Beating the close is the real test. TRAILING_LATE: 11.6% ROI over 324 bets

Most betting content is graded on wins and losses. A research desk has to meet a stricter standard: was the price we took better than the price the market settled on? The closing line is the market after the sharp money, the lineup cards and the weather updates have all been priced in. Beating it again and again is rare. That rarity is why this shop exists.

What we found

The bar is higher than it looks. A bettor needs a 52.38% win rate just to break even. Many services sell records that sit close to that line, over samples too small to separate skill from variance. A record with no sample size is only a claim.

Our proven signals are graded out of sample, and each result is stated with its n:

- TRAILING_LATE: 11.6% ROI over 324 graded bets - VELO_DROP: 8.9% ROI over 191 graded bets - SPIN_VS_AVG: 7.6% ROI over 234 bets

These are per-signal records. Each stands on its own sample. TRAILING_LATE has the deepest sample, and it is holding out of sample at 11.6% over 324 graded bets. Behind these signals are 9 agents and 428 indexed data points. The NBA predecessor to this engine was backtested over 1070 games.

The trends

The 2026-09-23 slate had 8 games on the board. It shows why the close is so hard to beat: the market moves, and concentrated money is part of what moves it.

The money is concentrated, and it arrives pregame. We tracked $1,181,632 across 12 campaigns on the slate. The top campaign was $170,075 on AZ @ COL Over 12, at 56% one-sided. Next came $122,149 on SD-LAD Over 9 at 77% one-sided, then $116,521 on ATL ML at 90% one-sided. The biggest single ticket on the board was $38,400. Money this large, landing this early, is part of what shapes the closing number.

Lines drift toward the money. Over the last 20h, the LAD side of SD-LAD moved from 54.5c to 62.0c (+7.5). This is the closing line doing its job, absorbing information. A price taken at 54.5c is a different bet from one taken at 62.0c. What we grade is whether we stood on the right side of that move before it happened.

Totals skew over. All 8 of 8 flagged total edges point over. The biggest is AZ @ COL: model fair 12.53 against a line of 11.5 (+1.03 runs). The model and the largest campaign lean the same way on that game, and $170,075 sits on the Over 12.

The model leans underdog in 2 spots, COL ML +413 and CHC ML +506. Our top moneyline read is COL ML +413, with the model at 22.4% and EV +14.9%. Worth reading that against where the money sits. The largest tracked campaign in that game is on the total, not the side.

Pitcher form is split at the extremes. Andrew Abbott (CIN) is trending the wrong way: a 9.38 ERA over his last 5 and 7.94 over his last 3, against a season mark of 4.70. At the other end, Aaron Nola (PHI) has a 1.74 ERA over his last 5 with 6.7 K/9. The market will price both of these runs by first pitch. The question is whether it prices them fully.

What it means

An edge on a single game is a hypothesis, not a result. COL at +413 with EV +14.9% is a model read. It becomes evidence only once hundreds of reads like it are graded against where the market closed. That is why we lead with n. We publish records for signals graded over 324, 191 and 234 bets, not for single-game reads.

The 8 of 8 over skew is itself something under test. When every flagged total points the same way, the working hypothesis has to prove it is a real read and not a calibration tilt. Next we are grading those over edges against the closing total, to see whether the model's numbers are beating the close or just following money that is already moving that way.

We publish probabilities and their track record, not guarantees. No price here is a promise about what a team or pitcher will do. 21+ where legal.

*What would change our mind: TRAILING_LATE's graded record failing to clear the 52.38% breakeven bar as its sample grows past 324.*