The 2026-27 season is a little over a week old, and on the 70 games so far that have a closing betting line, our three models have picked 56.7% of winners on average. That's well below the mid-60s they post over a full season. Is something broken?
Probably not. Over the last three seasons, October has been the weakest month of the year for our models, and for the betting market too. Early-season games are genuinely harder to call, and the reason shows up in the market's own numbers.
October is the hardest month for everyone
We took every regular-season game on our track record that also has a closing line, about 3,250 games across three seasons, and split them by month. Our models picked 56.8% of October winners and 69.8% of April winners. The market's closing line, the final consensus of bettors who have every piece of news, made the same climb, from 64% to 75.9%.
Figure 1. Winner accuracy by month: our models and the betting market
Source: HoopModel track record: 3,248 regular-season games with a closing line, seasons 2023-24, 2024-25, 2025-26. No lines were recorded for October and November 2025, so October covers two seasons.
Show the data
| Month | Games | Our models | Market | RF | NN | XGB |
|---|---|---|---|---|---|---|
| Oct | 111 | 56.8% | 64% | 56.8% | 56.8% | 56.8% |
| Nov | 439 | 63.8% | 67.4% | 67% | 61.5% | 62.9% |
| Dec | 470 | 66.5% | 68.1% | 67.7% | 65.3% | 66.4% |
| Jan | 689 | 62.5% | 67.8% | 62.2% | 62.7% | 62.7% |
| Feb | 516 | 61.4% | 70% | 59.3% | 62% | 63% |
| Mar | 707 | 67.6% | 71.1% | 65.1% | 68% | 69.6% |
| Apr | 316 | 69.8% | 75.9% | 66.5% | 71.5% | 71.3% |
The models trail the market in every month, which is expected: the market reacts to late scratches and lineup decisions that come after our predictions are made. October's gap, about seven points, is among the widest, but it rests on only 111 games. February's gap is wider still, at about 9 points. February is trade-deadline month, when rosters change and recent box scores briefly describe teams that no longer exist: the same problem as October, in miniature. That's our best explanation so far, not a tested one.
Games get more lopsided as the season goes on
Why would the market, with all its information, also be worse in October? The size of the closing spread tells you. In October the average spread is 5.8 points, and only about a quarter of games have a favorite by 8 or more. By April the average is 9.1, and about half of games do.
Figure 2. How lopsided the market expects games to be: average closing spread by month
Source: HoopModel track record: 3,248 regular-season games with a closing line, seasons 2023-24, 2024-25, 2025-26. No lines were recorded for October and November 2025, so October covers two seasons.
Show the data
| Month | Games | Average spread (pts) | Spread of 8+ points |
|---|---|---|---|
| Oct | 111 | 5.8 | 26.1% |
| Nov | 439 | 6.1 | 29.4% |
| Dec | 470 | 6.3 | 30% |
| Jan | 689 | 6.5 | 32.5% |
| Feb | 516 | 6.9 | 36.2% |
| Mar | 707 | 7.6 | 42.3% |
| Apr | 316 | 9.1 | 51.3% |
Two things push the spreads apart:
- Nobody knows yet how good the teams are. Rosters changed over the summer, players are returning from injuries, and last season's record is only a rough guide. Bettors price that uncertainty in with smaller spreads.
- Teams separate. Over the season, contenders pull away. By March and April, some teams are resting players or playing for draft position. More mismatches means more predictable games.
So most of October's lower hit rate is about the games, not the predictor. A coin-flip game between two teams nobody can rank yet is hard for anyone.
Where our models fall short early
Our models learn from box scores: recent form over the last handful of games, season-to-date numbers, and a running team-strength rating. In the first week, recent form is built from two or three games and season-to-date numbers barely exist. The market can lean on preseason expectations, roster news and scouting instead.
Would feeding the models more history help? Not more old history. We tested that directly: on a fixed test month, training on half a season versus everything since January 2022 (more than five times as many games) made no reliable difference.
Figure 3. Would more old data help? Accuracy on one test month by length of training history
Source: HoopModel experiment: fixed test month, five seeds per training length.
Show the data
| Training history | Model | Training games | Accuracy | 95% range |
|---|---|---|---|---|
| Half a season | RF | 819 | 74.5% | 68.8–80.2% |
| Half a season | XGB | 819 | 72.4% | 66.6–78.2% |
| 1 season | RF | 1,018 | 76.5% | 71–82.1% |
| 1 season | XGB | 1,018 | 75.3% | 69.7–80.9% |
| 2 seasons | RF | 2,064 | 74.5% | 68.8–80.2% |
| 2 seasons | XGB | 2,064 | 73.7% | 68–79.5% |
| 3 seasons | RF | 3,071 | 74.4% | 68.7–80.1% |
| 3 seasons | XGB | 3,071 | 75.1% | 69.5–80.8% |
| All since 2022 | RF | 4,548 | 76.5% | 70.9–82% |
| All since 2022 | XGB | 4,548 | 77.7% | 72.3–83.1% |
Older seasons describe rosters and a style of play that have since moved on. What the models are missing in October is this season's games, and the only way to get those is to play them.
How to read early-season picks
- Expect our models to hit in the high 50s to low 60s in October and early November. Even the betting market has historically managed only about 64% in October.
- Judge the models against the market line on the track record page rather than against a fixed number.
- By December, our models have historically been around 67% for the month, and the gap to the market has been at its smallest.
As for this season: on those 70 games the market has picked 67.1%, while our models are at 56.7%. That's a wider gap than in a typical October, but 70 games is a small sample, with a 95% range of roughly ±12 points around our number. We'll look again in our first monthly check-in at the end of November.