HermesLinesThe record
Every result here was graded against the lines sportsbooks actually posted, using only what was known before the game. Break-even at standard −110 odds is 52.4%.
NFL, 2026 season
| Game day | Record | Won | Units | ROI |
|---|---|---|---|---|
| Thu Sep 24 · Wk 3 | 8–11 | 42.1% | -3.46 | -18.2% |
| Sun Sep 27 · Wk 3 | 66–48 | 57.9% | +9.18 | +8.0% |
| Mon Sep 28 · Wk 3 | 7–3 | 70.0% | +2.73 | +27.3% |
| Thu Oct 1 · Wk 4 | 5–4 | 55.6% | +0.24 | +2.6% |
| Sun Oct 4 · Wk 4 | 110–74 | 59.8% | +21.56 | +11.7% |
| Season so far | 196–140 | 58.3% | +30.23 | +9.0% |
Updated automatically after every game; games not yet played aren't counted. Every posted pick, confidence 5 and under included: 267–211, +2.6% ROI.
2025 season, every game
All 272 regular-season games, with lines taken about 30 minutes before kickoff. The rule behind the board: take the under when our model gives it at least a 55% chance, at the highest line across the major books. Unders won 52.0% of 2025 props against the 50.3% their prices implied, and the best of several lines turns that into an edge.
| Test | Picks | Won | ROI |
|---|---|---|---|
| Line-shopped unders, ML model, 2025 season | 3,930 | 57.1% | +3.9% |
The rule was designed on the 2025 season it is scored on, so the 2026 record above is its real test.
Receiving overs In testing
The over on a player's receptions or receiving yards when our model gives it at least a 55% chance and confidence 6 or higher, at the lowest line across the major books. On the board since 4 October 2026 as a live test. Tested on past seasons it lost money (2024: −2.5%), so treat this record as early.
| Test | Picks | Record | Won | ROI |
|---|---|---|---|---|
| 2026 season so far, graded at the posted line and book | 81 | 56–25 | 69.1% | +26.5% |
Tight-end overs
A separate system: the over on a tight end's receptions or receiving yards when our model gives it at least a 55% chance, at the lowest line across the major books. It was found in the 2024 and 2025 seasons, then tested once, unchanged, on the 2023 season, which it had never seen, with lines taken about 30 minutes before kickoff.
| Test | Picks | Won | ROI |
|---|---|---|---|
| 2026 season so far, graded at the posted line and book | 27 | 63.0% | +14.3% |
| Tight-end overs, 2023 season | 376 | 66.0% | +19.0% |
NBA, 2025–26 season
The rule: take the under when our projection is 40–70% of the line and the under pays at least −143. It was designed on 30 nights of real-line data, then tested once, unchanged, on the other 134 nights of the 2025–26 regular season, which it had never seen. NBA picks return in October.
| Test | Picks | Won | ROI |
|---|---|---|---|
| Under rule, 2025-26 season | 1,965 | 55.0% | +3.8% |
Premier League, 2026–27 season
Free home, draw and away forecasts for every match, scored against the predictions Opta publishes for the same games. Log loss measures how much probability a forecast put on what happened; lower is better. These are forecasts, not betting picks.
Every season since Opta started publishing
Each season was forecast by a model trained only on the seasons before it. The bookmaker column is the average odds across major books with their margin removed.
| Season | Matches | Ours | Opta | Bookmakers | Closer |
|---|---|---|---|---|---|
| 2020–21 | 380 | 1.018 | 1.044 | 1.004 | Us |
| 2021–22 | 380 | 0.952 | 0.954 | 0.936 | Us |
| 2022–23 | 380 | 0.977 | 1.009 | 0.966 | Us |
| 2023–24 | 380 | 0.921 | 0.938 | 0.909 | Us |
| 2024–25 | 380 | 0.965 | 0.970 | 0.971 | Us |
| 2025–26 | 380 | 1.026 | 1.034 | 1.015 | Us |
| 2026–27, so far | 50 | 1.041 | 1.063 | 1.045 | Us |
| All | 2,330 | 0.978 | 0.993 | 0.968 | Us |
We were closer than Opta on 54% of these matches. The bookmakers' odds beat both of us over the whole test, which is why these forecasts are free and not sold as picks. See this week's forecasts.
On 29 September 2026 the model moved to ratings built on Opta's expected goals, blended with a smaller share of the earlier stats model. The table above is that model, and its settings were chosen on 2017–20 only, before any season shown here. The head-to-head record at the top counts every forecast as it was locked, so it includes the earlier model's.
How we grade
A result counts only if the pick was made before the game with data available at the time, against a line a sportsbook actually posted. Pushes are left out. ROI is profit per unit staked at the posted price.
NFL projections pool each player's per-game production over last season and this one, weight recent games, and adjust for opponent defense, the Vegas spread and total, weather and the injury report. Confidence (1–9) is how far our probability sits above the probability the book's price implies.