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.
Season Prize vs Weekly Prize: Shoot for the Season
Date: 2026-07-20 · Harness: pool_ev_analysis.py →
pool_ev_analysis_results.json · 4,000 simulated seasons per regime · Seed: 20260720
Stakes
| Contest | Cost | Prizes | Pot in | Pot out |
|---|---|---|---|---|
| Season-long | 1u/week → 18u | 270 / 180 / 90 to top three | 30 × 18 = 540u | 540u |
| Weekly | 1u/week → 18u | 30u to that week's winner | 30 × 1 = 30u | 30u |
Both pools are zero-rake — every unit collected is paid back. That makes this a pure skill question: the average entrant breaks even by construction, so any edge is transferred directly from the weaker players. Break-evens:
- Weekly: P(win a given week) > 1/30 = 3.33%
- Season: E[prize] > 18u, against a placement structure — not just first
The structural asymmetry
Winning a week means beating the maximum of 29 rivals on one 16-game slate — an extreme order statistic. The max of 29 draws lands far above any individual's expected score, so the mean-optimal slate rarely tops it; beating a maximum needs a fat right tail, i.e. variance. Winning the season aggregates 18 weeks, noise averages out, and accuracy wins.
The top-three payout is what settles it. A placement structure rewards consistency, which is precisely what accurate play produces and what a variance slate destroys. Baseline finishes top-3 48.8–86.9% of the time; the high-variance slate finishes top-3 1.8–4.2% of the time.
Results (credible rival regimes — see FINDINGS_POOL_2026.md)
| Rivals | Strategy | E[pts] | P(1st) | P(top3) | EV season | P(week) | EV weekly | EV both |
|---|---|---|---|---|---|---|---|---|
| K=3 | baseline | 1596 | 51.0% | 86.9% | +173.0 | 1.34% | −10.8 | +162.2 |
| K=3 | flip 2 @ top conf | 1484 | 1.6% | 4.2% | −10.3 | 6.47% | +17.0 | +6.7 |
| K=5 | baseline | 1595 | 22.6% | 60.6% | +94.9 | 0.75% | −13.9 | +80.9 |
| K=5 | flip 2 @ top conf | 1483 | 0.6% | 1.8% | −14.8 | 9.93% | +35.6 | +20.9 |
| K=8 | baseline | 1597 | 16.4% | 48.8% | +69.0 | 0.93% | −13.0 | +56.1 |
| K=8 | flip 2 @ top conf | 1485 | 1.1% | 1.8% | −14.1 | 13.71% | +56.0 | +42.0 |
| always-fav | baseline | 1597 | 22.2% | 51.2% | +81.5 | 3.51% | +1.0 | +82.5 |
| always-fav | + flip 1 tossup | 1596 | 23.4% | 48.6% | +79.0 | 5.89% | +13.8 | +92.8 |
| always-fav | flip 2 @ top conf | 1485 | 1.7% | 2.9% | −11.8 | 20.50% | +92.7 | +80.9 |
Verdict
Play the season. Skip the weekly.
Season with baseline is +69u to +173u on an 18u stake — a 4× to 10× return, positive in every credible regime. Only the magnitude depends on rival skill; the sign never does.
Weekly with baseline picks is −13.9u to +1.0u — negative in three of four regimes. Correct play wins a given week only 0.75–3.51% against a 3.33% break-even. Accurate play is actively bad at winning single weeks because it has no right tail: someone among 29 rivals spikes a lucky slate nearly every week, and the mathematically best slate is rarely the week's high score.
The variance play is now dominated outright. Under the old winner-take-all assumption it beat baseline on combined EV in sharp pools. Under the real top-three structure it does not win a single regime: baseline's combined EV beats it in all four (+162 vs +7, +81 vs +21, +56 vs +42, +83 vs +81). Chasing weekly lotteries forfeits a placement structure that pays three deep — and placement is where accurate play banks most of its value.
Recommendation
- Enter the season and play baseline — every favorite, ordered by devigged ML probability, plus the single tossup flip. The flip stays: it costs ~0.4 points a season, and it is the only thing that rescues you from the clone regime, where it is the best combined play (+92.8 vs +82.5).
- Do not pay for the weekly. With baseline picks it leaks ~11–14u a year. Save the 18u.
- Do not split the difference. A slate hedged toward weekly variance gives up far more season EV than it recovers.
Corollary on "I never won a week": at 0.75–3.5% per week, expected weekly wins per season is 0.14–0.63. Zero is the modal outcome for correct play — roughly a once-every-four-seasons event. It was never evidence of playing badly.
Honest bounds
- The rival model is unvalidated (see FINDINGS_POOL_2026.md). The season recommendation's sign is robust across the entire sweep; the weekly numbers are much more sensitive to it, which is itself an argument for the season.
- The sim is partly circular: outcomes are drawn from the market's
p_favand baseline picks byp_fav, so rivals can only add noise, never out-read the close. Defensible for this platform, but it inflates baseline's edge against a genuinely sharp opponent. - EV is not certainty. +94.9u expected still means finishing out of the money ~39% of the time at K=5. A zero-rake pool means the edge is real, not that the variance is small.
- Assumes one entry feeds both contests. If a separate weekly slate is allowed, the problem decouples: play baseline for the season and a variance slate for the weekly, banking both (+81 and +36 at K=5). Still worth checking.
- 2025 schedule structure; ties split prizes as
1/(co-leaders). - Not modelled: in-season switching — run baseline while season equity is live, then switch to variance once mathematically out of contention (~week 12+ if well behind). Captures most of both prizes; the natural next build.