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.
Do pick'em players get gutsier late, or when behind? (2026-10-19)
Pre-registered in desperation_test.py (commit 1fc4334, before any result).
Data: our Yahoo pool, 2025 wk 1-12 and 2026 wk 1-6, 7,237 usable picks from 62
member-seasons (week 1 dropped: no standings). Overall underdog rate 18.8%.
Controls: spread bin + member x season fixed effectsfixed effectsA design where each subject is compared against themselves, which cancels out their permanent differences. Here it means comparing a player to his own other seasons rather than to other players..
Results
| Test | Effect | 97.5% CI / p | Verdict |
|---|---|---|---|
| H1 time (week 2 -> 12) | +4.8pp dog rate | coef CI [0.013, 0.062], p=0.0006 | passes as registered |
| H2 standings (leader -> last) | -0.1pp | coef CI [-0.50, +0.48], p=0.97 | no effect |
Robustness: H1 positive in both seasons (2025 +0.39pp/wk, 2026 +1.4pp/wk) and in the LPM (+0.46pp/wk). H2 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. in every form: points-behind z (0.00pp/SD), each season alone, and with the member's prior dog rate controlled (+3.9pp, p=0.23). Late-season x standings (2025): nothing (p=0.63). Dog picks don't get more confidence when behind (p=0.21) or later (p=0.29).
Caveat on H1 (added after the registered run, flagged as such)
Everyone picks the same games, so a week with one popular dog moves all members together. The registered SEs clustered by member ignore that. Clustered by season-week (17 clusters) the same estimate gives p=0.064, CI spanning 0. Leave-one-week-out estimates stay positive (0.029-0.056/wk), so no single week drives it, but the evidence is "probably, modest", not conclusive.
External check (E3, Pickwatch public fans 2014-2025, 3,214 games, clustered by season-week): fan dog share rises +2.5pp from week 2 to 17, p=0.007. Same direction, about half the pool's slope.
Reading
- Players drift a few points toward underdogs as the season goes on, the public too. It is NOT desperation: the drift is the same for leaders and trailers. Raw dog rates by standing third: top 18.4%, middle 16.1%, bottom 21.6% -- the bottom third's excess is who they are (fixed effects absorb it), not where they sit.
- H2's CI is wide (roughly +/-7pp leader->last), so a moderate chasing effect isn't ruled out; a large one is.
- For the pool model: no standings term. A time drift of ~+0.4pp/week could be added to rivals' favorite rate; for a leader it slightly lowers rivals' expected points and raises their variance, roughly offsetting. Not worth a model term at this precision. Re-run after 2026 week 12 (registered script, unchanged) for a third look.