I'm really excited to be doing
NFL point spread picks for the 7th year in a row. My picks are purely computational and based on statistical algorithms that have been back-tested with data dating back to 1996. As you saw from my
previous post, we've had great success in past years and continue to learn and improve our NFL picks every year. My picks are only relevant when there is significant data explaining game outcomes because I use team-level data, I'll explain more in a second. Nevertheless, I still post my automated picks starting on week 1, so you should expect this week's picks tonight or tomorrow night at the latest. For this post, I'm going to tell you:
- how I started doing this
- a simple explanation of the methodology
- why it's only relevant after 4 weeks of data
- what to expect this year
How I Started Analyzing NFL Point Spread Data
In 2006, while I was a Statistics and Operations Research Ph.D. candidate, I met an econometrics professor who had published multiple articles on creating techniques for statistical modeling when there are significant outliers in the data. Some were regression-based and others were cluster-based. He had a few applications of his methodology in sports including figure skating and horse racing.
He gave the Statistics department a talk about a new research on applying some of his techniques to NFL point spread. I was intrigued and approached him. We spend the next 3 months refining his technique. He wrote his paper and I graduated.
Although back-testing (making point spread predictions without knowledge of the data) gives you a broader sense of how accurate an algorithm can be, it is not as fun as making real-time predictions. As a scientist, I needed a journal to keep track of my results and learn every week how well these algorithms were. That's when
NFLpickles.com was born.
After 6 years, I've revised a few of the original algorithms I was exposed to and continue to make adjustments every year. It's been a fun and successful ride and hope to continue this "journal" for years to come.
A Basic Explanation of the Methodology
I use team level offense and defense stats, game results, and spreads dating back to 1996. Having said that, the question I most often get is, how do you account for injuries? The answer is I don't. Weather? Neither. I only use the data I have, create a home-field advantage score, a strength of schedule metric, a point spread prediction, and a probability of how likely I'm on the right side of the spread. I then use the probability to pick the NFL games to bet on each week.
This is a one man operation which is why 1) I don't have money or time to get player-level data and 2) I only do NFL (it's weekly rather than daily!). Also, I'm satisfied with the ROI I get every year so there's no reason to go more granular at the player level data.
The only problem are injuries. What happens if Peyton Manning is not playing week 9? I simply would take that game out of the equation. It'd be a no bet. The only concerns are "main" players: QBs and RBs generally although I've made some exceptions. Throughout the years I've marked my data with games where there is a significant injury so they models "know" this is an outlier. Other times I've missed the mark and placed bets on games that I shouldn't have because I wasn't aware that a key player was out. So please, let me know if I'm putting a game for bet and I'm not aware that a starting QB or RB isn't playing.
As for the methodology per se, you weren't expecting this magician I mean statistician to reveal it's secrets right? I won't but I tell you that it is a set of algorithms that compensate for outliers in the data all working to make accurate predictions. Then I take the best of the best each week, like a democracy, to make predictions. I have a method to compute probabilities that a bet will be correct based on the algorithms that are pointing at a specific game, that is used to decide if the game will be a bet. Most times, we pick between 3 to 5 games each week.
Why Are You Starting Your Official Picks at Week 5?
Very simple. It's team level data. Teams change, players change, model needs data. If I start making picks at week 1, then they would be purely made based on last year's data. The models also create "team rankings" based on strength of schedule and those need at least 4 to 5 games to start stabilizing. Team momentum statistics also need a few weeks of data to be more precise. As for pre-season, you have varying player times which do not align well with our models. I just don't do it and I go more deeply regarding
betting NFL pre-season games here.
What to Expect This Year
Unofficial picks from week 1 to week 4. That means that I'll run my algorithms and publish the picks as I've done the past 6 years but won't tabulate the results on the right hand column. Official picks start on week 5 and results will be tabulated. We also make recommendations on NFL betting strategy based on
Kelly's formula and use the results to publish ROI. All picks and betting size recommendations will only be available to premium subscribers (more on that at a later post).
I expect to have a great season and hope to improve last year's 44% ROI!
