About JP+
JP+ is an independent college football power ratings model. It rates every FBS team using opponent-adjusted play-by-play efficiency data, then converts those ratings into projected game spreads, win totals, and betting edges.
What Makes JP+ Different
Most public ratings (AP Poll, CFP Rankings, even some computer models) are built on outcomes: wins, losses, and point margins. JP+ ignores outcomes and measures how teams actually play on each snap. Two teams can both win 35-14, but one did it with 55% Success Rate against a top-20 defense while the other did it with 38% Success Rate against a bottom-20 defense and got lucky with three turnovers. JP+ sees the difference.
The model is also market-blind. Vegas lines are never used as inputs or training targets. When JP+ disagrees with Vegas, that disagreement is the signal — not an error to correct.
How the Model Works
Efficiency Foundation Model (EFM)
The core engine is a Ridge Regression trained on play-by-play data from every FBS game. Each team gets an opponent-adjusted offensive and defensive rating built from two inputs:
- Success Rate — Did the play gain enough yards to stay on schedule?
- Isolated Points Per Play — How explosive is each play, measured in expected points added, isolated from field position?
Special Teams
A separate model rates field goal accuracy, punt net yardage, and kickoff returns. Special teams ratings feed directly into the spread prediction without a cap.
Preseason Priors
Before the season starts, each team's prior-year offensive, defensive, and special teams ratings are carried forward with mean reversion, then adjusted for coaching changes, recruiting talent, and transfer portal activity. As the season progresses and real play-by-play data accumulates, the priors are gradually replaced by in-season efficiency data.
Walk-Forward Design
Every prediction is made using only data available before that game. The model never sees future games when making a prediction — the same constraint you face when placing a bet. This is validated across the full 2022-2025 backtest with strict chronological guards.
Data Freshness
During the season, ratings and picks are updated each Sunday morning after play-by-play data becomes available from the previous day's games. Opening lines are captured when sportsbooks first post them (typically Sunday morning). During the preseason, ratings are static projections based on prior-year performance, recruiting, and transfer portal activity.
How to Use JP+
- 1
Check the Power Ratings
See where every FBS team stands on offense, defense, and special teams. The gap between two teams' ratings approximates the projected point spread — before home field is applied.
- 2
Browse the Picks Page
Weekly spread and total recommendations where JP+ disagrees meaningfully with Vegas opening lines. Filtered for the highest-confidence situations supported by governed historical validation.
- 3
Dig Into Team Stories
Click any team on the Resume Ratings page for a game-by-game season breakdown — scores, JP+ projections, and Vegas lines side by side. Great for understanding why a team is rated where it is.
Validated Performance (2022-2025)
Walk-forward backtest across 4 seasons (3,481 regular-season games, weeks 1-15). No postseason — opt-outs, coaching changes, and motivation variance make bowls a different sport. The model predicts each game using only prior data, then its line is compared to Vegas to determine if the pick covered.
| Market | Record | MAE | ATS | ROI |
|---|---|---|---|---|
| Blind holdout · tuned on 3 seasons, tested on 1 unseen season | ||||
| Spread PicksSpread Picks — leave-one-season-out | 636 bets | — | 61.0% | — |
| Game O/U PicksGame Totals — leave-one-season-out | 151-91-2 | — | 62.4% | +20.9% |
| Raw model diagnostics · walk-forward game predictions | ||||
| All SpreadsSpreads — all games (wk 1-15) | 3,481 games | 12.85 | 52.2% | — |
| 5+ Edge (Full Szn)Spreads — 5+ pt edge (wk 1-15) | 742-578 | — | 56.2% | +8.0% |
| 5+ Edge (Core)Spreads — 5+ pt edge (wk 4-15) | 515-372 | — | 58.1% | +11.9% |
| In-sample production · filters tuned on these same seasons | ||||
| Spread PicksSpreads — production bet list (2022-2025) | 385-232-6 | — | 62.4% | +21.0% |
| Game O/U PicksGame Totals — production bet list (2022-2025) | 86-46-0P | — | 65.2% | +26.8% |
| Team TotalsTeam Totals — production bet list (2022-2025) | 102-62 | — | 62.2% | +11.0% |
| Win TotalsWin Totals — Production | 84-47-4 | — | 64.1% | +31.6% |
Regular season only (weeks 1-15, no postseason). All ATS vs opening lines. MAE = Mean Absolute Error (average miss in points). Win totals: 2023-2025, 20%+ EV with probability margin filter. Break-even at -110 = 52.4%. Blind holdout rows are the honest unseen-data test. In-sample production rows show the historical result of the final pick filters on the seasons used to tune them; the validation details below state the more conservative forward expectation.
