Win Percentage Formula: Why It Falls Short as a Ranking

The win percentage formula — wins divided by games played — is the simplest possible summary of a season, and exactly because it is so simple, it cannot separate a team that beat strong opponents from a team that beat weak ones.

The formula itself

Win percentage is defined as

$$\text{Win\%}=\frac{W}{W+L}$$

with ties, where a sport allows them, typically counted as half a win and half a loss: W + 0.5T in the numerator, W + L + T in the denominator. It requires nothing beyond the final result of each game — no scores, no opponent identity, no schedule context — which is both its appeal and its central limitation.

Why win percentage alone can’t separate two teams with the same record

Consider two hypothetical 11-6 teams over a season. Team Ridgeline played six games against the league’s weakest four opponents and beat all six; team Harborview played a fully balanced slate against a mix of strong and weak opponents and went 11-6 against that tougher mix. Win percentage reports both teams identically — 0.647 — despite Harborview clearly having accomplished more given who it played. This is precisely the gap the ranking algorithms covered elsewhere in this series exist to close, using the structure of who-played-whom, not a bare win/loss tally.

Win percentage’s blindness to opponent strength is not a flaw that shows up only in contrived examples — it is structural. The formula has no term for who the opponent was, so no amount of real-world context can make it account for schedule difficulty without changing the formula itself.

RPI: patch win percentage with opponents’ win percentage

The Rating Percentage Index (RPI) is the most widely known attempt to fix this without leaving the win-percentage framework entirely. RPI does not replace win percentage; it blends it with two additional layers: the average win percentage of a team’s opponents, and the average win percentage of those opponents’ opponents. A common weighting is

$$\text{RPI}=0.25\times \text{WP}+0.50\times \text{OWP}+0.25\times \text{OOWP}$$

where WP is the team’s own win percentage, OWP is opponents’ win percentage, and OOWP is opponents’ opponents’ win percentage — together putting 75% of the final number on strength of schedule, not the team’s own record. The NCAA used RPI for years to help select and seed tournament fields across several sports, though it drew persistent criticism for lacking a rigorous statistical foundation and for occasionally rewarding a loss to a highly-rated team more than a win over a weak one; the NCAA discontinued RPI for Division I men’s basketball in August 2018, replacing it with the NCAA Evaluation Tool (NET), a system incorporating game results, scoring margin, and efficiency metrics instead of a pure win-percentage cascade.

Pythagorean win expectation: a different kind of fix

Where RPI keeps win-loss results as the foundation and adjusts for who they came against, Pythagorean win expectation takes a different route entirely: it ignores the actual win-loss tally and instead estimates what a team’s winning percentage should be, based only on total points scored and allowed. Bill James introduced the method for baseball, using the simple form

$$\text{Win\%}=\frac{P{F}^2}{P{F}^2+P{A}^2}$$

where PF is points (runs) for and PA is points (runs) against. Adapted to football, the exponent changes — the football version commonly uses a value close to 2.37, not baseball’s 2, reflecting football’s different relationship between scoring and winning.

The gap between a team’s actual win percentage and its Pythagorean win expectation is often more informative than either number alone. A team that has won several more games than its point differential would predict has typically won an unusual share of close, low-margin games — and because close games are less predictable than blowouts, that kind of gap is commonly treated as a signal more likely to shrink than to persist, a soft form of the same idea a formal ranking algorithm handles more rigorously by working directly from the full schedule graph.

A worked example (hypothetical numbers)

Suppose a hypothetical team, Fenwick, finishes a shortened season 6-2, for a raw win percentage of 0.750. Its eight opponents combined for a modest average win percentage of 0.45, and those opponents’ own opponents averaged 0.50. Plugging into the RPI formula:

$$\text{RPI}=0.25(0.750)+0.50(0.45)+0.25(0.50)=0.1875+0.225+0.125=0.5375$$

Fenwick’s strong 0.750 win percentage pulls its RPI up only modestly, to 0.5375, because three-quarters of the formula is weighted toward a schedule that was not especially demanding — precisely the correction a raw win percentage cannot make on its own.

Now suppose Fenwick scored 210 points on the season and allowed 160. Its Pythagorean win expectation, using the football exponent of 2.37, is

$$\text{Win\%}=\frac{210^{2.37}}{210^{2.37}+160^{2.37}}\approx 0.66$$

An expected win percentage of about 0.66 translates to roughly 5.2 expected wins out of 8 games — noticeably below Fenwick’s actual 6 wins. Read together, the two numbers tell a coherent story: Fenwick’s point differential does not fully support a 6-2 record, and its schedule was not unusually difficult either, so both independent corrections point toward the same conclusion even though they were built from completely different inputs (one from wins and opponents, the other from point totals alone).

Two different questions, not two competing answers

RPI and Pythagorean win expectation are easy to conflate because both exist to improve on plain win percentage, but they are actually answering two different questions. RPI asks “who did this team beat, and how good were those opponents” — it still counts wins and losses, just weighted by schedule strength. Pythagorean win expectation asks “how much better was this team on the field, independent of which particular games it happened to win or lose” — it discards the win-loss tally entirely in favor of point totals. A team can score well on one measure and poorly on the other; neither result makes the other wrong, since they are not measuring the same thing. The regression- and network-based ranking algorithms covered in the rest of this series are, in a sense, more rigorous descendants of both ideas at once — using the full schedule graph the way RPI gestures at, and often margin of victory the way Pythagorean expectation does, solved simultaneously, not layered on top of a plain win percentage. Massey least squares is the clearest example of the margin-of-victory side of that lineage, built as a regression on point differential, not a ratio of winning percentages; an Elo-based rating takes a third route again, updating incrementally game by game without solving a season-wide system at all.