College Basketball Efficiency Handicapping: The Four Factors and Where Ratings Mislead

Adjusted efficiency ratings are the best public tool in the sport. They are also the same tool the book used to price the game. Your edge is not in reading them. It is in knowing exactly where they break.

Every serious college basketball bettor uses adjusted efficiency ratings. So does every sportsbook. That symmetry is the central problem of this post, and confronting it honestly is more useful than pretending a public rating is a secret weapon.

If you and the book are reading the same numbers, you cannot beat the book by reading them better. You beat the book by knowing what the numbers cannot see, and by recognizing the specific, repeatable situations where a rating is describing something other than team quality.

In Post 65 we used efficiency as an input without examining it. This post opens the box.

01What Adjusted Efficiency Actually Measures

Offensive efficiency is points scored per 100 possessions. Defensive efficiency is points allowed per 100 possessions. Both are tempo-free by construction, which is the entire reason they exist. A team averaging 84 points per game at a frantic pace can be less efficient than one averaging 66 in a grinding system, and points per game cannot tell you which is which.

The word doing the heavy lifting is adjusted. A raw efficiency figure tells you what a team did. An adjusted figure estimates what it would do against an average Division I opponent. That correction is essential, because a team that played four guarantee games in November and a team that played four high-major road games have raw numbers that mean nothing next to each other.

The adjustment works by simultaneous estimation. Every team's rating depends on its opponents' ratings, which depend on their opponents' ratings, and the system solves for a set of values that best explains all the results at once. It is elegant, it is well-tested, and it has a specific vulnerability we come to in Section 09.

02How the Ratings Are Built

The public systems differ in detail but share an architecture worth understanding.

• Inputs are possession-level. Points, possessions, and location for every game.
• Opponent adjustment is iterative. Ratings are solved against each other rather than assigned in one pass.
• Margin is used but capped. Most systems damp the influence of blowouts so that a 40-point win does not count four times as much as a 10-point win. This is deliberate and it matters more than most bettors realize.
• Location is weighted. Home, away, and neutral are treated differently, usually with a league-average adjustment rather than a venue-specific one.
• Recency may be weighted. Some systems weight recent games more heavily. Others treat the season as one sample. Know which one you are reading.

The NCAA's own NET rating sits alongside these for selection purposes. It uses game results, strength of schedule, location, efficiency, and a capped scoring margin. Note that NET is a selection tool, not a predictive one. It is designed to sort resumes, not to forecast Tuesday's game, and using it as a predictive rating is a common and costly category error.

03Factor One: Effective Field Goal Percentage

Dean Oliver's four factors framework explains most of what happens in a basketball game, and shooting is by far the largest. Effective field goal percentage weights three-pointers at 1.5 times a two-pointer, correcting the obvious flaw in raw field goal percentage.

Roughly forty percent of the explanation for who wins a basketball game sits here. It is also, as we established in Post 65, the least predictable of the four from one game to the next.

That combination is the single most important tension in basketball handicapping. The biggest driver of outcomes is the one you can forecast least reliably.

The resolution is to separate the two components. Shot selection is stable. How many threes a team takes, how many rim attempts it generates, how many contested mid-range jumpers it settles for. These are system properties and they persist. Shot conversion is unstable. Whether the open threes go in on a given night is close to random around a true mean.

04Factor Two: Turnover Rate

Turnovers per possession, roughly twenty-five percent of the explanation. A turnover is the worst possible outcome of a possession because it produces zero points and frequently gifts the opponent a transition opportunity.

Turnover rate is more stable than shooting and less stable than tempo. It is heavily guard-dependent, which makes it one of the properties most disrupted by roster turnover, and it is the factor most responsive to opponent scheme. A pressing defense as described in Post 68 attacks this factor directly.

Two handicapping applications. First, a team with a single reliable ball handler carries a specific vulnerability that season-long ratings partially mask, because the vulnerability only manifests against certain opponents. Second, a defense that forces turnovers at a high rate produces variance in both directions, since turnovers create transition points but pressure also concedes easy baskets when broken.

