Bracket Science: The anatomy of an upset

The past three tourneys have been remarkable in the 25-year history of the 64-team era for their extraordinary lack of madness. No other three-year period has seen so devoid of upsets. The 2007 dance featured only three, the fewest since 1985. And 2008 marked the only tournament since 1985 to have all top seeds reach the Final Four. Then last year, long shots sprung just five surprises -- second only to 2007 for the fewest upsets.

What do three consecutive unpredictably predictable tourneys portend for 2010? Last year, after the sanity of the previous two years, I boldly proclaimed that 2009 would be crazy. But it didn't happen. So if I say we're in for a tourney of double-digit upsets this year, don't take it to the bank.

Still, the law of averages is the law of averages. There couldn't possibly be four consecutive calm, chalky dances -- could there? One thing's for sure: There is no prohibitive favorite like North Carolina this year. All the country's top teams have shown signs of vulnerability.

So while 2010 might not reach the sort of insanity that reigned in 13-upset years like 1990 and 2002, it will almost certainly be wilder than last year. It's a good bet that your bracket pool will be won by the person who can identify the right favorites to fall and long shots to advance. But Cinderella spotting is tricky business. Settle on the wrong high-seeded victim, and your bracket could collapse in the first weekend.

Good thing that picking the right underdogs and paper tigers isn't all guesswork. The upset victors and victims of the modern era have shared common attributes. When you know what they are, it's easier to sniff out the upsets. Let's examine the factors that correlate with upsets. Right after Selection Sunday, I'll provide an update that identifies the 2010 dark horses and vulnerable powerhouses.

When is a win an upset?

Not every game in which a lower-seeded team knocks off a higher seed is an upset. Nobody's going to fit a glass slipper on a nine seed that beats an eight in Round 1. (Actually, nine seeds are 54-46 against their higher-seeded opponents.) It's only when you get a gap of at least four seed positions between opponents that a game has upset potential.

Surprisingly, two-thirds of the 64-team tourney games have met this condition. Of the 1,575 games that have been played in the past 25 years, 1,043 of them have pitted long shots against favorites -- and the underdog has won about 20 percent of the time. That's an average of 8.3 upsets per tourney, or roughly one in every 7.5 games. (Now you know why the 16 upsets over the past three tourneys is so surprising -- it's about three fewer than usual per dance.) This chart shows the round-by-round results of upset games in the 64-team era.

Because of the way the brackets are set up, most of the upset matchups occur in the first three rounds of the tourney. Of the 1,400 games played in the first three rounds, 1,002, or 72 percent, have been upset matchups. Picking these upsets correctly -- or at least minimizing the number of victims you advance -- is essential to building a winning bracket.

While first-round upsets grab most of the attention, the upsets in the second and third rounds do the most damage to your bracket. Only 19 of the 105 opening-round upset victims are teams seeded one, two or three. More important, the top three seeds are a dominating 281-19 (.937) against first-round dark horses. Only about one in 16 top-three seeded teams will fall victim to an upset in Round 1. To put it another way, less than one top-three seed will lose in the first round per tourney. The last victim was third-seeded Iowa, which was humbled by 14 seed Northwestern State in 2006.

It's a different story in the second round, where the top three seeds are just 172-52 against Cinderellas. That's a solid .768 winning percentage, but it's nowhere near the lockdown .937 rate of the opening round. Instead of only one in 16 teams losing, the top three seeds lose one in every four games -- and about two per year. They also comprise 78 percent (52 of 67) of the second-round upsets. Last year, none of the top three seeds got upset. But in 2008, there were two high-seeded upset victims: second-seeded Duke lost to seven seed West Virginia, and two seed Georgetown lost to 10th-seeded Davidson.

