Which devig method matches results? (2026-10-31)

Answer: use exact Shin. For a two-way market it gives exactly the same numbers as the additive method (take half the vigvigThe bookmaker's built-in cut. It is why a coin-flip bet at -110 needs you to win about 52.4% of the time just to break even. off each side). Our production "Shin" was a shortcut that over-corrected: it shaded favourites about 0.8pp too far on lopsided lines. Proportional devigdevigRemoving the bookmaker's cut from odds to recover the market's actual implied probability. errs the other way by the same amount. Fixed in [redacted].

On near-even lines (spreads, totals, most props) every method agrees to within about 0.1pp, so the choice barely matters there. It matters on lopsided moneylines and longshot props.

Data

There are 134,560 closing two-way lines that have results, built by build.py into data/closes.parquet. The close is each book's last price for both sides within 6 hours before the start.

Pushes and ties are dropped.

Methods compared

• mult: proportional.
• add: additive.
• power.
• shin_exact: solves for z so the fair probabilities sum to 1.
• odds_ratio: Cheung's method.
• shin_prod: the old production code. Its closed-form z is roughly twice the exact z, and it then renormalised the result.

Results

Pooled, all 134,560 lines. The table shows log losslog lossA score for probability forecasts that punishes confident wrong answers harshly. Lower is better. relative to the best method, with standard errors clustered by game.

Lopsided lines, where the methods disagree (favourite raw-implied ≥ 65%, n = 21,385):

NFL sides, 2006–2025 (n = 3,025 favourites at 65%+): the favourite won 74.3% ± 0.8. Proportional said 73.8%, exact Shin 74.5%, and old production Shin 75.1%. This explains the earlier note that "Shin over-corrects NFL sides" (memory: pickem strategy): that was the shortcut, not Shin itself.

By sport and market: no market has a winner that clears noise once you allow for about 30 groups and 5 comparisons each. - The largest gaps are NCAAB moneylines, where mult is worse by 2.3 ± 1.3, and NFL rushing-yards props, where the old Shin is worse by 0.68 ± 0.27. - Choosing a method per market would be fitting noise, so one method is used everywhere.

Changes made

• devig_shin now solves for z exactly. Every caller of devig_auto/devig_shin changes from 2026-10-31: EV alerts, closing fair probability, paper trading CLVclosing line valueWhether you got a better price than the market settled at. Widely used as a faster signal of skill than profit, which takes ages to measure. and line shopping.
• The 10 paper-trading bets that already had a close were recomputed. Most moved by less than 0.5pp of CLV. The longshot anytime-TD bets moved most: +360 went from −8.7% to −3.5%.
• EV alert history rewritten 2026-10-31 with scripts/clv/rewrite_fair_exact_shin.py.
• Backup: [redacted].bak_pre_exact_shin_20261031_181223.
• Each stored closing_fair_prob was matched to the snapshot pair that reproduces it under the old formula, then re-devigged with exact Shin. 25 early rows had used proportional or power devig and were matched the same way.
• 4,537 of 4,870 rows were rewritten. The other 333 have no pair that reproduces them and were left unchanged: 252 February–March props, 73 May spreads and totals, and 9 point-adjusted props.
• Alert-time fields (pinnacle_fair, true_prob, edge_pct, ev_pct) were left as they were alerted.

• Effect on reliable rows (n = 3,495): mean clv_fair_pct went from +0.30 to +0.14 pp, and the share with positive CLV from 59.3% to 54.7%. Alerts lean toward favourites (the push gate is favourite-only), and the old shortcut gave favourites too high a fair probability at close.

  Sport	Before	After
  MLB	−0.34	−0.42
  NBA	−0.42	−0.54
  NCAAB	−0.15	−0.45
  NFL	+1.25	+1.24
  NHL	+0.76	+0.54
  WTA	+0.79	+1.11
  - Left proportional by design: model inputs and model grading (nfl/models/joint_score.devig_home, scripts/nfl_model_clv.py, fantasy market_projection). The joint model tested proportional, power and Shin as a tie.
  - Untested: markets with more than two outcomes (3-way soccer, first-TD groups) still use proportional, because there were too few graded rows.

Side note

Heavily juiced player props did not consistently lose or win more than priced. - Receptions favourites at 65%+ won 55.6%, against 63.4% priced (n = 809, 108 games, about 2.4 SE). - Passing yards and passing TD favourites won more than priced.

With about 25 groups tested this is consistent with noise. It isn't a lead without a fresh sample.

Run: python3 backtests/devig_calibration/build.py && python3 backtests/devig_calibration/evaluate.py