05Factor Three: Offensive Rebound Rate

The share of a team's own missed shots it recovers, roughly twenty percent of the explanation. College basketball has far more variance here than the NBA, because size and effort differentials between programs are enormous.

We covered the critical technical point in Post 65 and it bears repeating, because getting it backward is common. An offensive rebound does not create a possession. It extends one. The effect is on efficiency, not on tempo, and a dominant offensive rebounding team does not play faster as a result.

Offensive rebounding is a partly philosophical property. Some staffs send everyone to the glass. Others retreat immediately to prevent transition. That choice is stable year over year and it is visible on film long before it is visible in the numbers.

06Factor Four: Free Throw Rate

Free throw attempts relative to field goal attempts, roughly fifteen percent of the explanation and the smallest of the four. It measures how consistently a team gets to the line.

Its importance for a bettor exceeds its weight, for reasons that have nothing to do with efficiency. Free throws stop the clock, which affects pace. They come with fouls, which affects rotation availability. And they concentrate at the end of close games, which is the game-script effect we built into totals in Post 65.

A foul-drawing offense meeting a foul-prone defense is one of the cleanest over-leaning matchups available, and it is a matchup that season-long efficiency ratings capture only in aggregate.

07Where Ratings Mislead: The Sample Problem

An adjusted rating in late November is built on six or seven games against wildly uneven competition. It looks like a precise number. It is an estimate with error bars wide enough to drive a bus through.

Ratings need roughly ten to twelve games before they carry real signal, and they need games against varied competition rather than ten repetitions of the same mismatch. A team that opened with eight home games against overmatched opponents can rank in the national top twenty-five having proved nothing.

The practical stance is not to avoid November. It is to recognize that both you and the market are working from the same thin sample, so the disagreements you find are less likely to be edges and more likely to be shared uncertainty. Bet smaller, or spend the month building the roster knowledge that pays in January.

08Where Ratings Mislead: Blowouts and Margin Capping

Rating systems cap the influence of large margins deliberately, to prevent a team from inflating its number by running up scores against inferior opponents. This is correct design and it creates a specific blind spot.

A team that wins by 35 and a team that wins by 22 against the same opponent receive similar credit. Usually that is right. Occasionally it conceals a genuine difference in quality that only manifests in blowouts, such as elite rotation depth or a bench that is itself high-major caliber.

The related distortion runs the other way. Garbage time is played by benches, and bench minutes are included in the efficiency calculation at full weight in most systems. A team that plays eight minutes of garbage time in ten different games has roughly a game and a half of its season rated on lineups that will never play in a competitive situation.

09Where Ratings Mislead: The Connectivity Problem

This is the most important and least discussed limitation, and it is unusually relevant this season.

Opponent adjustment requires the graph of games to be connected. Teams need to play opponents who play opponents who eventually play everyone else, so the system can triangulate. When those connections are sparse, the adjustment has less to work with and the resulting ratings carry wider error.

Mid-majors are chronically under-connected. A team that plays fourteen non-conference games against similar competition and then eighteen conference games has very few edges linking it to the top of the sport. Its rating is not wrong, but it is less certain, and the uncertainty runs in both directions.

As we covered in Post 67, this season adds a second layer. Twenty-seven programs changed conferences. Old connections were severed and new ones created. For one season the graph is sparser than usual and every rating built on it inherits that extra uncertainty, including the book's.

10Where Ratings Mislead: Roster Change and Injury

A season-long rating describes the team that played the season. It does not describe the team taking the floor tonight if that team is missing its best player.

The gap is largest in exactly the cases you care about most. A star who has missed the last four games is still fully represented in a season-long rating built on twenty-five. Some systems offer roster-adjusted variants. Most public displays do not, and a rating that has not been adjusted for a significant absence is describing a team that no longer exists.

Same logic in reverse for a returning player, for a team that has settled a rotation after an unstable start, and for the first-year coaching staffs from Post 68 whose season-long number includes two months of installation the team has moved past.

11Recency: Should You Weight Late Games More?

Yes, with discipline. Teams genuinely improve and decline within a season, so a rating that treats a November game and a February game identically is throwing away information.