The Sweet 16 isn't as treacherous for the top three seeds as Round 2, but it isn't a first-round cakewalk either. Of the 27 third-round upsets of the modern era, 17 (or 63 percent) have involved teams seeded one, two or three. But they do have a .833 winning rate against long shots (85-17). In fact, upsets among the top three seeds are so rare that less than one occurs per tourney. None of the higher seeds got upset in the Sweet 16 last year. In 2008, though, Davidson was able to spring a surprise against second-seeded Georgetown.

Upset prediction value: Balancing forecasting accuracy with frequency

Before we take a deeper dive into the anatomy of upsets, you need to ask yourself: What kind of an upset prognosticator do you want to be -- accurate or prolific? You could create an elaborate rule that has never failed to predict an upset. But the rule would almost certainly apply to just a handful of games. By the same token, you could pick every single upset -- if you're willing to be wrong 80 percent of the time -- and end up dead last in your pool.

In determining which factors have the biggest influence on upset prediction, you have to balance how much the rule increases the odds of picking an upset (accuracy) with how many upsets it describes (frequency). Here's an example: Long shots have won 17.5 percent of the 600 games in which an upset could happen in Round 1. These giant killers are rarely 15 or 16 seeds. By eliminating these teams from consideration, you improve your chances of picking an upset by 44.3 percent -- a 25.2 percent winning rate (101-299) vs. 17.5 percent. Just as important, 101 of 105 first-round Cinderellas satisfy this rule. By multiplying the increase (44.3 percent) above the typical upset rate by the percentage of upsets described (96.2 percent) we arrive at the upset prediction value (UPV), for the "15- and 16-seed exclusion" rule: 42.6. When you evaluate rules by their UPV, you're better able to compare their relative worth in predicting upsets.

Seven basic guidelines for Cinderella spotting

If the deadline for finishing your bracket is approaching, you may not have time to dig into the nitty-gritty of upset probabilities. Still, it's worth knowing a few general guidelines to boost your odds of picking an upset. Remember these seven rules, and you'll dramatically improve your odds of spotting an upset.

1. Don't pick any long shots lower than a 13 seed. Sure, the bottom dwelling seeds do spring an occasional upset, but it's at a much lower rate than higher-seeded teams. These long shots are 21-297 (.066) against opponents with a seed position at least four rungs higher than them -- hardly worth risking your bracket on.

2. Never pick a top seed to be an upset victim in the first three rounds. You don't have to look any further back than Duke's loss to fifth-seeded Michigan State in 2005 to know that top seeds can be toppled before the Elite Eight. But it's not worth predicting. In the 252 games they've played against prospective Cinderellas, one seeds have shattered the slipper 235 times (17-235). That's a 93.3 percent success rate -- much too strong to bet against.

3. In the opening round, pick solid-scoring veteran Cinderellas that aren't struggling heading into the dance and don't rely too much on their frontline. OK, so maybe this isn't such a simple guideline ... but it could help point you to the best 2010 underdog candidates. Restrict your choice of first-round Cinderellas to teams that:

• Score more than 66 and beat teams by more than five points per game.
• Have no worse than a two-game losing streak heading into the dance.
• Get at least 25 percent of their points from guards.
• Have a winning percentage better than .600.
• Have a starting unit older, by average, than all sophomores -- but not with all seniors.

Teams with these characteristics are 72-126 in Round 1, for a 36.4 percent upset rate. That's more than twice as frequent as the typical 17.5 percent first-round upset rate.

4. In the second round, focus on solid scoring dark horses that aren't too hot or too old. The most reliable Cinderella teams in Round 2 score more than 65 and win by more than 3.5 points per game, haven't won all of their final 10 pre-tourney games and don't have more than three senior starters. Underdogs with these qualities are 44-61, for a 41.9 percent upset rate.

5. In the Sweet 16, avoid one and two seed victims -- and only pick senior-led, balanced-scoring long shots. You're already steering clear of top seeds. And after the first two rounds, it no longer pays to pick two seeds as victims. They're a dominant 25-6 against underdog six and 11 seeds in Round 3. By ignoring upset matchups involving the top two seeds and restricting your long shot candidates to teams seeded 12 or higher, you'd be 16-21 in picking third-round upsets. If you further restricted your upset choices to teams that have won at least five of their past 10 pre-tourney games and got between 35 and 80 percent of their points from guards, then you'd improve your prediction success to 16-12 (.571).

