Data Dishin' with DJ Dimes: Gamblers Anonymous (Week of 12/28/15)
It's the [2nd] mosttttt wonderful timeeeeee, of the yearrrrrrrr... Although no dates on the calendar hold a candle to the opening weekend of March Madness, the beginning of conference action still brings an immense amount of joy to us here at 3MW. Coincidentally, this yearly milestone also marks the beginning of the New Year, which raises the almighty question I know we are all losing SOOOOO much sleep over: What is your 2016 New Year's resolution?
I speak for all of us here at 3MW when I urge you to dream bigger than the cookie-cutter choices of "losing 15 pounds" or "re-connecting with old friends" or "[insert cliche resolution that will be followed through for two weeks at the most]". This chapter of the Data Dishin' series is dedicated to the real "go-getters" out there, the true 1% of the 1%, who wish to do something meaningful in 2016: Yes, I am talking about the men and women all across this great nation who are focused on one thing, and one thing only: Winning Monies. It is you fine degenerates who are the true heroes, and the primary source of my inspiration...
Over the past week, I began documenting Against-the-Spread (ATS) results for all D1 mens college games . Along with final score results, I am also capturing the opening and closing lines, as well as the public consensus percentages (how many bets were placed on each team). Armed with this data, I will post weekly updates, which will call out any significant trends observed in the collegiate hoop betting market in the week prior (see "General Trends" below).
However, if you are a 3MW lifer (or groupie), you know the real motive behind this process is for tracking the week-to-week performance of Green Magic, which will turn two weeks old in a few days. To those of you who are now completely lost, Green Magic is a single game prediction model recently developed by some of the world's most esteemed mathematicians (me). It's purpose is to identifying teams with the most value when picking against-the-spread. If you're thirsty for the dirty details of how the model works, I'd highly recommend reading the prequel, "The Birth of Green Magic", which can be found on the DJ Dimes main page (http://three-man-weave.com/3mw/2015/12/14/dishin-data-with-dj-dimes).
Here is quick (and probably poorly explained) synopsis of what Green Magic does:
- For any two teams playing today (one home and one away), the model will search through 10 years of regular season game data and identify historical games which featured teams that are most similar to the two teams playing today. Based on the outcomes of the most similar historical matchups, the model determines what is most likely to happen in todays matchup. It will predict the final game score, which in turn reveals what the spread should be. At a high-level, its concept is similar to that of the TeamRankings model, which can currently be found on ESPN Insider game pages.
Before pulling back the curtain to reveal the wizardry genius of Green Magic, I wanted to first present some general or "macro" betting trends from the past week in college hoops.
General Betting Trends: 12/28/15 - 1/3/16
Below is a summary of the Against-the-Spread results for the 223 D1 games tracked over the past week:
- "Public" means a team received 60% or more of all bets (# of bets, not total $ of bets)
- "Non Public" means a team received 40% or less of all bets (# of bets, not total $ of bets)
- Games that ended as a push are not captured here
- Games in which lines ended in a pick'em (line equal to zero) are not captured here
Takeaway: "Home Public" plays (teams playing at home, with the majority of all bets on that team) appeared to be the only profitable betting trend that emerged in the final week of the holidays. Betting on all of these teams would've generated you a 4.1 unit return (36-29 record and assuming the juice is 10%). To put this in context for those of you who are not degenerate gamblers, if you bet $25 a game (in other words if 1 unit for you is equal to $25) on all 65 games, you would've won just over $100 (yay!).
These results actually contradicted my hypothesis that home teams don't have as much value over the winter break. My theory is since most students are generally gone for the holidays, the majority of the home crowd fans consist of older alumni, resulting in a much more bland environment (no offense to you "Baby Boomer" or "Gen X" readers out there). Initially, this theory looked to maybe have some clout, as road dogs began the week 6-0 on 12/28. Each of these 6 matchups featured a major conference team playing their first game back from Christmas break against a lesser non-conference opponent.
However, my initial excitement proved to be premature, as this was clearly too small of a sample size to react to. As the week progressed, and as conference play began, home teams protected their home floor relatively well, and quickly wiped away any rust that may have developed during their time off.
Beyond that, there really wasn't a major trend that emerged this past week. I will continue to track these results throughout the year, in order to test the legitimacy of some ancient and cliche gambling strategies, including: 1) Bet against the public & 2) Bet home dogs .... each of which have been considered to be profitable in the long run.
Now lets get to the juicy stuff ...
