game's earned runs deviate from the game-level mean. A pitcher could allow 0 runs in three games and 12 in one game, yielding a moderate ERA but high variance. Another pitcher could allow 3 runs in every game, yielding the same ERA but zero variance. The two metrics are independent of each other in the sense that the same ERA can pair with any non-negative variance.
Why Variance Is Not in Runs
A common mistake is interpreting variance in the original units. Variance is always in squared units, runs² in this case. Saying "the variance is 27 runs" is incorrect. It is 27 runs². The standard deviation, which is the square root of variance, is in runs. That distinction matters when you report results. An analyst who says "the variance is 27 runs" has made the single biggest newcomer error with dispersion measures.
What a Variance of Zero Means
A variance of zero means all game-level earned runs are identical. This is mathematically possible but practically rare over a full season. Pitcher B's four-game example achieves it because every game produced exactly 3 runs. Over 30 starts, a zero variance would require the same earned runs in every single game, which is unlikely given the randomness of contact allowed and defense played.
Limits of the Variance Comparison
The variance measure has two important limits. First, variance is sensitive to outliers. A single game where a pitcher allows 12 earned runs massively inflates the variance. Change that 12 to a 6 in Pitcher A's line, and the variance drops from 27 runs² to 6.75 runs² even though the ERA changes only slightly. The ERA, being an average, smooths out the outlier; the variance highlights it. This is what Casella and Berger call the influence of extreme values on the sum of squared deviations. Second, variance cannot be computed for a single game. You need at least two data points for a sample variance, and ideally many more for a meaningful estimate.
Failure Case: Zero Innings Pitched
If a pitcher has zero innings pitched in a season, the ERA is undefined because you cannot divide by zero. The variance also cannot be computed because there are no game-level earned runs to measure. This is a structural limit of both metrics, not a flaw in the comparison. Any season with at least two games and positive innings pitched allows the calculation.
When Variance Misleads
Comparing variances across datasets with very different means can mislead. The coefficient of variation (CV = σ/μ) is the correct tool for relative variability, but it is meaningless if the mean is near zero. For earned runs allowed, the mean is rarely near zero over a season, so CV can be useful. But the variance itself remains an absolute measure in squared units, and it should only be compared when the means are similar. Pitcher A and Pitcher B have the same mean (3 runs per game), so the direct variance comparison is valid.
| Metric | Pitcher A | Pitcher B |
|---|---|---|
| Game 1 earned runs | 0 | 3 |
| Game 2 earned runs | 0 | 3 |
| Game 3 earned runs | 0 | 3 |
| Game 4 earned runs | 12 | 3 |
| Total earned runs | 12 | 12 |
| Total innings | 24 | 24 |
| ERA (per 9 innings) | 4.50 | 4.50 |
| Game-level mean (runs) | 3.0 | 3.0 |
| Population variance (runs²) | 27.0 | 0.0 |
| Standard deviation (runs) | 5.20 | 0.0 |
Who This Explanation Serves and Who Should Skip It
This explanation suits anyone who needs to understand why a season rate does not tell the whole story about a pitcher's game-to-game performance. It serves statistics students who need to compute variance by hand and interpret it, analysts using Excel who need to choose between VAR.S and VAR.P, and anyone who needs to explain to a manager why two pitchers with the same ERA are not interchangeable. It also serves fantasy baseball or sports analytics enthusiasts who want a concrete, step-by-step worked example using real MLB official statistics.
Skip this explanation if you need a deep proof of unbiasedness for sample variance; that belongs in Casella and Berger's Statistical Inference. Skip it if you need a full course in probability theory; that belongs in Blitzstein and Hwang's Introduction to Probability. Skip it if you need the Cramér-Rao bound or any advanced estimation theory. Two pitchers can have the same season rate but very different variability in earned runs allowed, using variance as the consistency measure and ERA as the rate measure.
Common Questions
Can two pitchers have the same ERA but different variance in earned runs?
Yes. ERA is a rate per nine innings; variance measures the spread of game-level earned runs around their mean. The two metrics are independent. One pitcher can allow 0, 0, 0, 12 runs; another 3, 3, 3, 3. Both have an ERA of 4.50 but variances of 27 runs² and 0 runs².
What is the unit of variance for earned runs allowed?
Squared runs (runs²). Variance is always in squared units of the original data. The standard deviation, which is the square root of variance, is in the original units (runs). Reporting variance in runs is the most common mistake newcomers make.
Why does sample variance use n-1 instead of n?
The n-1 denominator is Bessel's correction. It makes the sample variance an unbiased estimator of the population variance. Without it, the sample variance tends to underestimate the true spread. For a sample of four games, the population variance would use 4; the sample variance uses 3.
What is a failure case for this ERA and variance comparison?
If a pitcher has zero innings pitched, the ERA is undefined and variance cannot be computed. If there is only one game, variance cannot be computed because you need at least two data points. The comparison also fails if the means are very different, as variance is an absolute measure.