6. In the Elite Eight, narrow in on offensive-minded Big Six conference squads with experienced coaches and respectable records. Long shots from the six power conferences that score more than 70 points per game, have a winning record greater than .650 and a coach who has been to the dance before are 4-3. All the other would-be Cinderellas are 1-21. (George Mason is the only exception.)

7. For the Final Four and championship, pick underdogs with veteran tourney coaches. In the 12 matchups of the past two rounds, long shots with coaches that have gone to the dance at least seven times are 4-1. The rest are 0-7. Those four wins come courtesy of Rollie Massamino's 1985 Villanova Wildcats and Larry Brown's 1988 Jayhawks. Tom Izzo was the veteran coach who failed to spring an upset.

So how strong of an upset predictor would you be if you observed these guidelines? How does a 41 percent accuracy rate grab you? In upset matchups meeting all the above conditions, the underdog is a reasonable 140-205 (.406) against the favored seed. That's 103.5 percent better than the typical 19.9 percent upset rate and it describes more than two-thirds of the upsets (140 of 208). That translates into a strong UPV of 69.7.

Upset rules for matchups in the first two rounds

If you're willing to examine potential upset squads more carefully, you can boost your upset-spotting success rate to the point where you're right more often than you're wrong. Here's what you need to do in the first two rounds:

Round 1: Take high-scoring 11 seeds, frontcourt-dominant 12 seeds with tourney-tested coaches and 13 seeds with experienced coaches that soundly beat opponents. Even if you're going to examine the opening round more closely, it doesn't pay to pick a Cinderella lower than a 13 seed. That restricts your choices to 6 vs. 11, 5 vs. 12 and 4 vs. 13 matchups. In those games, look more favorably on underdogs with these attributes:

• Take 11 seeds that score at least 73 points a game and beat their opponents by seven or more points. These teams are 22-19.
• Take 12 seeds that have been to the previous dance and get more than 49 percent of their points from the frontcourt. These 12 seeds are 15-7.
• Take 13 seeds with experienced tourney coaches who aren't snakebitten -- with five or more trips and no Elite Eight appearances. Make sure their scoring margin is more than 6 percent better than opponents and that they get between 33 and 67 percent of their points from guards. Finally, avoid streaking 13 seeds that have won more than eight of their past 10 games. Teams fulfilling these conditions are 11-14.

By applying these three rules, you'd be 48-40, or 54.5 percent accurate, in predicting first-round upsets. That's 211.7 percent better than the typical first-round upset rate of 17.5 percent, and it describes 45.7 percent (48 of 105) of the upsets -- for a hefty UPV of 96.8.

Round 2: Pick experienced or battled-tested seven through 12 seeds -- but avoid ninth-seeded underdogs. With all apologies to ninth-seeded UAB, which upset top-seeded Kentucky in 2004, there's no reason to pick a nine seed Cinderella in Round 2. They're an abysmal 3-51 against the big dogs. And 13th- through 15th-seeded teams should be ignored as well, since they're only 6-34. That leaves seven, eight, 10, 11 and 12 seeds. Here are the upset rules for these seeds in their matchups:

