Simulation results through NCAA tournament round 2

On Saturday, I posted the round 1 performance of my NCAA tournament simulations, using data from Teamrankings.com. I did pretty well: 23/32 games correct, similar to some other prediction methods that I tested. Before round 3 kicks off, I wanted to go through my results through round 2. For comparison, in the first four rows of the table below, I took the better team in each game as indicated by the ranking listed on the left (i.e., the higher seed, the team with the better RPI, etc.). For the “Causal Sports Fan” row, I took the team that won the majority of my game simulations.

My picks to make the sweet sixteen did a little bit worse than my first round picks — only 9 out of 16 are still in. However, some sweet sixteen picks dropped out in the first round for each method (I lost 2 Missouri, 2 Duke, 4 Michigan, and 11 Texas). If I forgive first round mistakes and examine the actual round two match ups, all methods did quite well in the second round:

These results are similar to my simulation performance in previous years (over 70% correct in the first two rounds). I missed two upsets (13 Ohio over 12 South Florida and 11 N.C. State over 3 Georgetown) but got 6 Cincinnati over 3 Florida State right. I incorrectly predicted 5 New Mexico over 4 Louisville and 5 Vanderbilt over 4 Wisconsin, but got all the other favored winners correct.

I learned something from this last game: the 200 simulations that I ran per game are not enough for close match ups. Although Vanderbilt won 76% of the 200 simulations, 1,000 or 5,000 simulations flips the game to Wisconsin at about 55% odds. This game was a big outlier — more simulations don’t change any other second round picks — but I’ll stick with my predictions from before the round occurred.

The next rounds get tougher, because the games get closer. I still pick the team that wins the majority of my simulations in each game, but with estimated odds near 50% for many of the games, my confidence in these picks is not as strong. Here are my predictions for the South region (I list the odds of winning the previous game next to each team, based on 8,000 simulations per game for rounds yet to be played):

Duke’s exit gives Baylor an easier route to the elite eight (76.9% odds), though I had Baylor over Duke anyway. Kentucky is still the favorite to get through this region, but their toughest game is Indiana (only 55.3% odds). If Indiana can upset Kentucky, they also hold a slight edge over Baylor, as indicated in parentheses (55.3% odds).

Next, the East:

Florida State was a fairly weak three seed in my simulations; St. Bonaventure had decent odds against them (45.5%), and I had Cincinnati and Texas favored, though slightly. Florida State probably got such a high seed because they beat Duke and North Carolina en route to the ACC tournament championship but didn’t really deserve it. The beneficiary is Ohio State, who has 65.7% odds against 6 Cincinnati, but had similar odds against the 3, 11, and 14 seeds. They were a pretty safe bet to make the elite eight from the outset.

The top half of this bracket is messier because of the absence of Fab Melo for Syracuse. I run some modified simulations for the Syracuse – Wisconsin game in my round 2 submission for TeamRankings’s blogging competition. Syracuse only has 56.6% odds with Fab Melo; without him, they might not even  be favored. Even if they win, a game against Ohio State is a virtual coin flip (51.3%), before accounting for Melo. I’d say Ohio State is the right pick to escape the East from our current vantage point.

The West:

My champion in initial simulations was the 2 seed in the West, Missouri. They bowed out in the first round, of course, but the lesson from that pick is that Michigan State is not a strong 1 seed. Indeed, I give Florida 52% odds of beating Michigan State. Florida is also my only upset pick in the third round (apart from the caveats for Syracuse given above), at 56.3% odds over Marquette. Lucky for Michigan State, they face a weak 4 seed in Louisville, so they still probably have the best odds of making the final four for this region.

The Midwest:

Because of the upsets in the middle of this bracket, the 1 and 2 seeds (North Carolina and Kansas) are strong favorites to reach the elite eight. While some might argue that these upset teams have proven they are better than their regular season performance suggests, I’m inclined to go with the larger sample still and trust the stats. Although I slightly favor Kansas over North Carolina in the elite eight (55.3%), North Carolina’s odds of making the final four are probably a bit better, since they have 87.1% odds of winning in round 3.

