Sports analytics got smarter, but the underdogs kept winning

Over roughly the same period, the Kansas City Chiefs kept winning games they were supposedly less likely to win.

A Sports Illustrated analysis of betting data from 2016 to 2026 compared underdog performance across the NFL, NBA, MLB, NCAA basketball and NCAA football. Kansas City finished first across all five sports, winning 20 of the 30 games in which it entered as the underdog. That works out to 66.67%.

The Baltimore Ravens were next at 61.9%, followed by the Pittsburgh Steelers at 60%.

This doesn't mean bookmakers spent a decade believing Kansas City was a bad football team. The Chiefs became an underdog only 30 times in the sample, and many of those games would have involved another strong opponent or circumstances that shifted expectations away from them.

Still, the timing is interesting. The decade in which sports became much better at collecting and processing data was also a decade in which some very familiar teams repeatedly beat pregame expectations.

So what exactly got smarter?

An NFL team plays relatively few games compared with teams in other major American sports. Even after the league expanded its regular season, 17 games leaves analysts with a small annual sample compared with 82 in the NBA and 162 in MLB.

Then there is the amount that can change between those games.

Quarterback availability can transform an entire offense. One injury on the offensive line can alter how a team approaches pass protection. Weather can matter. So can travel, rest and the strengths and weaknesses of the next opponent.

There are also fewer scoring events than in basketball. An interception, missed field goal or fourth-down conversion can have a disproportionate effect on the result without necessarily telling us much about which team would perform better over a much larger sample.

The SI rankings produced an unusual result in that context: the Chiefs, Ravens and Steelers took all three places at the top of the overall table.

Pittsburgh offers a particularly interesting contrast with Kansas City. The Steelers' percentage was slightly lower at 60%, but they accumulated 45 outright victories as underdogs during the period. Kansas City had 20.

For an analytics system, those are two different stories. One team was rarely an underdog but won at an exceptional rate when it was. Another encountered that position far more often and kept producing upsets.

Simply feeding more historical results into a model doesn't resolve the difference. The circumstances behind those results matter.

The NBA should, in theory, provide analysts with a richer dataset.

There are far more games. There are dozens of possessions in each one. Individual shooting locations, lineup combinations, defensive matchups, player movement and on-off performance can all be measured.

The league has become one of the most data-heavy environments in professional sport.

SI's NBA results were noticeably different from the NFL figures. Boston led at 43%, with 92 victories from 214 games as an underdog. Miami followed at 38.8%, winning 156 of 402.

Neither team fits neatly into the traditional image of an outsider.

Boston reached the playoffs throughout the period covered by the study and won the 2024 NBA championship. Miami made multiple deep postseason runs. There were nevertheless hundreds of individual nights when the market preferred the team on the other side.

Some of those variables are relatively easy to quantify. Others become useful only when combined with context.

Three-point shooting is a good example. A model can know exactly how frequently a player makes a shot from a particular location. It can measure the quality and volume of the attempts a team usually creates. What it cannot establish in advance is whether a 36% shooter will make two of eight attempts tonight or six.

Over an 82-game season, much of that variance has time to settle. In one game, it can decide the result.

Player availability introduces another problem. Knowing that a star is active isn't the same as knowing exactly how effective he will be after returning from an injury. A minutes restriction changes the calculation again. So does a late scratch elsewhere in the rotation.

The data can be accurate while the situation it describes keeps moving.

Sports prediction isn't built around one permanent assessment of a team.

Kansas City's rating changes. So does Boston's. New information arrives, old performances become less relevant and players improve or decline.

That creates a continuous feedback loop.

A rookie performs better than expected and future projections incorporate the new evidence. A team changes its defensive system and opponents gradually accumulate enough film to respond. An injury removes an important player, who later returns under different circumstances.

Odds move with those changes.

This also helps explain why a successful underdog can eventually stop appearing as an underdog so often. Once the market has enough evidence that a team is stronger than previously thought, future expectations rise with it.

Kansas City's sample of only 30 games is telling. As the Chiefs established themselves as one of the NFL's strongest teams, there were fewer occasions when opponents were favored against them.

The challenge for predictive systems is getting from the old assessment to the new one quickly enough.

Historical data is valuable precisely because the past contains information about what may happen next. Sports teams make that task awkward by refusing to remain the same long enough to be treated as fixed datasets.

A coaching change can alter the usefulness of older results. A trade can change several lineup combinations at once. A young quarterback may improve substantially between September and December. College programs can turn over major parts of their rosters between seasons.

Adding another million rows to a database doesn't make those problems disappear.

The rest of SI's rankings reinforce the point.

Houston led MLB with a 51.3% underdog win rate. Duke led NCAA basketball at 50%. Clemson topped NCAA football at 57.1%.

Those percentages shouldn't be compared as if each came from the same experiment. The schedules, sample sizes and structure of the games are completely different.

Baseball gives analysts an enormous number of games but places unusual importance on the starting pitcher in any single matchup. College basketball combines a relatively large schedule with young rosters and postseason games where one poor shooting night can end a season. College football offers far fewer games and, in SI's dataset, much smaller underdog samples for some of its leading programs.

A useful sports model therefore isn't simply a machine with access to a huge quantity of information. It needs the right information for the sport and the situation being analyzed.

That distinction has become more important as machine learning has entered sports analytics. Processing capacity makes it possible to examine combinations of variables that would have been impractical for a human analyst to calculate manually. Real-time feeds can update assessments almost immediately when new information arrives.

But the output still has to describe an event that hasn't happened yet.

Most fans never see the infrastructure behind those calculations.

They don't see every historical result being processed, the player data feeding a model or the adjustments made when a lineup changes. By the time the information reaches a sportsbook app, much of that complexity has been compressed into a handful of numbers.

Kansas City +120. The opponent -140. A total. A spread.

Modern sportsbook platforms build a much larger digital product around those numbers, combining live markets, real-time data, mobile interfaces and promotional offers. The experience also varies by jurisdiction, with availability and terms changing depending on where a user is located. Sports Illustrated's Canadian betting coverage illustrates this with its guide to the Stake promo code, which outlines the offer alongside the geographical restrictions that determine where the platform can be accessed. The odds may originate from increasingly sophisticated analytical systems, but how users encounter them is shaped just as much by the platform and market they are using.

That consumer-facing simplicity can hide just how much technology sits underneath the market.

A number beside a team name is the end of a process, not the process itself.

For the decade covered by SI, that process became faster and more data-rich. Teams gathered more information. Analysts gained better tools. Betting markets could react to news almost immediately.

Kansas City still went 20-9-1 in games where it started on the underdog side.

Boston still won 92 times as an NBA underdog. Miami did it 156 times. Pittsburgh managed 45 NFL victories from that position.

There is no contradiction there that needs solving. Better analytics were never going to remove the turnover that bounced the wrong way, the unexpected breakout performance or the shooter who suddenly couldn't miss.

What changed over the decade was the amount of information available before those things happened.

What didn't change was the game itself.