• In 1 vs. 8 games, take battle-scarred eight seeds with tourney experience. Eighth-seeded squads that: 1) went to the tourney the previous year, 2) have a coach whose made at least three trips to the dance, 3) have less than a .750 record, 4) beat teams by eight or fewer points a game and 5) won at least five of their past 10 pre-tourney games are 7-7. Every other eight seed is 2-30.
• In 2 vs. 7 games, take seven seeds from the Big Six conferences that beat opponents by more than six points a game and have won at least four of their past 10 pre-tourney games. These seven seeds are 9-6. All other seven seeds are 8-37.
• In 2 vs. 10 games, take 10 seeds that average a starting unit of at least juniors (freshman=1, sophomore=2, junior=3, senior=4). These older teams are 12-9 against their favored opponents, while all other 10th-seeded teams are 3-12.
• In 3 vs. 11 games, take 11 seeds that are from Big Six or mid-major conferences, have a tourney-tested coach, have won fewer than nine of their past 10 pre-tourney games and get between 40 and 70 percent of their points from guards. Teams satisfying these conditions are 6-3. All other 11th-seeded squads are 2-17.
• In 4 vs. 12 games, take Big Six or mid-major 12 seeds coming into the tourney winning no more than three pre-tourney games in a row that beat opponents by fewer than 10 points a game. They're 7-4 against four seeds. The rest of the 12 seeds are 3-12.

If you applied all these rules, then you'd accurately predict 41 upsets in 70 games for a 58.6 percent success rate. That's 143.0 percent better than the typical 24.0 percent second round upset rate, and it describes 61.2 percent (41 of 67) of the upsets. That works out to a UPV of 88.1

Putting all the upset rules together

If you followed all of the rules listed above for each round of the tourney, then you would've picked upsets correctly in 113 of 198 games, for a 57.1 percent success rate. That's a 186.2 percent improvement over the typical upset rate of 19.9 percent. And it describes 54.2 percent of the 208 upsets that have occurred in the 64-team era for an impressive 101.2 UPV -- over 45 percent better than the UPV achieved from the seven basic guidelines. Let's face it: If you could identify 113 Cinderellas (almost five per tourney) with nearly 60 percent accuracy, you'd be the envy of your fellow bracketeers, whether you won your pool or not.

The value of identifying high-seeded upset victims

While it might be a nice ego boost to be an expert Cinderella spotter, it's probably not going to help you build a winning bracket. If your main goal is to climb to the top of your pool, it's more important to know the upset victims to avoid than the perpetrators to advance. Just as Cinderellas possess common traits, there are definite qualities that first- through sixth-seeded victims possess. Steer clear of the following high-seeded squads and you'll stand a better chance of nailing the Final Four:

Top-seeded underachievers: 56 of 100 one seeds have failed to win four games and reach the Final Four. What separates these underachieving top seeds from their more reliable counterparts? If a one seed possesses any of the following attributes, consider penciling them in for a premature tourney exit:

• An average scoring margin less than 10.5 points per game.
• Fewer than 72 scored points per game.
• A coach with more than four tourney appearances but no trips to the Elite Eight.
• Didn't go to the previous year's dance.
• Gets 73 percent or more of its points from the frontcourt or backcourt.
• Gets outrebounded on average.
• Shoots worse than .465 from the field.
• Has a strength of schedule lower ranked than 80.

The 30 top seeds that possess any of these qualities have overachieved just twice. The 1997 Minnesota squad beat expectations despite not having gone to the tourney in 1996. And Texas overachieved in 2003, even though Rick Barnes had logged 11 tourney trips without going to the Elite Eight.

Two seed-defying runs in 30 tries works out to an abysmal 6.7 percent seed overachievement rate (SOAR). Contrast that with the top seeds that don't possess any of the eight factors above. Of the 70 squads without these disqualifiers, 42 have advanced to the Final Four. That works out to an overachievement rate of 60 percent, nearly 10 times better than the rate of the pretenders.

Second-seeded underachievers: 54 of 100 two seeds have failed to reach the Elite Eight. How can you spot the paper tigers? Avoid two seeds that:

• Have won less than seven of their past 10 pre-tourney games.
• Have lost two or more pre-tourney games in a row.
• Allow no more than 56.8 points a game, suggesting an grind-it-out playing style.
• Have an average scoring margin fewer than eight or more than 19 points a game.
• Shoot less than .450 from the field.
• Have a combined rebounding and turnover margin of less than 4.0.
• Get 10 percent or fewer of their points from guards -- or more than 76 percent.
• Rely on their starters for more than 90 percent of their scoring.