The discipline is that recency weighting is also how you talk yourself into chasing streaks. A team that won its last four is not necessarily better than it was. It may have played four weaker opponents, or shot 41 percent from three across a fortunate stretch.

The rule that keeps this honest: weight recent games more heavily only when you can identify the mechanism. A rotation change, a returning starter, a scheme adjustment, a freshman who has visibly figured it out. If you cannot name what changed, the recent stretch is probably variance and weighting it is chasing.

12The Public Ratings the Market Already Uses

Several public systems publish adjusted efficiency, each with its own construction: KenPom, Bart Torvik, EvanMiya, and Haslametrics are the most widely referenced, and the NCAA's NET sits alongside them for selection purposes.

Use more than one. When they agree closely, the estimate is well-supported and there is unlikely to be an edge in disputing it. When they disagree meaningfully, that disagreement is telling you the underlying data is ambiguous, and ambiguity is where a specialist with genuine information can find something.

What none of them can do is watch a game. They cannot see that a starter is playing through something, that a freshman has stopped turning it over, or that a building is genuinely hostile in a way the league-average adjustment misses. That list is your job.

13Building an Improvement on Public Ratings

You are not going to out-model the book. You are going to adjust a shared model with information it does not contain. The realistic improvements, in rough order of value:

1. Roster adjustment. Manually discount a rating for absences the season-long number still includes. The largest and most reliable adjustment available.
2. Venue-specific home court. Your own table from Post 64, replacing the league-average constant.
3. Garbage-time discount. For teams with many blowouts, mentally widen the error bars rather than treating the rating as precise.
4. Mechanism-based recency. Weight recent form only where you can name what changed.
5. Style interaction. The matchup logic from Post 68, applied on top of the ratings rather than instead of them.
6. Connectivity humility. Treat mid-major and newly-realigned ratings as less certain in both directions.

None of these are exotic. All of them are things a model running 5,000 games cannot do and a person following thirty teams can.

14Common Efficiency Mistakes

• Treating the rating as proprietary. The book has the same number. Reading it is table stakes, not an edge.
• Using NET as a predictive rating. It is a resume-sorting tool designed for selection, not a forecast.
• Trusting November numbers. Six games against uneven competition is an estimate, not a measurement.
• Weighting recent shooting. The least stable factor is the one bettors most often extrapolate.
• Ignoring roster changes. A season-long rating describes a team that may not be taking the floor.
• Treating offensive rebounds as tempo. They extend possessions rather than adding them.
• Reading decimal places as precision. A rating of 112.4 is not meaningfully different from 111.8.
• Using one system. Disagreement between systems is itself information about how solid the estimate is.

15Your Efficiency Workflow

1. Pull adjusted offense, adjusted defense, and adjusted tempo for both teams from at least two systems.
2. Check agreement. Close agreement means a solid estimate. Divergence means ambiguity worth investigating.
3. Check sample adequacy. Fewer than ten to twelve varied games means treat everything as provisional.
4. Adjust for roster. Absences, returns, and rotation changes the rating has not absorbed.
5. Check the four factors individually. Two teams with identical net ratings can be built completely differently, and the matchup depends on how.
6. Apply style interaction. Does either scheme attack the other's weakest factor?
7. Estimate possessions, then project scores. The method from Posts 64 and 65.
8. Apply venue and situation. Your table, not the league default.
9. Now compare to the market. If you cannot name the specific thing the rating missed, you do not have a bet.

16The Bigger Picture

There is a version of analytical handicapping that consists of finding the best public rating and betting whenever it disagrees with the number. It does not work, and the reason is simple: the book read the same rating and priced accordingly. Any disagreement is either information the book has that you do not, or it is noise.

The version that does work treats the rating as a well-built baseline that is blind in specific, knowable ways. It cannot see tonight's roster. It cannot see this particular building. It cannot see that the last four games were a different team. Those blind spots are stable, they are enumerable, and they are exactly where a specialist operates.

Which returns to the argument this section has made since Post 63. The rating covers 350 teams. You cover thirty. That is not a disadvantage. It is the only arrangement under which you know something the number does not.