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.
Survivor strategies vs the REAL ESPN field, 2020-2025 (2026-10-24)
survivor_sim.py had to model the other players. ESPN's free survivor game
publishes every week's pick count per team (public gambit API, 2020+:
scripts/nfl/espn_survivor_picks.py → [redacted],
survivor_pick_shares; ~0.2-1.2M entries in week 1). survivor_real_field.py
replaces the modeled field with the measured one.
The real field is casual, and herds less than the sim assumed
- Most-picked team averages 29% of the field (the sim's "mild" field had 42%, "heavy" 64%).
- Best-fitting rival model: beta ≈ 2 (cross-entropy 2.84; beta 3: 3.05; beta 6: 4.29; perfect = 2.27). The model fits only loosely.
- The field's average pick wins 71% of the time in weeks 1-4; the biggest available favorite wins 80.5%. The gap holds in later weeks too (72 vs 82, 68 vs 77), so it is not an artifact of eliminated players still picking.
Pot equity (x entry fee, zero rake)
Outcomes drawn from the closing moneyline (4,000 seasons each), the field picking with that week's real ESPN shares, pool of N rivals drawn from it.
| Strategy | N=10 | N=50 | N=500 |
|---|---|---|---|
| Greedy (biggest favorite) | 2.11 | 3.42 | 7.44 |
| Plan ahead (F) | 2.01 | 3.25 | 7.20 |
| F + fade popular, w=0.5 (C50) | 2.04 | 3.46 | 8.26 |
| F + fade popular, w=1 (C100) | 1.97 | 3.30 | 7.74 |
| C100 with modeled shares | 1.94 | 3.00 | 6.66 |
| Copy the field's top pick | 1.51 | 1.61 | 1.30 |
vs greedy, season-level (6 seasons): C50 +0.04 (SE 0.13) at N=50 and +0.81 (SE 0.31, better in 6 of 6) at N=500. F −0.16 (SE 0.23) at N=50.
Reading
- Against a field like ESPN's, the edge is just discipline: take a big favorite every week. Any of the tested strategies earns 2-3x the fee in 10-50 person pools, because the typical entry picks a 71% team instead of an 80% one.
- Planning ahead and fading the public add little here. The sim's big gains for those came from heavy herding; this field doesn't herd enough. Fading popular picks is worth it only in large pools (N=500).
- Literal replay on real outcomes is a lottery: in 2022 and 2024 every strategy and most of the field died in week 1. Six seasons can't rank the strategies on actual results; the table above uses simulated outcomes.
Caveats
- ESPN's free field is the softest field there is. Paid pools (office pools with money, DraftKings/Circa survivor) are sharper; expect much smaller multiples. Yahoo's field is more concentrated (week 7 2026: KC ~47% per RotoWire vs ESPN's top pick DEN 20%), which makes fading the public more valuable there.
- Shares are reused under simulated outcomes, though different results would change who is alive and what they pick.
- ESPN counts are entries that picked, not strictly entries alive (some weeks week-to-week counts exceed the previous week's survivors).
- Rake, buybacks and multi-pick weeks are not modeled.
Forward paper test
scripts/nfl/survivor_paper.py → survivor_paper_picks in pickem_field.db.
G, F, C50 and C100 each run one entry from 2026 week 7 (all teams unused, so
it's a late start). Cron: picks Sun 12:45 UTC (live no-vig moneyline, ESPN
live shares), grades Tue 12:30 UTC. Week 7: all four took HOU (80%, 14.5% of
ESPN). One entry per strategy is an operations check, not evidence.