Two seeds that possess any of these qualities have fallen short of seed expectations of 2.43 wins per dance 27 times in 30 tries. That's a 10.0 percent SOAR. Two seeds that don't possess any of these qualifications are 43 for 70 in overperforming -- a 61.4 percent seed overachievement rate. The three teams that beat expectations while saddled with one of these attributes were: 1995 North Carolina (six wins in their past 10 games), 1997 UCLA (scoring margin less than 5.6 per game), and 2009 Oklahoma (six wins in their past 10 games).

Third-seeded underachievers: Forty-eight of the 100 three seeds never see the second weekend of the tourney. Here's how to spot a third-seeded squad that won't reach the Sweet 16; just avoid any team that:

• Has lost five or more of its last 10 tourney games.
• Has scored 67 or fewer points per game.
• Has a margin of 4.2 points per game or worse.
• Has a coach making his first March Madness appearance.
• Didn't go to the previous year's tourney.
• Has a road winning percentage of .500 or worse.
• Gets outrebounded on average.
• Has a turnover margin of 1.0 or worse.
• Shoots less than .440 from the field.
• Has the 25th or worse RPI among the tourney field and a strength of schedule ranked worse than 30.
• Gets 18 percent or less -- or 85 percent or more -- of its points from the backcourt.
• Has five senior starters.

That's a lot of disqualifiers, but they clearly separate the third-seeded contenders from pretenders. Only three of 19 teams with any of these qualities have overachieved -- a 15.8 percent rate. (The exceptions are North Carolina in 1986, Massachusetts in 1992 and Mississippi in 2001.) Contrast this with the 49 of 81 three seeds without any of these conditions that overachieved. That works out to a 60.5 percent overachievement rate, nearly four times better.

Four seeds that lost to 13 seeds in Round 1: Four seeds hold a commanding 79-21 advantage over 13 seeds in their opening-round match-up. How do you steer clear of the four seeds most likely to get upset? Avoid teams that:

• Have a winning percentage less than .680.
• Get less than 20 percent of their points from guards.
• Have an All-American.
• Have a coach who's been to the tourney but is leading a team that didn't go the previous year.
• Have an average starting unit class age less than 2.4, where freshmen=1, sophomores=2, juniors=3 and seniors=4.

Four seeds with any of these qualities are just 30-18, for a .625 success rate. All other four seeds are 49-3, for a winning percentage of .942.

Five seeds that lost to 12 seeds in Round 1: Five seeds have won nearly two-thirds of their first-round matchups with 12 seeds. But five seeds with a certain makeup lose more often than they win. Beware of favorites with any of these five qualities:

• They didn't go to the previous tourney.
• They've lost two or more pre-tourney games in a row.
• They've won less than five of their past 10 games.
• Their strength of schedule is ranked higher than 90.
• They score fewer than 66 points per game.

Five seeds possessing any of these attributes are just 20-21 (.488). Meanwhile, all other five seeds are a safe 46-13 (.780).

Six seeds that lost to 11 seeds in Round 1: Six seeds actually have an easier time against 11 seeds than five seeds do against 12 seeds. They're a solid 69-31 against their dark horses long shots. What separates the six seeds that get upset from the ones that get past the first round? Avoid any team that:

• Has a winning percentage less than .600.
• Has scoring margin less than 4.8 points per game.
• Has strength of schedule ranked higher than 100.
• Gets fewer than 22 percent of its points from the backcourt.
• Has an average starting unit younger than all sophomores, where freshmen=1, sophomores=2, juniors=3 and seniors=4.

Six seeds plagued with any of these characteristics are only 12-13 (.480); the rest of the six seeds are 57-18 (.760).

Peter Tiernan has been using stats to analyze March Madness for 20 years. His insights into the NCAA basketball tournament can help you build a better bracket. E-mail him at bracketscience@comcast.net or visit bracketscience.com .