NBA Model Beta Test Picks 2.20.25
We are turning our NBA beta test live post All-Star Break! As mentioned in the announcement before the Super Bowl, our goal is to broaden our predictive models to more sports and not just the NFL. We hope to eventually capture and replicate the success our football models have produced so far across a year-round product offering rather than waiting for the NFL to kick off again in September.
The next step of this journey is to flip the switch on our NBA model beta test from internal behind closed doors to live for you to see. While this doesn’t change much from a data and modeling perspective, it will drastically help us improve and refine our process of where, when, and how we get picks out. These lessons will be learned through a series of trial and error, so thank you in advance for your patience and support.
I will repeat it again since it is so counterintuitive, but accuracy is not the main goal of our beta test. Of course, we would love it if our models were profitable against the spread between now and the playoffs, but we are not setting any performance goals so early in the process. So, hitting above 52.4% ATS would be a bonus for our testing but not a success-determining goal like it is for our official NFL model each season. We are not ready for that quite yet.
If you have read our work before, you know that transparency with our readers is a main focus for us. This applies to model strategy as much as it does to model results. So let’s dive into our strategy for the NBA model as it stands now for our testing. The NBA training dataset has been the main project for our team of interns, led by the ‘Orb Apprentice’ (we are keeping names anonymous out of respect for people’s privacy). He and I have each developed our own predictive models and are treating the overlapping consensus picks as the ‘official’ model picks, similar to what we do with our NFL multi-model strategy. So for now, we have two models and are working on a third to see how well it interacts with the others. Due to the inherent variance of NBA scoring, each of our models is classification-based rather than linear. While this has limitations, so far it has led to higher accuracy and given us a better chance to capture the randomness of NBA games.
All that being said, live testing so far has been very promising. Since both models were developed on January 15th, here is how our models have performed against the spread when overlapping on picks:
During this stretch, favorites have covered 43.4% of all games, while underdogs covered 56.6% of the time. Based on the distribution of picks shown in the confusion matrix above, our models may be over-indexing on underdogs and adapting to the real-world result class imbalance. A positive sign is that so far, the models are performing just as well when predicting either class (predict favorite vs predict underdog). While these results are from live testing rather than backtesting, I do not expect them to maintain this level of accuracy for the remainder of our beta test. Keep in mind, that the results include plenty of losing nights which is one of the few performance trends I can promise will continue. We also are starting by only giving picks out a few days of the week, so the models could happen to perform better or worse on those days than they do overall. As always, if you are looking for a model that ‘guarantees’ nothing but winning performances each night when predicting spreads (50% expected accuracy over the long term) sports gambling and Orb Analytics aren’t for you, especially when doing a test of new models.
One thing, in particular, I will be looking out for is how favorites perform ATS down the stretch. During the NFL season, we saw favorites dominate heading into the playoffs as better teams were fighting for playoff seeding while bad teams were losing for draft position. Will we see the same thing happen in the NBA? If so, how will our models perform with no class imbalance or even if favorites tilt the scale in the opposite direction?
Now to officially turn this beta test live, here are the consensus model picks for tonight’s slate:
Spreads:
Memphis -2.5
Atlanta +1.0
Clippers -2.5
Charlotte +16.5
Portland +3.5
You can find more Orb NBA content throughout the season on our NBA podcast ‘Dishing and Swishing’ hosted by Mickey Swish and The Vibe Navigator:
2025 All Star Weekend - Kia Skills Challenge
Image Credit: Kyle Terada-Imagn Images
2025 All Star Weekend - AT&T Slam Dunk Contest
Draymond Green vs. The Vibe Navigator
Image Credit: clutchpoints.com
EP12 - Is the all star game supposed to be bad?
Even just in a test, releasing predictive model picks for a 2nd sport is a huge day for the Orb project so I wanted to acknowledge and thank everyone who has supported us so far along our journey! I would also like to give a special shoutout to our entire team of data interns, both past and present, who have done incredible work to get us to this point on the NBA side of things.
- Team Orb Analytics
.
.
.
.
.
.
DISCLAIMER:
The information provided on this website is for informational purposes only. It is not intended to be gambling or financial advice, and should not be relied upon as such. We are not responsible for any actions or decisions taken by readers based on the information provided on this website.
The picks and predictions provided on this website are based on our own research and analysis, and are intended to be used for entertainment and informational purposes only. We do not guarantee the accuracy or completeness of the information provided, and we are not responsible for any losses or damages incurred as a result of using this information for gambling or other purposes.
By accessing and using this website, you acknowledge and agree to the terms of this disclaimer, and you assume all risks and liabilities associated with your use of the information provided on this website.
 | A guest post by
|
 | A guest post by
|