Betting Strategy
Setting the Target
If you can’t quantify your edge or prove that you have some sort of advantage then no amount of betting strategy is going to make you money in the long term. So the challenge is figuring out whether or not you can get picks right consistently enough to beat the bookmaker.
On this site we focus mainly on making bets against the spread on a week by week basis. On average, these bets can be found at your local casino or bookmaker at a line of -110. This means that if we are betting $100 on the game, we will have the chance to win $90.91. Working out the math as shown below, we need to be able to achieve a betting success rate of 52.4% or better to break even.
Determining Your Edge
Now that we know our target, how do we know if we can beat that on a consistent basis? The only true way to do that is to rack up experience over time and keep detailed win/loss records of all bets made. The problem with that is you will be devoting capital now not knowing whether or not you have a true edge. So the best alternative then becomes finding a way to use history to determine how you would have done in the past. This can be tricky and must be done diligently so as to not skew things in your favor. That is the main downfall of back testing - using the data that you already know to build a model that produces the results you would like to see. You must avoid the trap of modeling to the data. The only honest and fair way to do back testing is to build a model first based on what you feel is important without looking at the data. Once the model is complete, then historical data can be used to see if the model would have worked in the past or not. Obviously when doing this, only data that would have been available during that time and place of the pick can be utilized. It is critical not to let any outcome bias creep into your back testing!
Building the Model
Thankfully we live in the internet age and all kinds of information is right at our fingertips. Our goal was to build a model using only easily available data from reliable sources we found online. A big thank you to the following sites for their diligence in providing accurate and easy to access information that we utilized in building our model.
www.pro-football-reference.com
Any stat you can think of, they have it. It’s the ultimate database for NFL teams, games, and players.
Great site for advanced statistics like the famous DVOA (Defense-Adjusted Value Over Average) taking into account quality of opponents, game situations, and various other important factors that go into the raw numbers of any football game box score. Please check out their site for thorough explanations of their methods and great insight during the season.
http://www.aussportsbetting.com/data/
The final piece of the puzzle for us was finding a way to get reliable betting line information from past seasons. The best resource we located was the Australian betting site above who provides an excellent, easy to access and download, excel file with all the opening lines, closing lines, and final game scores dating back to 2006. These guys are champs for compiling this so please have a look at the link above if you would like to check out the data for yourself.
So with the information in place, the real fun begins. The main approach for us was to figure what we felt the most important aspects of the game were and start to try and quantify them. At the end of the day, we wanted to have a score or rating for each team and represent that in point values. This way it would be easy to compare any two teams in the NFL and predict what we thought the point spread for a game between them should be. As an example, one important aspect of any team is obviously the quarterback. So let’s say Joe Flacco (incredibly average) is going up against Aaron Rodgers, the Packers and Aaron Rodgers would have a several point advantage at QB over the Flacco led team. These types of comparisons are done for each of the critical aspects of the game and when you add it all up you have a predicted point spread for that match up.
I would love to go more into the details of how we built the model, what weightings were given to the different aspects of the game, and how we calculated the final point values but for now that is something we are choosing to keep proprietary. Once we figure out more about the direction of this site and how it will work best, it could definitely be something we would love to share in the future.
So once you have the comparisons complete and predicted point spread for a given game calculated, that should tell you what team the model suggests you should bet on. So the Bears and Packers are playing on opening night of the 2019 season the point spread is hovering around the Bears giving 4 points at home (Bears -4). If the model suggests that the point differential between the teams based on the comparisons we ran shows that the Bears are a better team by 6 points, then the model is essentially suggesting that a pick for the Bears -4 is the smart play. We are gaining two points of value based off of the model by betting them -4 when we think the link should be closer to -6. That’s how all of the picks are made in the next section when we give some details on the back testing results.
