Track record

The short version

• Where the method works: picks the market prices well above break-even hit more often than break-even needs (56.8% vs 54.1%), and picks priced below break-even lose as predicted (49.1%, -9.1% per leg, n=5731). That asymmetry is the strongest result so far.
• Where it is not proven: the flagged group's 95% interval (51.3%–62.1%) still includes break-even, and once split by sport, stat, side or app, 0 of 295 segments clear the bar. Treat every individual edge as a market comparison, not a promise. As an entry, a 6-pick Flex Play built from legs hitting at 56.8% would have +26.1% ROI before any real-world friction.
• Biggest limit: the free history is MLB (173 games, 2026-09-15 to 2026-09-27) and NFL (32 games, 2026-09-18 to 2026-09-29). There is no free graded history yet for NBA, NHL or college basketball, so those sports show "history still building" until live grading collects enough.

Validated edges

None yet. To be called proven, a segment must have 100+ out-of-sample legs, a 95% lower bound above break-even, and a hit rate above break-even in the earlier training half too. The Board's "Top picks · proven edge" section stays empty until something passes.

Closest to proven (not proven)

Ranked by the out-of-sample lower bound. These are the segments to watch as more graded data arrives.

Running record · backtest

Flagged picks (2+ pts edge on PrizePicks, Underdog and Dabble), cumulative by game date, with the 95% interval and break-even.

Live graded results

Published picks

A separate record for the legs EdgeLeg locked as Tonight's picks on the board. This is not the same as the broader live +EV sample above.

AI-suggested slips

A separate record for the 2–6 leg slips EdgeLeg suggested from the board. Graded after every leg finishes, using each app's verified payout table.

By edge size

Edge = fair win probability minus the break-even of the app's default entry. Cumulative rows (2+, 4+, 6+) overlap the bands.

By app

By sport and app

By stat type

By stat type and side

Overs vs unders

Split by side, overall and then by sport and stat. Unders are also broken out by edge size. A row is only labelled proven if it clears the same out-of-sample bar as every other edge on this page.

Overall

Every graded +EV-eligible leg in the sample, all edge sizes.

By sport

By sport and stat

Unders by edge size

Pooled across sports, then the same unders split by how far they sat from break-even.

Overs vs unders by app

Same split, further cut by app and the flagged or below-break-even buckets.

Blowouts

What happens to props when a game runs away. A game is a blowout only after it ends, from the final margin. Pre-game risk is a separate call, from the sharp spread or moneyline when we have it. Nothing here is proven; losing segments are listed with the rest.

• Nothing is proven. No blowout split cleared the out-of-sample bar. Treat every number below as history still building.
• NFL, directionally what you would expect, but thin. In the 9 of 32 games that ended as 17-point blowouts, unders hit 52.4% (n=246) and overs 47.8% (n=255). Rushing-yard unders in those games hit 56.9% (n=65) while overs hit 43.3% (n=67). Starter-volume passing unders hit 55.6% (n=27) in blowouts, versus 41.5% (n=65) in competitive games. Every interval still includes break-even.
• MLB, the other way for batters. A 5-run blowout usually means a lot of offense, so batter unders got worse, not better: all unders in blowouts hit 47.2% (n=519) against 57.2% (n=1,795) in competitive games. Total-bases unders in blowouts hit 43.3% (n=203), a losing segment. Starting-pitcher strikeout unders were a bit better in blowouts (55.2%, n=67) but that sample is small.
• A big pre-game spread did not prove much. Only 2 of 32 NFL games had a sharp spread of 10+ points, and only one of those finished as a 17-point blowout. Legs in those two games hit about 51.5% on both sides (n=132). That is not enough to say a big spread predicts starter unders. We now store the live pre-game spread and the final margin on each graded pick so this test can fill in. NBA, NHL, college basketball and soccer still say history still building — there is no free graded history for them yet.
• Coaching tendencies: unverified. Garbage time, pulled starters, clock-killing runs, empty-net goalie pulls and late soccer subs are the mechanisms we are testing. We did not have a free, citable source for coach-by-coach habits, so we do not name any.

