Split-half reliability, variance decomposition

Form, and how much of it is real

Each club's season as a rolling ten-game run differential, month by month, and a split-half test of how much of the table is talent.

Computed on request from the record book (Oct 13, 2026) and 720 finals in the game log.

The finding

After about 144 games, run differential per game runs from KT's +1.22 to Kiwoom's −1.75. How much of that spread is real? Single-game margins swing so much that chance alone would spread the clubs' averages by about 0.41 runs a game; the observed spread is 0.88, which makes roughly 78% of it talent. Winning percentage splits much the same way (78%).

A second check splits each club's games into odd and even: the halves correlate at r = 0.50 (95% −0.19–0.86), implying a full-season reliability of 0.66 — consistent with the variance split, but ten clubs make the interval wide. First-half run differential predicts the second half at r = 0.63.

Ten-game form is another matter. Over ten games, chance alone moves run differential by about ±16 runs (one standard deviation), and 62% of all the rolling values below sit inside that band. Most hot and cold spells are weather, not climate.

Run differential: talent share

78%

Win%: talent share

78%

Odd–even reliability

0.66

run differential

10-game noise, 1 SD

±16

runs

Small sample. Every correlation on this page is across 10 clubs; a value of r = 0.5 from ten points has a 95% interval running from about −0.2 to 0.86.

Rolling 10-game run differential

Each panel is one club's season: the sum of run differential over its last 10 games, after every game. The shaded band is ±1 standard deviation of what chance alone produces over 10 games. Panels share a scale.

±1 SD of chance

KT+1.22/G

Samsung+0.95/G

KIA+0.73/G

Doosan+0.40/G

LG−0.02/G

Hanwha−0.10/G

NC−0.37/G

Lotte−0.44/G

SSG−0.63/G

Kiwoom−1.75/G

Run differential per game, month by month

Tinted by size and sign. Months with fewer than five games are shown faint — a handful of games says almost nothing.

Club	Mar	Apr	May	Jun	Jul	Aug	Sep	Oct	Season
KT	+3.3	+1.5	+0.3	−0.6	+1.5	+1.2	+1.8	+5.3	+1.22
Samsung	−2.3	+0.6	+2.4	+0.8	+1.4	+2.0	+0.8	−3.2	+0.95
KIA	−0.3	+0.1	+1.3	+1.4	−0.9	+2.0	+0.5	+1.6	+0.73
Doosan	−1.0	−0.4	+0.2	+0.4	+1.2	+1.7	−0.5	+1.6	+0.40
LG	−3.3	+1.3	−0.4	+1.0	−1.6	+0.3	−0.2	−0.8	−0.02
Hanwha	+0.7	−1.0	+2.4	+0.6	+0.6	−3.6	−0.6	−0.1	−0.10
NC	+3.3	−0.1	−0.9	+0.8	−1.3	0.0	−0.6	−1.9	−0.37
Lotte	0.0	−1.8	−0.2	−0.4	+0.8	+0.6	−0.9	−1.2	−0.44
SSG	+4.0	+0.7	−2.1	−2.5	−1.4	+1.0	+0.7	−0.4	−0.63
Kiwoom	−4.3	−0.8	−2.7	−1.5	−0.1	−4.2	−1.1	−0.6	−1.75

How much of the table repeats?

Correlations across clubs between two halves of their own season, with 95% intervals.

Method

The variance split compares the observed spread of clubs' run differential per game with the spread that single-game noise alone would give a season of this length:

σ²(noise) = σ²(single-game margin) ÷ G · talent share = 1 − σ²(noise) ÷ σ²(observed)

For winning percentage the noise variance is 0.25 ÷ G, a coin flip's. The split-half test correlates each club's odd-numbered games with its even-numbered games, which cancels any trend through the season, and steps the result up to a full season with the Spearman–Brown formula, 2r ÷ (1 + r). The first-versus-second-half correlation includes genuine change — trades, injuries, slumps — so it is expected to be lower.

Ten clubs is a small population for any correlation; the variance split is the steadier of the two estimates because it uses every game. The noise model treats games as independent and equally variable, which a schedule with uneven opponents, rest days and weather does not quite honour. Months are calendar months, so March and October are short.