New to these reports? Start here
- Dotted-underlined words have a plain-English definition — hover or tap them. Every term is also on the glossary page.
- "Null" means we found nothing, not that something broke. Most reports here are negative results, on purpose — knowing an idea doesn't work is the point.
- Two questions get asked separately. First, is the effect real? Second, is it already priced into the betting odds? An effect can be completely real and still useless to bet on.
- A "calibration" row is a self-check. It runs the same method on something already known to be true. If that fails, the whole report is unreliable — so it's reported alongside the findings.
- If a confidence interval includes zero, the real effect might be nothing at all, so no claim gets made.
Best Pick'em Strategy for 2026: Pick Every Favorite
Date: 2026-07-20 · Harness: strategy_search_2026.py →
strategy_search_2026_results.json · Data: 2,119 REG games, 2018–2025 · Seed: 20260720
The question
Which confidence pick'em strategy should actually be run in 2026, given the data pipeline this platform has built?
The constraint that shapes every answer
Confidence pick'em scores only the ranking of picks. The rebuild
(FINDINGS.md) established that any confidence scheme which is a monotonic
transform of spread magnitude is byte-identical to always-favorite — spread^1.5
and tiered multipliers both scored +0.00pp exactly, not approximately.
So a candidate can only differ from the baseline by:
- (a) reordering games non-monotonically with respect to the spread, or
- (b) flipping which side is picked on some subset of games.
Anything else is the baseline in disguise. All five candidates below do (a) or (b).
Pre-registered candidates
Fixed before running; all reported, win or lose. No candidate was added after seeing results and no threshold was tuned.
| Strategy | Mechanism | |
|---|---|---|
| A | always-favorite, ordered by spread magnitude | BASELINE |
| B | always-favorite, ordered by devigged moneyline probability | (a) reorder |
| C | B + confidence damped ×0.85 in divisional games | (a) reorder |
| D | B + confidence damped ×0.85 when favorite on ≤6 days rest | (a) reorder |
| E | B, but flip to the home underdog when spread ≤ 2.5 | (b) pick flip |
Multiple comparisonsmultiple comparisonsTesting many ideas at once. Test twenty things and one will usually look significant by luck alone, so results need a stricter bar.. Five candidates against one baseline — at a naive 5% level roughly a one-in-four chance something looks significant by luck. A Bonferroni-adjusted 99% interval is reported alongside the 95% and is the one that decides.
Results — % of maximum confidence points, per season
| Candidate | 2018 | 2019 | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 | ALL | SU |
|---|---|---|---|---|---|---|---|---|---|---|
| A always-favorite / spread | 72.3 | 71.2 | 74.7 | 70.8 | 72.1 | 70.2 | 75.7 | 72.1 | 72.39 | 66.4 |
| B always-favorite / devigged ML | 72.0 | 71.3 | 75.2 | 70.6 | 72.1 | 70.0 | 76.0 | 72.0 | 72.39 | 66.4 |
| C B + divisional damping | 71.5 | 70.1 | 74.2 | 70.1 | 70.8 | 70.8 | 75.5 | 72.1 | 71.86 | 66.4 |
| D B + short-rest damping | 71.5 | 70.6 | 75.3 | 69.3 | 70.7 | 69.0 | 75.4 | 72.2 | 71.73 | 66.4 |
| E B + small home-dog flip | 72.2 | 71.5 | 74.1 | 70.4 | 71.6 | 69.1 | 75.0 | 71.2 | 71.86 | 65.2 |
Season-clustered bootstrapbootstrapRe-running a calculation on thousands of resampled versions of the data to see how much the answer wobbles. The spread of those answers becomes the confidence interval. (4,000 resamples of seasons, not games — weeks within a season share teams and schedule structure and are not independent):
| Candidate | Δ vs baseline | 95% CI95% confidence intervalThe range the true value is plausibly in. If this range includes zero, we cannot rule out that the real effect is nothing at all. | Bonferroni 99% CI | Verdict |
|---|---|---|---|---|
| B devigged ML prob | +0.00pp | [−0.16, +0.18] | [−0.20, +0.24] | tie |
| C divisional damping | −0.52pp | [−0.92, −0.11] sig | [−1.01, +0.01] | fails adjustment |
| D short-rest damping | −0.66pp | [−1.12, −0.15] sig | [−1.23, +0.00] | fails adjustment |
| E small home-dog flip | −0.52pp | [−0.78, −0.22] sig | [−0.85, −0.10] | survives — robustly WORSE |
The interesting result: B ties, but not for the boring reason
B scoring exactly +0.00pp looks at first like the monotonic-transform trap again — as if devigged moneyline probability were just a relabelling of the spread. It is not. Checked directly:
- SpearmanSpearman correlationA measure of whether two rankings agree, from -1 (opposite) through 0 (unrelated) to +1 (identical). Cares about order, not exact values.(p_fav, spread_mag) = 0.99144 — high, but not 1.0.
