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.
NFL first TD scorer (started 2026-10-21)
Why: obscure markets might be soft. The owner put all Odds API credits on the NFL the same day, and first TD capture started then (player_1st_td).
The market's cut
Probe, NE @ CHI (Oct 22): - 7 books post the market, with 24–27 players each. - Listed implied probabilities sum to 1.22–1.63 (DK 1.30, FD 1.29, Caesars 1.22, BetMGM 1.35, BetRivers 1.35, Bovada 1.31, Fanatics 1.63). - Some first-TD probability falls outside the list (other players, defence/ST, a game with no TD), so the true cut is roughly 30–40%. - Book-to-book spread is large: Stevenson +600 to +750, Swift +360 to +450.
Implication: a single-book model edge has to beat ~30%. The only plausible route is the best price across books against a fair estimate — that is, outlier books.
Timing test: do some players score early? (timing_skew.py, no odds)
Method: - 2,761 regular-season games, 2016–2025. - If TD order were random, a player with k of the game's N TDs would score the first one k/N of the time. - Pricing first TD from the anytime-TD market assumes this.
Results (observed / expected first TDs, CI from a 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. over games):
| ratio | 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. | 2016–20 | 2021–25 | |
|---|---|---|---|---|
| RB | 1.059 | 1.013–1.107 | 1.088 | 1.030 |
| TE | ~1.01 | 0.937–1.086 | 1.039 | 0.978 |
| QB | 0.978 | 0.866–1.102 | 0.989 | 0.969 |
| WR | 0.966 | 0.923–1.006 | 0.938 | 0.993 |
| Defence/ST | 0.887 | 0.760–1.011 | 0.818 | 0.962 |
Reading: - RBs score the first TD ~6% more often than their share; the effect is smaller in the recent half. - Against a 30% cut this is tiny: it moves a 14% first-TD chance to ~15%. - Timing alone is not an edge. It is a correction for a fair-price model.
Plan
- Fair first-TD price per player: - λ_i from Pinnacle's devigged anytime TD (captured), with P(first) ≈ λ_i / Λ_game × P(≥1 TD), times the position timing factor above. - Compare with the best of the 7 book prices. - Grade forward from week 8 of 2026.
- Historical check once credits reset (Nov 1):
- Odds API historical props for 2025: one pre-kickoff snapshot of
player_1st_tdandplayer_anytime_tdper game. - Cost: ~10 credits × 285 games per market, ~5,700 total. - Answers whether best-price-vs-fair bets won, without waiting a season.
⚠ Variance: these are ~+400 to +2000 bets. A few hundred is needed before any ROIreturn on investmentProfit as a percentage of the money wagered. +2% means $2 profit per $100 bet. number means much; judge on calibrationcalibrationA deliberate sanity check on the method itself: run it on something already known to be true. If it fails to detect the known thing, the method is broken and its other results mean nothing. of the fair price first.
Historical check, 2025 (2026-11-02): no edge
Odds bought 2026-11-02 (backtests/historical_odds_nov2026/): player_1st_td + player_anytime_td,
one snapshot 10 min before kickoff, all 285 games (5,716 credits). Script grade_2025.py →
results_2025.json.
Fair price as planned, with two fixes found while building it: - Pinnacle lists only a few players in many games. Players it doesn't list get the other books' anytime price × that book's season ratio to Pinnacle (0.91–0.95; margin removal only). - The first scorer is matched on the board's own names. Timing factors and the defence/ST share (5.9%) come from 2016–24 only.
| Value | |
|---|---|
| Player rows priced / games | 8,699 / 285 |
| Log losslog lossA score for probability forecasts that punishes confident wrong answers harshly. Lower is better.: fair vs DraftKings' devigged board | .1328 vs .1331 (tie) |
| Calibration | too flat: fair 1.4% → actual 1.0% at the bottom; fair 14.2% → actual 18.1% at the top |
| Best-price bets with fair × price > 1 | 426 bets, 4 wins, ROI −55% ± 24 (median price +8000) |
| Same at claimed EV > 10% / > 25% | 1-207, −75% / 0-72, −100% |
Reading: - The fair price ties the books on accuracy, but it's too flat: it rates longshots above what they hit and favourites below. - Every "+EV" bet it finds is therefore a +5000–10000 longshot, and those lose. - The best-price-vs-fair route doesn't work with this fair price. Don't bet first TD. - Not tested (it would be a new, post-hoc model): a sharper fair price that steepens the anytime → first-TD mapping. The top bin's 18% vs 14% hints at it, but that's 199 rows found after looking.