When I wrote my last NBA model performance post in early January, the season sample was still relatively small and I was mostly trying to answer a simple question: is the model showing enough signal to keep trusting it?
Now the sample is much larger, and the answer is more nuanced.
As of March 27, 2026, the public NBA model performance page shows the model hitting 61.8% on sides across 1,033 tracked games, 52.1% against the spread across 1,008 games, and 49.0% on totals across 1,000 games. That gives me a much better sense of where the model is actually helping, where it needs tighter filters, and where I need to stay cautious.
The short version is this: I still think the model is most useful as a winner-prediction tool, but I’m being more selective about how I want to use that information in actual betting decisions.
The big-picture performance update
The strongest part of the model remains straight-up winner selection.
The live performance page currently shows 638 correct sides out of 1,033, good for 61.8% accuracy overall. That is down a bit from the 63.0% overall figure I wrote about in the January post, but the sample has more than doubled since then, which makes the current number more meaningful.
When the model agrees with the market, results are much stronger. Those picks are hitting 70.1%, and book favorites above 55% implied win probability are hitting 71.3%. Even the 56–65% confidence band is converting at 72.9%, although the very highest-confidence buckets still have samples too small to lean on.
Against the spread, the model is more modest. The current ATS number is 525 correct out of 1,008, or 52.1%. In January, that number sat at 54.9%, so this is an area where the model has cooled as the season has gone on.
Totals remain the weakest part of the board. The model is 49.0% overall on totals, with overs at 46.5% and unders at 49.7%. That is basically in line with what I wrote in January: totals still are not a model strength right now.
What I’m focusing on now for sides
This is where my thinking has become more specific.
Earlier in the season, a lot of the conversation centered around whether the model could identify winners at a high enough rate to matter. I think the answer there is yes. But prediction accuracy alone is not the same thing as betting value, especially in a market where favorites can inflate raw win rate. That was a core point in my January write-up, and I think it still applies now.
So at this point, the main thing I’m looking for on sides is pretty simple:
model-predicted winners that are plus money, really at any confidence level.
That does not mean I suddenly trust every underdog. In fact, the dashboard says the model is only 38.7% when it disagrees with the market, and the dog (<45%) bucket is only 32.1%. Those numbers are not good enough to justify blindly backing every contrarian side.
But that is not the same as saying plus-money winners are useless.
What it means to me is that I care more about price plus signal than I do about forcing an arbitrary confidence threshold. If the model is willing to pick a team to win and the market is still offering plus money, that is the kind of spot I want to pay attention to. The hit rate may be lower than it is on favorites, but the payout structure does more of the work. That is the lens I’m using more now than I was a couple of months ago. This is partly an inference from the dashboard’s split data and partly a strategy choice, not a claim that the public page has already proven this subset profitable on its own.
How I’m treating spreads now
Spreads are where I’ve become more selective.
The overall ATS number of 52.1% is not bad, but it is not strong enough for me to treat the model as having a broad, automatic edge against every spread. Some smaller pockets are better than others. Spread ≤3.5 games are hitting 54.7%, and the WinConf 50–55% group is at 53.8%. But the 56–65% win-confidence ATS bucket has fallen to 48.0%, which is a clear sign that stronger moneyline confidence does not automatically translate into a better spread play.
Because of that, I’m not looking to fire on every ATS lean.
What I’m looking for now is a more specific setup: spots where the model is close to the predicted spread as well. In practical terms, that means I want a tighter fit between the projection and the market number, rather than just using the model to force a side because it happens to lean one way. That is a strategy filter I’m applying based on the way the current ATS results have flattened out; the public page does not directly publish a “close to model spread” performance bucket yet.
So for me, spreads have shifted from a broad model output to a more selective secondary tool.
EV results still need work
One of the most interesting — and frustrating — sections on the page is still expected value.
The live page currently shows Negative EV picks hitting 70.2%, while Positive EV picks are hitting only 40.9%. The higher positive-EV bands are even worse: EV >5% is at 39.2%, EV >10% is at 36.4%, EV >15% is at 35.7%, and EV >20% is at 34.4%.
That is obviously backward from what I would want to see.
I said in January that the EV section was giving puzzling results, and that still appears true now. At minimum, it tells me I should not treat the current EV buckets as a standalone green light. Whether that means calibration needs work, the implied probability handling needs work, or the bucket definitions need work, the takeaway is the same: I’m not leaning heavily on the current EV segmentation until it makes more intuitive sense in the results.
Totals remain a watchlist area, not a strength
I do not think there is any reason to oversell this category.
Totals are 49.0% overall. Overs are 46.5%. Unders are 49.7%. Larger total-gap buckets are not producing strong enough numbers either.
So for now, totals remain more useful as a diagnostic area than a core betting lane.
That may change later as the model evolves, but right now I do not think the honest takeaway is anything other than: this is still the weakest part of the current version.
My takeaway right now
At this point in the season, I think the clearest summary is this:
The model still looks most useful for predicting winners. That part is real. The overall side accuracy is still strong enough to keep paying attention to.
But from a practical betting standpoint, I’m narrowing the focus.
Right now, I’m most interested in:
- model-predicted winners that are plus money, at basically any confidence level
- spread spots where the market line is close to the model spread, rather than forcing broad ATS action
- staying cautious on totals until the performance gives me a better reason not to
That is really where I am with it today.
I’m less interested in pretending every category is beating the market and more interested in being honest about what the model is actually doing well. The point of the public page is not to sell certainty. It is to show the numbers, track the outcomes, and keep tightening the process over time. The page itself says it best: focus on the more stable splits and treat tiny samples as noise.
That is still the approach.
No hype. No blanket claims. Just a better idea of where the model is helping, where I want tighter filters, and how I’m thinking about the numbers as the season moves forward.
— Dr. Cover