The numbers aren't cherry-picked
A common trap in sports analytics: a model looks great on the same data it was built on. To guard against that, JP+ uses blind holdout testing: tune the model on three seasons, then judge it on the fourth season it has never seen. Every season takes a turn as the blind year.
For spread picks, that test landed at 61.0% ATS on unseen data, profitable in all four holdout years.
Game totals are tested the same way, separately from spreads. The blind totals test landed at 62.4% ATS (151-91-2). The public game-total picks have run 65.2% ATS (86-46-0) in the historical record, but JP+ treats ~60.0% ATS as the more realistic forward expectation.
Realistic forward expectations: ~61.0% ATS for spreads and ~60.0% ATS for game totals. Break-even at standard juice is 52.4%.
Signal Flags & Sizing (2022–2025)
Every JP+ pick carries one or more signal flags — structural patterns that explain why the model likes the side. Flags serve two purposes: they label the thesis behind each bet, and when a flag represents an independently validated edge, they increase stake sizing.
Most flags are 1.0× (standard stake). The model's core engine already qualified the pick — the flag adds context, not the conviction. Roster Alignment is in this group: it is an early-season trust cue, not a premium sizing signal. A few flags earn 2.5× sizing when the pattern is strong enough across all four backtest seasons.
| Flag | What It Means | Record | ATS | Size |
|---|---|---|---|---|
| Def Dom | Backed team top-15 FBS at limiting opponent Success Rate | 20-7 | 74.1% | 2.5× |
| Away Conviction | Away pick, 5+ pt edge, weeks 4–15. Market chronically overvalues home field | 111-66 | 62.7% | 2.5× |
| SR Dominant | Backed team top-25% FBS in offensive Success Rate | 31-17 | 64.6% | 1.0× |
| Market Lag | Market anchored to preseason prestige — JP+ sees in-season efficiency the market hasn't priced yet | 45-30 | 60.0% | 1.0× |
When multiple flags land on the same pick, sizing increases to 2.5× (the cap). Most flags are applied to 5+ pt edge picks in core weeks (4–15). Roster Alignment applies only in Weeks 1-2 and is reported on its admitted raw pre-selection validation surface. Records are vs opening lines across all four seasons.
How Picks Are Selected
Not every game where JP+ disagrees with Vegas makes the pick list. Games pass through three filters before surfacing as a recommendation:
- Edge threshold — JP+'s projected spread must disagree with Vegas by at least 5 points. Smaller edges don't reliably overcome the vig.
- Sigma qualification — A game-specific uncertainty model estimates how confident JP+ is in its own prediction. Games where the model's uncertainty is high relative to its edge are filtered out — even a large edge means little if the model isn't confident.
- Expected value — Each remaining game is priced against the vig to confirm positive expected value. If a pick doesn't clear EV after accounting for the -110 juice, it's dropped.
Three structural vetoes remove game profiles that historically neutralize the edge regardless of magnitude. Vetoed games are shadow-tracked but not staked.
Six Betting Markets
Spreads
Game-by-game spread picks from the Selection Engine (highest-conviction bets).
Game Totals
Over/Under picks on combined game scores. Filtered to extremes where the model has informational advantage.
Win Totals
Preseason season-win projections via Monte Carlo simulation. Compared to Vegas lines for EV-positive over/under bets.
Conference Futures
Simulated conference championship probabilities. Identifies mispriced conference title markets across all 10 FBS conferences.
National Championship
CFP appearance and national title probabilities from Monte Carlo simulation. Compared to sportsbook futures odds.
Dog ML Picks
Home underdogs where JP+'s implied win probability exceeds the market's by 15%+. Best used for line shopping across multiple books.
How to Read the Ratings
Each team's JP+ Overall rating is measured in points versus FBS average. A team rated +20 is roughly 20 points better than average per game. The gap between two teams' ratings approximates the projected point spread before home field is applied.
The overall rating is the sum of three components:
- Offense — Higher is better. A +10 offense generates ~10 more points per game than average.
- Defense — Higher is better (fewer points allowed). A +10 defense allows ~10 fewer points per game than average.
- Special Teams — Field goals, punting, and kickoff returns. Typically ranges from -2.5 to +2.5.
JP+ vs Other Models
vs SP+ (ESPN): Both use play-by-play efficiency and publish projected lines vs Vegas. SP+ optimizes for prediction accuracy (minimizing spread error) and positions itself as a measure of team quality rather than a betting model. JP+ optimizes for ATS edge — a fundamentally different objective that prioritizes finding market mispricings over minimizing MAE.
vs FPI (ESPN): FPI blends efficiency with win probability and recruiting. It's designed for editorial rankings, not spread prediction. JP+ is purpose-built for betting edge.
vs Sagarin / Massey: These are margin-of-victory models that use final scores. JP+ uses play-level data, which is more predictive because it separates process from luck.