Testing the Model
So the model is built -let’s see if it works! We have the information we need going back to 2006 but it turns out back testing takes time, and we still have our day jobs. Organizing the data properly and making sure that only data available at that particular point in time is utilized can be a bit time consuming. That means we haven’t had the chance just yet to go through and back test all the way to 2006. For now, we have run the back test for 5 seasons starting in 2014 and running through the 2018 season. This gives us a decent starting point and enough games where we can start to see some semblance of statistical significance. As this season progresses, we plan to back test a few more seasons and post the results as we go along. Having more and more information to base our bets on will only help in the long run and further cement whether or not we have an edge. One final clarification before we take a look at the results, the closing lines from the Australian sports betting site above are the ones we utilized for all back testing. This should be the most fair approach to ensure we are not testing against any lines that wouldn’t have had all of the most up to date game time information incorporated into them.
All right so the graph above and the table below show a break down of the results of 5 years of back testing. There is good news and bad news - let’s take the bad news first. If we look at the overall record of the model when making picks on all of the games, it does select slightly more winners than losers but not to the extent that we can make money off of those picks. It is not above the 52.4% rate of winning threshold calculated above as it comes in at 51.8% with a 643-599-38 record over those 5 years. That means if we just make bets on every game each week based on the model, we are going to lose money over the long term. Not what we are looking for.
Time for the good news. So as you can see, we decided to segment out the results based on how big the points differential was between the model and the actual closing line
Important Note: the vertical axis above and left column of the table represent the difference between the closing line point spread and the point spread predicted by our model. This is not representing the actual line of the game - only the difference in the actual line and our predicted line. For our example above about the Bears and Packers, the point spread differential is 2 and that would be captured accordingly in the table/graph above. It is irrelevant what the actual line of the game is (in this case Bears -4) when we are talking about the above graph/table).
For the games that the model is more aligned with the actual closing line, our picks are basically coin flips. This is a definite losing proposition with the advantage (hold) a bookmaker typically keeps. But there is hope! When our model predicts that the point spread should be further and further away from the actual closing point spread of the game, we start to get some very positive results. The difference is seen best in the horizontal bar graph. The segments where the winning percentage is above 52.4% are highlighted in green while the segments below are in red. You will notice an increasing trend in winning percentage as the point spread differential rises from o to 10. So maybe we can’t bet on all the games and expect to make money, but maybe we still can bet on some of the games for a profitable outcome. So based on the graphs above, things start to turn in our favor when the model is greater than 4 points off of the closing line (the 4 point line is a bit arbitrary and certainly may change some as we get more data loaded into the system in the future but for now that’s what we will be using to separate what looks like a good bet versus a bad bet for us based on the back testing completed so far). If we were to only bet on these games, we would have an overall record of 240-192-12. Throwing out the pushes, that’s good for a winning percentage of 55.56%. Now we’re talking. That’s a win rate that I think most sharps would be perfectly happy with, and we still get the fun and excitement of betting on close to 5 games per week on average (444 games bet / 5 seasons / 17 weeks per season = 5.22 games bet per week). The chances of achieving that record randomly are around 1 in 109. This record lies outside the 95% confidence interval (+/- 2 standard deviations) that most statisticians use to feel comfortable saying something is most likely not a random occurrence. We’ll dive into this a bit more in a featured post sometime in the future.
There is, of course, always still a chance that our model has had a pretty lucky 5 season run. But at this point, the 5 seasons of data are enough for us to begin devoting a small amount of capital to this endeavor to see where it leads. Let’s not forget, it’s still called gambling for a reason! There are never any sure things!
Practical Application (a.k.a. now how do we make money?)
So this is great - we seem to have a way to identify bets where we can expect to achieve a profitable winning percentage. With unlimited budget and unlimited bet maximums, we could expect to grow our money infinitely by making these bets time after time, season after season. Early retirement here I come!
Oh so you don’t have an unlimited budget? Bookmakers don’t allow you to bet as high as you want on a given game? Oops - maybe we should set down our margaritas and apologize to our bosses for going Les Grossman on them.
The next part of this sports gambling puzzle is probably even more important than everything we just covered here. We need to figure out how much to bet on these games to both maximize our chances for profit and insulate ourselves from the inevitable run of bad luck that comes with any game of chance. Even at a 55.56% win rate, losing seasons will happen (see the 2016 season in our NFL historical bets section). If we are getting too cocky and betting too big, we will be destined for failure. That’s where we will leave it for this Betting Strategy section and kick it over to our Bank Roll management page to continue this story.