- The two orderings disagree in 138 of 141 weeks.
So B genuinely reorders the confidence slate in essentially every week of the sample, and after eight seasons it lands on the identical point total. The disagreements are pure noise with respect to outcomes.
That is a cleaner market-efficiency demonstration than a nullnull resultA test that found nothing. "Null" is the starting assumption that there is no real effect; a "null result" means the data gave us no reason to abandon that assumption. It does not mean the data was missing or the test failed to run. usually gets. Spread and moneyline are two separate quotes the market publishes, they encode the same information about who wins, and the places where they disagree carry no exploitable signal about by how much you should be confident. The hypothesis motivating B — that a low total makes a given spread more decisive, and that the moneyline prices this while the spread ordering ignores it — is real in mechanism and worth exactly zero in points.
The other three failed, and one failed robustly
C and D encode public folk wisdom (division games play close; short rest hurts). Both lost at the naive 95% level and neither survives Bonferroni — the honest read is "no evidence they help, weak evidence they hurt." The closing lineclosing lineThe final odds right before a game starts. It reflects everything the betting market knows, which makes it the hardest benchmark to beat. already prices divisional and rest effects; damping confidence on them is adding noise to a calibrated number.
E is the one firm finding. The small-home-dog flip is the only candidate whose interval excludes zero after multiple-comparison adjustment, and it does so on the wrong side: −0.52pp [−0.85, −0.10], with straight-up accuracy dropping 66.4% → 65.2%. The classic "take the small home dog" heuristic is not merely unprofitable here — it is measurably destructive in confidence pick'em. It costs picks and wastes the confidence points assigned to them.
Recommendation for 2026
Pick every Vegas closing favorite. Order confidence by spread magnitude, or equivalently by devigged moneyline probability — they tie.
Expected performance: ~72.4% of maximum points, ~66% straight-up, which is what the market alone gets you. Marginal preference for the devigged-ML ordering (candidate B) on two non-statistical grounds: it is the market's actual probability rather than a proxy for it, and it has no fitted parameters to rot. But the choice is cosmetic — pick whichever is easier to wire.
Do not ship C, D or E.
Honest bounds
- This is a null resultnull resultA test that found nothing real. The starting assumption is "there is no effect here", and a null result means nothing in the data argued against it., and nulls are the expected outcome here. The closing line is the market's own probability estimate; this platform has yet to find a sport where it can be beaten. The value delivered is knowing what to run, not an edge.
- The search space is small. Five pre-registered candidates, not an exhaustive sweep. A wide search would have found something "significant" and it would have been noise — that is precisely why the space was fixed in advance and why the adjusted interval decides.
- The model hybrid is still untested. The advertised "trust XGBoost when spread < 3" strategy cannot be evaluated: no week-by-week model predictions were ever archived. Reporting a hybrid number without them would repeat the exact failure the rebuild documents. Archiving pick'em predictions at generation time is the prerequisite for ever answering this, and remains the single highest-value change to the pick'em surface.
- Winning a pool is a different objective. These strategies maximize expected points. In a large pool, beating N opponents may reward variance — deliberately contrarian picks that lower the mean and raise the ceiling. That was not tested and would need pool size and opponent behaviour to model.