Plus EV Analytics

+EV in the SuperContest

Advertisements

“It’s like the nature channel. You don’t see piranhas eating each other, do you?”

– Mike McD, Rounders

Last year, I was a part owner of an entry that placed 79th in the Westgate NFL SuperContest. After paying 30% withholding tax for non-US citizens on the GROSS cash amount (WTF, America?), I barely broke even. Yes I know there are ways to get the tax back, don’t DM me about it.

I’ve been trying to convince some friends to partner with me this season (US citizens this time, can’t get fooled again). “Yes it’s a high-variance luckfest, but it can be huge +EV with proper strategy”, I say. “You’re full of crap, prove it”, they say. Challenge accepted…let’s get started.

The contest rules are simple. Each week of the NFL season, each entry picks 5 games against the spread. 1 point for a win, 0.5 points for a push. Top 100 scores at the end of the season get paid on an escalating payout scale.

There are two elements to be considered in a good SuperContest strategy:

  1. Contest point spreads are published on Wednesdays. Entries aren’t due until Friday evening – and it’s being pushed to Saturday evening in 2020. Often, the live lines change between Wednesday to Friday and sometimes significantly if there’s an injury or other late breaking news. By picking games where there has been favourable movement between the contest spread and the current spread, one can capture closing line value which is highly predictive of win probability. If you can make picks that have 52-53% win probability when everyone else is taking 50% coin flips, that adds up to a large advantage over the 85 picks that make up a SuperContest season.
  2. To cash in the SuperContest, you need to finish in the top 100 out of 3,000+ entries. To win serious money, you need to finish in the top 10. Even with the best strategy, you don’t accomplish that without a LOT of good luck. The top-heaviness of the prize structure means that many strategies that would work if you’re betting individual games are not appropriate here. If the 100th place finisher ends with 50 points, a score of 49.5 is as good as a score of zero. To really climb that leaderboard, you need to differentiate yourself from the pack. That means going contrarian and making picks that you believe will be shared by as few other entries as possible. With a top-heavy structure, you’d much prefer a 50% chance of gaining a point on the pack and a 50% chance of losing a point on the pack, rather than being a part of the pack.

Notice that I didn’t include “NFL handicapping skill”. Do you know what skilled handicappers do? They bet. And their bets influence the live lines. So when you look at the difference between the contest spread and the current spread, you are in fact using the collective wisdom of every handicapper in the world. Wait, so we’re going to find +EV in an NFL handicapping contest without handicapping a single game? Yes we are.

I have a data set from the 2018 NFL season, containing:

The first task is to calculate each team’s win probability against the SuperContest spread for each game. Doing this during the season, there are enough alternate point spreads available that no conversions needs to be done…but in this data set, all I have is the Friday point spread, so I used this simple method: I looked at the historical frequency of a game landing on each number. For example, 2.6% of NFL games land on 4. So if the SuperContest spread on a game is 3.5 and the live line is 4.5, there’s a 1.3% probability that the -3.5 wins and the -4.5 loses. Assuming the 4.5 line is efficient and has 50% win probability on each side, the -3.5 would have 51.3% win probability. I’m making a pretty terrible assumption here that the 2.6% of 4s are divided equally into 1.3% “favourite wins by 4” and 1.3% “dog wins by 4”. In doing so, I’m actually underestimating the closing line value – so any EV estimates that come out of this method will be too low. For the sake of simplicity, we chalk it up to “conservatism” and move on.

Next, we need to quantify the “contrarian” effect. We can see every entry’s picks after the fact, but that doesn’t help us in real time when we’re making our picks. So, we need to build a model to forecast the crowd’s behaviour as accurately as possible.

Let’s define a pick’s “selection ratio” as the ratio of the number of times the pick was made to the average, where the average is total picks that week / (2*total games that week). More popular picks have a selection ratio higher than 1, etc.

We’re going to model the selection ratio using another one of those tricks I picked up in actuarial school, the Generalized Linear Model – basically a type of regression that works in places where you can’t use a normal regression. I’m going to predict the selection ratio as a function of four data elements:

1. Closing line Value. The idea that picks with favourable line movement are more likely to win is not earth-shattering, and it’s to be expected that at least some of the contest players are using that to at least partially inform their decisions.