Blowouts vs competitive games

Every graded +EV-eligible leg with a final score, split by whether the game met the final-margin bar.

By sport and stat

Starter-volume props

Passing yards, passing TDs, passing completions, and pitcher strikeouts — the props that usually belong to a starter. Other markets stay in the unknown-role rows, not labelled as role players.

Pre-game spread risk (NFL)

A game is in the risk zone when the sharp home spread is 10+ points. Only two games in this sample qualified, so these rows are small. MLB, NBA and other sports are not in this table yet.

Form, defense and pace: do they add an edge?

Each rule below picks a side using only what was known before kickoff (last 5 and last 10 games, season average, opponent defense tier, game total, home or away) and is graded on the 1,723 NFL lines the apps actually offered across 32 games (2026-09-18 to 2026-09-29). The first part of the sample trains, the later part tests. A rule is adopted into the edge only if it clears the same bar as everything else here: test n of at least 100, test 95% lower bound above break-even (54.2%), and train hit rate above break-even.

84 rule and stat combinations were tried; 1 cleared the bar, and 0 still clear it once the bar is widened for that many tries. None is adopted. Two weeks of games is a short window, so a lone pass among this many tries is what chance alone produces. We will re-run it as more graded games arrive.

Real over vs under hit rates on offered lines

Rules tested

All 84 rows tested are in the data file behind this page. Defense tier uses each defense's last 8 games; there is no free, terms-allowed source for who guards whom, so primary-defender matchups are not tested.

Closing-price blend: weighted mean vs median

Does taking the median of each source's fair price beat the weighted mean we use? Lower Brier score is better; the interval is a bootstrap by game.

On the lines you can actually play the median is no better, so the live edge keeps the weighted mean. What is adopted is the safety net: a source is ignored if its fair price is outside 2%–98% or more than 15 points from the median of the others (changed 4 of 1873 offered lines, no measurable cost).

By agreement size

By line type

De-vig method: power

Same picks, different way of removing the vig.

De-vig method: worst case

The most conservative of the three methods.

Does the fair % match reality?

Both sides of every line included. If fair % is honest, the hit rate in each row should sit near it.

Method, limits and credit

• Data: Graded player-prop sample. MLB: 173 games, 19,223 legs, 2026-09-15 to 2026-09-27; NFL: 32 games, 4,904 legs, 2026-09-18 to 2026-09-29. EdgeLeg re-computed every number here; the underlying sample is not redistributed. Required attributions are on Data credits.
• Method: closing sharp prices from several books and exchanges are de-vigged and blended; the favored side's fair % is compared with each app's break-even for its default entry (PrizePicks 6-pick Flex, Underdog 2-pick Standard, Dabble 6-pick All-In). A leg is "flagged" at 2+ pts of edge.
• Out-of-sample: Chronological per sport: first half of games = train, second half = test (out-of-sample). A segment is proven only if: Test n >= 100 legs AND test hit-rate 95% Wilson lower bound above break-even AND train hit-rate above break-even.
• Lines: Closing lines (last price at or before kickoff), not the line a user saw earlier. Whole-number lines excluded. PrizePicks standard lines only; Underdog per-pick multipliers applied to break-even. Underdog picks with a payout multiplier other than 1x are excluded from the backtest because their price cannot be tested here, and Dabble is limited to near-even prices.
• Multiple comparisons: 295 segments were tested. By chance alone a few would look good, which is why the bar uses the lower confidence bound on held-out data.
• Not covered: esports (no stat history or sharp prices), NBA, NHL and college basketball history (no free graded source yet), and apps without a verified payout table (fair % only, no EV claim).
• Blowouts: final-margin cuts and pre-game risk zones are listed in the Blowouts section. Live grading now stores the final margin and the pre-game sharp spread on each graded pick so sports without history can fill in. Coaching-style claims are marked unverified unless a free, citable source supports them — none did for this sample.
• Live results start from the day grading was switched on. Results come from final box scores; a pick whose player cannot be matched exactly is voided, never guessed.

EdgeLeg is analytics only. It takes no bets and past results do not guarantee future results. How EdgeLeg compares with other tools