Aaron Brown's NFL Sports Betting System.

[AaCBrown]

Actually, I would like to give an example using stock earnings. Before the last earnings event the implied move was 10%. The event was a 25% move. This earnings the implied move is 20%. There was an increase of 10% which is = S + N. How do I measure S and N?

S + N = 20%, this is what you observe. You can't measure S and N, you can only guess about them, hopefully with better than random accuracy.

I this case you have conflicting information. The fact that this quarter's implied move is greater than last quarter's implied move inclines you to guess it is an over-reaction and the actual expected move is between 10% and 20%. But the fact that it is less than last quarter's actual move inclines you to bet that the actual move will be between 20% and 25%. Both of these predictions are things you might expect to be borne out when averaging over hundreds of events, not something reliable every time.

This is ultimately an empirical question to be investigated from past data. There could be reasons you have the opposite effect, many moves tend to be under-reactions. The only value of thinking about shrinkage is the knowledge that it works in many similar situations, so it's less likely to be a random pattern thrown up by data mining (that is, if it turns out to exist in the data) than some pattern without broader support.

 

[AaCBrown]

I am interested in your application of R^2 despite the autocorrelated nature of the data in both sports betting and finance. Linear regression provides an at-best poor predictor (theoretically) for data that is serially correlated (good teams tend to keep doing good, etc) and yet you seemed to have cracked the code without touching any of the "mathematically pure" techniques such as ARIMA. Would you mind giving me a taste of your use of R^2 with this data and how you were able to use it despite the math telling you autocorrelation is going to destroy your predictors?

I don't think in terms of R^2 for this application. I'm trying to get a 55% win rate, meaning I need a 0.01 R^2 between my prediction and the outcome. In most textbook problems, an R^2 of 0.01 would not be considered useful. But it can make you rich if it's real. High R^2's usually describe relations that are obvious without statistics--and often if you want to understand those relations you need domain knowledge rather than statistical expertise.

In an efficient sports betting market, there would be no autocorrelation. Sure, good teams keep winning, but if the market correctly evaluates how good teams are, the spread should be set such that good teams and bad teams beat it 50% of the time (in fact, underdogs win slightly but consistently more than 50% of games, but this autocorrelation is the sort that models exploit to win, not autocorrelation that destroys the models).

My model actually bets on negative autocorrelation in spreads. If the spread moves to favor Team A, it moves too far on average, and you should bet against Team A.

 

[TheBigShort]

So the more related your measurements are, the more benefit from shrinkage.

Aaron, can we go over this a bit more?In your example Ed George found out it was better to shrink to the average of each position. What if we find that your dominant hand explains some variance in batting average. How would I go about shrinking to both: position and dominant hand?

The answer to your earnings question is ultimately empirical.

How might you go about finding a predictor variable on a data set that has many levels. For example, in an Earnings data set where our DV is |Move| we have many levels/groups such as Stock, Industry, Sector, Mktcap etc.. We might find a relationship between |Move| and Mktcap only once we group by sector. The amount of combinations start to get very large. This is a serious problem I have when it comes to modelling.

Above I posted a link to cross validated where @whubber showed a negative relationship between 2 variable once we grouped and used the centroid of each group.

 

[AaCBrown]

Your first question is easy. You could build a model of the form:

second half of the season batting average =
A * (first half of the season batting average for the player) +
B * (first half of the season batting average for all players at the same position) +
C * (first half of the season batting average for all players of the same handedness) +
error

Normally in shrinkage, A + B + C = 1, but you do not have to enforce this. You also don't have to use a linear model, for example, you might want to shrink batters close to the average more or less than batters far from the average; or you might want to shrink batters with more at bats less than batters with more at bats.

I don't think it's wise to look for shrinkage everywhere, with respect to every variable and combination of variables. I would look for it only where (a) you had a lot of data from different markets and time periods suggesting there was some broad, fundamental effect, or (b) you had some economic or behavioral reason to expect over-reaction.

 

[TheBigShort]

Aaron, its great to see you have turned ET notifications on . It looks like people are really enjoying your posts. Keep them coming!

Another concept you mentioned in the paper is "Feeding The Hungry". Do you think this concept would apply to an earnings event? Eg. The long vol earnings bettors/hedgers have lost money over the last 4 LULU events, so they will most likely not be willing to bid up the option prices this quarter. I have actually tested this because of your paper, but I did not find a signal.

Unlike sports, The Hungry might remain hungry like in the index vol space (ivol usually overstates hvol). Which is unlikely to change. I would love to hear your opinion on that.

p.s. Where do you do your data analysis? I could not figure it out from your posts on NP and Qoura.

 

[TheBigShort]

Data Analysis Using Regression and Multilevel/Hierarchical Models

I am 100 pages in and I absolutely love it! The math is simple, the examples are fantastic. For anyone who is interested in learning about shrinkage/partial pooling, this is the best beginners guide!

 

[MrKobalt]

I would like to try it as well.

 

[PoopyDeek]

Who has Aaron Hernandez in their FF roster?

 

[newwurldmn]

Who has Aaron Hernandez in their FF roster?

duzos kid's friend.... for his NBA fantasy team.

 

[Freedance]

This information should be helpful to anyone placing bets. No matter how much I tried to figure out the odds, winnings, and other things, it was always tricky.