The CWS Tracker: Where the Model Disagrees with DK
The CWS Tracker has one job: show where the model and DraftKings actually disagree on this year's bracket. The page runs the 64-team field 50,000 times and lines the result up next to DK's implied prices. Six rows are where the model has a real opinion this year. Three at the top of the bracket, three among the 2-seeds, zero among the 3-seeds.
The CWS Tracker is live at /cbb-cws-tracker. It has one job: show where the model and DraftKings actually disagree on this year's bracket, and by how much. The page runs the 64-team field 50,000 times and lines the result up next to DK's implied price on every team.
Below: how the numbers get made, how not to use the page, and the six rows where the model and DK disagree this year.
How the numbers get made
Each team carries a composite rating with five layers.
| Layer | What it is |
|---|---|
| Anchor | PEAR composite, the underlying talent rating |
| Form | recent results, time-decayed at the game level |
| Injury | per-pitcher rating deltas with a per-day decay schedule |
| Resume | strength of record, capped at ±1.5 runs, weight 0.072 |
| Home field | a small per-game term |
In basic terms, you are summing five inputs to get one rating per team, where one rating point is worth about nine percentage points of single-game win probability. A confirmed Tommy John on a Friday starter nets about -0.9 to -1.0 points, which moves a regional final about eight points the other direction.
When running the sim, it is also important to know how pitching workload carries between games. Ideally, you want the model treating a Game 1 ace as tired going into a Game 3 start, not fresh. The sim does this with an α=0.35 fatigue carryover between rounds and α=0.75 between games of a single series.
Pitcher news flows in through a Brave search layer with TTLs by category. Tommy John holds 90 days. Out-for-season 30 days. Day-to-day clears in 3 days. The field itself is the official NCAA bracket as released Selection Monday, and probabilities recompute on each game finalization.
How not to use the page
Three things the tracker is not built to do.
The modal walk is not a futures pick at full size. The walk shows the most-likely path through the bracket. The joint probability of that exact path actually happening is 3.97 in a million. The model's real edge on any one champion sits in the 4 to 8 point range above the market, which is the number to use for futures math.
The Total column on the per-game cards is empty for a reason. We walk-forward-audited the totals model. MAE 5.85 versus DK 4.15. The over/under win rate matches an always-UNDER baseline at 53.8%. In the bucket where the model claims its biggest edge, MAE balloons to 9.97 versus DK 2.94. Putting our totals number next to a clean DK line might look better on paper, but that should not be the only consideration. The cell stays empty until the model gets rebuilt.
Don't grade the model on host prediction. D1Baseball nailed 15 of 16 hosts this cycle. We nailed 13. Host prediction is committee-resume modeling. Our model predicts game winners. They are different jobs, and 13 of 16 is not a signal of model quality either way.
Where the model disagrees with DK
Three rows at the top of the bracket.
| Team | Model | DK | Edge |
|---|---|---|---|
| Georgia Tech | 22.5% | 14.3% | +8.2pp |
| UCLA | 20.1% | 14.3% | +5.9pp |
| Georgia | 11.9% | 7.1% | +4.7pp |
The model takes less of Mississippi State (-6.5pp), North Carolina (-3.9), Texas (-3.5), Texas A&M (-3.2), and Auburn (-2.0). Three 1-seeds the model wants more of and five it wants less of is an unusual shape. Most cycles, market and model lean the same direction at the top of the bracket.
Three rows among the 2-seeds.
| Team | Regional | Model | DK | Edge |
|---|---|---|---|---|
| Tennessee | North Carolina | 28.5% | 22.2% | +6.3pp |
| Cincinnati | Mississippi State | 19.5% | 14.3% | +5.2pp |
| Oregon State | Oregon | 42.1% | 37.7% | +4.3pp |
Tennessee and Cincinnati are the cleanest reads. Oregon State carries Dax Whitney's confirmed Tommy John on the injury layer, and the resume layer keeps them above the market price after netting that hit. Two more 2-seeds run +2 to +3 points and round out the same lean: Oklahoma at Georgia Tech (+3.0) and Miami at Florida (+2.8).
Zero rows among the 3-seeds. We checked all 16. Arizona State at Nebraska is the only non-negative read and sits at +0.6pp, which is noise. Every other 3-seed prices at or below the model number, most by four to nine points. The biggest negative gaps are Missouri State at Kansas (-9.6pp), USC Upstate at Alabama (-8.3), and Liberty at Georgia (-6.9). DK has the 3-seed line dialed in this year. The upset path the market is underpricing is not through the 3s. It is through the 2s.
The biggest takeaway is that the model and DK are crossed at the top of the bracket more than usual, and the 2-seeds are where the market has the cleanest underprice. About 75 tournament games clear between Friday and the end of Omaha. Whether the model is right on the six rows above will be on the calibration page by the end of June. Either it shows up or it doesn't.
Model champion probability, top 6
Two-seeds the model likes more than DK
More from The Margin
All articles →We rebuilt the quarterback rating from scratch. Accuracy won.
Success isn't wins — it's points per drive, and the QB controls twice as much of that as he does winning. Accuracy is the engine. 'Clutch' is noise. QBR is the least stable number in football. And the rating that beats it credits the thing the old stats ignore: running.
We rated every college quarterback. Then we learned the Heisman isn't a quarterback award.
There's no public EPA in college past 2021, the schedules are a free-for-all, and the biggest award is decided by story as much as by snaps. So we built an opponent-adjusted rating that survives all three — then learned that predicting the Heisman means measuring the team, not just the quarterback, and that even then a stubborn slice of the trophy is pure narrative.
Omaha is set: eight teams in, and the model likes Georgia
Georgia leads the model's board at 41.8 percent, North Carolina sits second, and Troy is the 0.4 percent story nobody bracketed. The full title board, plus where to watch the model disagree with the book.