The Final Four:

It’s hard to say too much here. Missouri was a pretty strong favorite over most of these teams (most odds over 57%, and many over 60%), but none of the remaining teams really distinguish themselves. Kentucky is my pick to win, but they have odds around 55% for every game starting with the third round. It’s not likely that they’ll win four in a row with that percentage. With 200 simulations, I had previously predicted that Kansas would top Syracuse, but 8,000 makes it clear that Syracuse is the favorite (55.3% odds). Again, I think Wisconsin and Ohio State have a great chance to beat them sans Melo, and Ohio State is also a slight favorite over Kansas (52.2%).

If you’re looking for crazy picks, 4 Indiana and 7 Florida both have decent chances to make the final (Indiana wins 50.9% over Florida, and Florida wins 43.7% over Kentucky), but neither show well against the likely finalists from the other half.

I’m excited to see what happens in the next two rounds. There are lots of 50-50 games in the simulations, so I don’t expect to get as many picks “right,” but I hope that my success percentages come close to my simulated win percentages.

Advertisements

7 responses to “Simulation results through NCAA tournament round 2

  1. Always nice to see someone post a successful formula. Too often the public immediately dismisses algorithms as nonsense. We’ve seen a lot of success with them.

  2. Jeff Sagarin and Ken Pomeroy are two highly respected guys in the field. Making a formula that competes with theirs is highly impressive.

    • Thanks for the praise, Ryan. Indeed, I was excited by the results. I’m not being totally fair to Pomeroy, since I think he publishes match up specific odds that incorporate how the teams fit together (I just use his main rankings). Still, this was my first go at this, and I think I can make some improvements still (better adjustments for strength of schedule, for example). Your RLDinvestments site got me in a nice little debate about Kelly betting with my buddy – good stuff. I’ll try to post my performance against better odds later.

  3. …your picks are math based and from a computer.but from a reality perspective-they are weak. This explains totally why your accuracy rates are so low. This is not your fault. It is merely a statistical expression of your formula. You will find greater success if you impart point values to human/group attributes in, say, ten performance areas of your choice. Factor in home court advantage, bench relativity, etc and then mesh these to integrate with your current approach. I feel you will then have a more comprehensive math-based tool. Thanks for your work.

  4. Thanks for reading, Greg. I completely agree that there’s more that can be added to these simulations (the Fab Melo situation for Syracuse is a perfect example, since my input stats don’t account for his absence). However, I think that my performance is pretty good for a first try. Over 70% for the first two rounds is as accurate as using the Vegas lines to pick the winner. Also, I’m not sure why home court advantage would matter in the tournament, where games are played on a neutral floor.

  5. How would say it changes the optimal bracket strategy knowing that you’re probably competing against a pool of noise traders who have some small random deviations from the higher-seeded bracket? Is there some threshold less than 50% at which it makes sense to pick a lower-seeded underdog by your method?

    In my pool with high school friends, I looked at Sagarin Predictor and picked OSU to win and Florida to make the Final Four, based on who was likely to be undervalued by the rest of the pool. I’m guessing I have the best chance to win right now. Almost no one has OSU, while a lot of the rest of the win probability will be split between Kentucky brackets.

  6. Your intuitions are absolutely right. In terms of filling out your own bracket, Florida and Ohio State are great options (and my predicted champion, Missouri), because they weren’t super popular. I talked about St. Mary’s in this light as well – although they lost in the first round, they had odds over 45% against seeds 1, 2, 3, and 4 in their bracket. In terms of a threshold, it will depend on your pool. For example, in a small pool, I’d say take the favorites the whole way. The bigger the pool (and the more you care about being 1st versus just doing well), the lower you should drop your thresholds.

    I wish I’d had time to sim all the possible (or most likely) combinations of games, so that I could determine probabilities of advancement at each stage for every team. This info would help you figure out where an underdog might have higher than expected odds.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s