Green Magic Results: 12/28/15 - 1/3/16
Ta daaaa. We finally have preliminary results for the infant baby Green Magic (GM). The betting strategy is to play the side that GM indicates is undervalued (or fade the side that is overvalued). This is done by comparing the GM projected spread (dervied from similar historical games) with the actual Vegas spread. Below are the summary results of the win-loss record and unit returns for all Green Magic plays from the past week:
- Overall: Refers to the spread projection based on historical matchups of teams with similar Adjusted Offensive Efficiency & Adjusted Defensive Efficiency. This acts as a baseline spread in the system.
- Four vs. Overall: Refers to the spread projection based on historical matchups of teams with similar Four Factor characteristics (Effective FG%, Turnover Rate, Rebounding Rate, Free Throw Rate), in comparison to the "Overall" spread projection
- Style vs. Overall: Refers to the spread projection based on historical matchups of teams with similar style of play (tempo and defensive identity), in comparison to the "Overall" spread projection
- Combined vs. Overall: Refers to games in which the "Four vs. Overall" and "Style vs. Overall" projections pick the same side
Takeaway: We're all gonna be rich!!!!!
Actual Takeaway: It is still much too early to make any assertions about the success of the model.
Though you probably sensed the sarcasm associated with the 1st takeaway, this is still an encouraging start. I'm actually furious knowing that I would be up ~ 10 units (conservatively) had I bet on every team that Green Magic labeled as "undervalued". Given my unit is $50 (cause I'm rich), that means I would be up ~$500 dollars had I bet on all 200 games.
I don't want to bore the vast majority of those who have actually read this far, but I will show one game example and how the Green Magic output reveals which team is "undervalued". Here one row from my Excel summary sheet that is fed from the code script ...
I specifically chose the Northern Iowa @ Southern Illinois game because I naively bet on the road favorite Northern Iowa in this one (without firing up GM first). Had I consulted Green Magic before making that wager, he/she/it would've probably said something along these lines .... "Hey you idiot, the Salukis are undervalued! Don't be a dummy and bet on that purple and yellow team!" He/she/it was right ...
To explain the picture in more detail, the 4th and 5th columns, capture the opening and closing line for this game. The 0 and 2 mean the game opened at a pick'em, and closed with the home team Southern Illinois as 2 point dogs (+2). Apparently I wasn't the only person betting on Northern Iowa, as the line moved two whole points in favor of Southern Illinois.
For the baseline spread projection (indicated in the "Overall" column), the most similar 150 matchups, based on Adjusted Offensive Efficiency and Adjusted Defensive Efficiency alone, indicated the spread should be -0.26 (essentially zero or pick'em). Using these same 150 historical matchups and re-ordering them by how similar those teams Four Factors characteristics are to the Southern Illinois & Northern Iowa, the projected spread shows the home team should be actually be favored by 2.5 (-2.54). And when re-calibrating the same way with the Style characteristics, the projected spread also shows the home team should be favored (-2.43). Finally, since both Four & Style tests agree that the home team is undervalued, this would qualify as a "Combined" play, in favor of the home team Southern Illinois.
Sure enough, Green Magic was right, and Southern Illinois beat Northern Iowa by 2 on their home floor, costing me, as well as millions of other fools, a BOATLOAD of money (aka $50). This game is chalked up as a big fat W for all three tests shown above.
I'm sure a small number of you who read the first chapter in more detail ("The Birth of Green Magic") will be wondering where the "Other" similarity test is incorporated in the above results. The answer is that I made an executive decision to remove that third component of the model for the time being. I did some one-off analyses (basic regressions) to more closely measure the impact the "Other" factors (i.e Effective Team Experience & Effective Team Height) had on game predictions. Put simply, these more qualitative factors simply don't influence affect how good teams will perform in certain matchups as much as the "Four" factors and "Style" factors. As I begin to incorporate more team variables going forward (I allude to these examples in chapter 1), there will certainly be some qualitative factors that prove to be relevant in determining game outcomes. When I identify which factors those are, and quantify their impact, they will be merged with the "Style" similarity test OR given their own "Other" test, as I initially intended. For now, the near-term testing will focus on the "Four" test, "Style" test and "Combined" test (when the Four & Style projections lean the same way).
Hopefully this gives a clear enough understanding of how the model is operating in its current state. I am now beginning to use the model as a primary (and sometimes even exclusive) input to my college hoop wagers. But let's be honest... I'm fully expect my irrational confidence to jinx this early success, so stay tuned for what may be a much more depressing update next week ...