Pretty strong relationship here.

2. The spread itself. Do contest players prefer favourites? Dogs? Close spreads? Mismatches?

Two clear patterns here – favourites are overweight compared to dogs, and tight spreads are overweight compared to mismatches.

3. Team bias.

With a couple of weird exceptions like Minnesota and Dallas, there’s a pretty clear picture of where the “public teams” are. This one may change quite a bit from year to year as teams get more and less popular with the public.

4. Primetime games

Thursday makes sense because if you pick the Thursday game, you have to have all 5 of your picks for the week in by Thursday kickoff – that’s a big disadvantage. I’m guessing Sunday night is more popular because people want a rooting interest and Monday night is less popular because contest players want “closure” on their week after Sunday, or because they don’t like the uncertainty from the extra day of information that could change.

We can combine these four elements in our GLM to obtain a decent estimate of where the distribution of picks will be for any given week.

So now we can build our contest strategy. We can pick the top 5 closing line value picks each week, or the top 5 most unpopular picks each week, or some combination of both strategies. For each possible strategy, we can determine the EV by simulating the 2018 NFL season 100,000 times using the actual set of picks that each contest entry made and the outcome probabilities for each game based on the Friday lines (I’d use the Sunday closers if I had them, it doesn’t make much difference because the contest participants wouldn’t have had them either). For each simulation, I can add up my score and see if I would have cashed vs everyone else’s scores, allocating myself an appropriate portion of the prize pool which is made up of $1,380 per entry x 2650 entries. There were actually 3126 entries in the 2018 contest, but I’m only including the 2650 of them who submitted all 85 picks. The EV for each strategy is the average winnings over the 100,000 simulations, divided by the $1500 entry fee.

Because the contest only puts $1380 of the $1500 entry fee into the prize pool, the “rake” is 8% and if this contest was pure luck each participant could expect to realize an expected return of $1380, a net result of -$120 and an EV of -8%.

To test how much of a skill contest this is, we can test a strategy that picks games at random to see how much worse it would do than the -8% benchmark:

A strategy that picks games at random returns an average of $1485, for an EV of -1%.

Wait, WTF? If you can outperform the average by picking games at random, that would mean that the average skill level of the SuperContest players is…negative???

That can’t be right, must be a bug in my code. Before I scrap it and start over, let’s do one more test. What if I followed the crowd using a strategy that picks the 5 games each week with the highest projected pick percentage?

A strategy that’s 100% based on popularity returns an average of $1271, for an EV of -15%.

It’s not a bug. Handicappers are likely to make the same picks as other handicappers. Maybe they’re following the same stupid narratives, or listening to the same stupid touts on TV. And in doing so, they’re sinking each other and themselves. The piranhas are eating each other. Handicappers are dead money in a handicapping contest…the irony is delicious. If the contest had no rake, like this one from the good people at Circa, it would be +EV to pick all games completely randomly.

OK enough screwing around, on to the results of our strategies over 100,000 simulated seasons. Because I gave myself the benefit of knowing the Friday odds, I did not allow myself to pick any Thursday games in this test. Here it is:

A strategy that’s 100% based on closing line value returns an average of $3718, for an EV of +148%.

A strategy that’s 100% based on unpopularity returns an average of $1787, for an EV of +19%.

And the piece de resistance:

A mixed strategy that’s X% based on closing line value and 100-X% based on unpopularity returns an average of $4566, for an EV of +204%.

No I’m not telling you what X is, because that’s MY strategy and you can’t have it.

With my strategy:

If I had unlimited free time, I could build a dynamic strategy where X changes from week to week depending on your standing. If you’re near the top of the leaderboard you should be more inclined to go with the value picks and grab as many wins as you can. If you’re near the bottom, you should go contrarian to make up ground. This dynamic strategy would do even better than +204%.

Make no mistake, the contest is still extremely high variance – with the best possible strategy you still lose $1500 11/12 of the time. Also you have to tie up your money for months, it’s a lot of work to figure out your entries each week and it’s not really scalable so you’re likely not going to be able to get rich off this. But if you’re looking for +EV and a fun way to stay engaged this NFL season, you’ve got it.

Good luck!

Advertisements

Advertisements