Loyal readers will recall I’ve made a few detours to delve into sports where there was an interesting parallel to financial markets, particularly in the use of analytics. Today, I provide an update on those posts. (Hint: I’m going to brag some about my outstanding bracket!)
First, is there collusion in baseball? While I concluded there wasn’t, at least explicitly, a new academic paper cited by the Wall Street Journal suggests perhaps otherwise.
Then I’ll revisit my bracket predictions now that the tournament is over. Spoiler Alert: My methodology did exceptionally well!
There’s No Collusion In Baseball!
My original stance was baseball teams weren’t colluding in using models to value players since there was no explicit communication involved. However, the end result did resemble collusion.
New research suggests that some algorithms do indeed produce outcomes similar to actual collusion depending on how they are implemented.
The short story is “learning algorithms”, where the AI is given an outcome to target but no instructions for how to get there, are more likely to produce collusion than “adaptive algorithms” where the AI starts with a model upon which it tries to improve through iteration. If you are interested in the details, read the paper.
Given we live in a world where bias is seen everywhere, could algorithms also be prone to bias? Indeed, they can! Could this be enough to convince regulators that machines are, in fact, colluding with each other? Perhaps. If so, we might see public policy evolve to ban AI that produces tacit collusion.
If this were to occur, it would be a big victory for baseball players hitting free agency. Yet for those young players who, fearing a bad outcome in their future free agency, settled for a cheaper contract today, they would have a reason to start crying in baseball!
Bracket Challenge Thesis “Proven”
My advice for picking your brackets boiled down to three words: “Don’t Pick Duke”! Not because they weren’t a good team but because of the “lottery number effect” where you pick common numbers and have to split the pot if you win. And that advice looked prescient pretty quickly as they almost lost in the second round, again in the third, and finally fell in the Elite 8.
The rationale to pick against Duke was too much else had to go right for you to win your pool. With nearly 40% of entries taking Duke to win, you would need to do incredibly well on all the other games to beat all the other Duke entries.
However, if you picked Virginia or Gonzaga, which had similar odds in the computer models, you would be competing against a lot fewer people in the other rounds giving you more margin for error.
The other advice was not to pick too many favorites. Even though most years, only one or two of the 1 seeds make the Final Four, most entries will choose three or even four of the #1s. Even those who pick against the favorites, tend to go with #2 seeds or if you’re bold a #3.
However, history tells us that there is typically one team seeded 4 or lower in the Final Four as Auburn did this year.
Bracket Contest Results
Recall I created five different bracket strategies to test my ideas. The two that were the most interesting were the Analytics and Contrarian approaches.
The Analytics approach followed the advice not to pick too many favorites. It replicated the historical distribution of seeds throughout the tournament and did wonderfully for the first two weeks.
It got 7 of the Elite 8 correct (only missing Purdue) and after getting both Virginia and Texas Tech into the Final Four Saturday was in the 99.8% percentile nationally before Duke and Kentucky lost the next day.
However, because this entry went with Duke as champion it ultimately fell short of the Contrarian approach which had Virginia as champion instead of Duke.
The Contrarian entry lagged the Analytics entry by three wins in the first round, two more in the second, and one in the third and both had two Final Four teams. Yet, it pulled ahead by getting Virginia right in the end and ended up in the 98th percentile overall.
Note, how this proves the thesis! The Analytics entry was as good as one could have hoped through 58 of the 63 games. If Duke had won, it was probably good enough to beat out most or all of the other Duke entries. However, the Virginia entry was able to make mistakes early and still win because it had less competition.
It’s worth noting that in a year where the experts largely predicted favorites would do better than normal this year, the Consensus bracket finished last!
Finally, in a small redemption for the value of active managers, the Stock Picker bracket also had Virginia winning and finished 95th ptile nationally.