Are AI Sports Predictions Accurate? The Full Validation Behind a 79% Model Win Rate
Backtests show that the Mysports.AI Pro plan (v4.8), on high-confidence games, reached a 79% model win rate in the NBA with an ROI of roughly 28% (across the 9 backtested leagues, the hit rate at the ā„70% confidence threshold is 77.6%). Following the model with flat stakes last year produced +1,240 units of net profit across all leagues. This page explains how the numbers are defined, how the model works, where it falls short and how to verify it, losing games included.
An Honest Definition of the 79% Win Rate: It Does Not Mean Every Bet Wins 79% of the Time
79% is the historical hit rate of high-confidence, model-selected games. It does not mean every bet is profitable, and it must be read together with the selection criteria and the sample definition.
79% Is What You Get After the AI Filters the Slate
Each day the model screens hundreds of games and keeps only the few high-confidence games with the clearest signal. 79% is the historical hit rate of those selected games: the AI does not bet every game, but the games it does act on hit roughly eight times out of ten. If you pick games on your own, this number does not apply.
+1,240 Units Is About Long-Term Accumulation, Not a Single Game
Staking a flat 1 unit per bet (for example $100) and following the model's picks over the long run produced +1,240 units of net profit across all leagues last year. There were losing days and losing weeks along the way; the equity curve reflects the long-term cumulative result.
13,000+ Games Monitored Live, 53 Model Iterations
The database covers the full 2025-26 season, with every line movement in the NBA, MLB and NHL fed into the computation. Through repeated backtesting and retraining, the model has evolved from a single model into the current stack of Gradient Boosted Decision Trees (GBDT), ensemble learning and deep learning, and new features are added continuously.
Real Prediction Screens: Daily Picks and Past Results, Game by Game
This section shows the actual screens of the Mysports.AI prediction page. Every game carries a confidence score, the best available market odds and the expected value (EV). Past predictions are marked game by game as win / loss / push, the hit rate updates in real time, and every result sits in the same table.
Model Backtest Validation Curves: Three Charts for Long-Term Win Rate and Calibration
The three charts test long-term profit and loss, probability calibration and generalisation, shown as the equity curve, the calibration curve and the learning curve. All data is historical backtesting and does not represent future results.
Equity Curve Ā· +1,240 units cumulative
Cumulative net profit at flat stakes, drawdowns included.
Calibration Curve: Confidence vs Actual Hit Rate
The closer the green line tracks the grey ideal line, the more reliable the confidence values (yellow dot = 79%).
Learning Curve: Training vs Validation Hit Rate
The two lines converge, indicating no significant overfitting and a model that generalises to new games.
Why This Model Architecture? Three Edges Traditional Models Cannot Match
The architecture is built around gradient boosted decision trees, combined with deep learning (TabNet / LSTM / GNN), probability calibration and ensemble learning, balancing efficiency on structured data with time-series context. For the full design seethe AI algorithm breakdown; per-league details are in the NBA model breakdown, MLB model breakdown and NHL model breakdown.
Edge 1: Contextual Data Correlation
A traditional linear model treats each variable independently: a star player is injured, so the team's win probability drops. This architecture reads two signals together, "key player out" and "line movement". If a star is confirmed out but the market odds fail to move by a matching amount, the model treats it as a mismatch between market pricing and true strength (a so-called trap line) and adjusts the direction of its prediction accordingly. Interactions between variables like this are exactly what linear models struggle to capture.
Edge 2: Feature Weights That Adapt by League
Predictive performance depends on feature quality and weighting, not on the volume of data. The model adjusts feature importance to the context of each game, much like an experienced analyst asking "which variable matters most tonight": in the NHL it raises the weight of the starting goaltender and lowers home-ice advantage; in the NBA it raises player fatigue (back-to-backs) and recent shooting trends. Each league runs on its own weighting.
Edge 3: 1,500+ Curated Features Plus Calibration Turn "Luck" Into Explainable Variables
The model gathers 1,500+ features, but quality is what matters: overlapping or weakly correlated noise is removed first, the core features with real predictive power are kept, and probability calibration ensures that games predicted at 60% hit close to 60% over time. In MLB, wind direction and the park's run environment are included as well; these variables often decide pitchers' duels settled by one or two runs.
Cross-League Feature Importance Heatmap: How the Model Weights Each League
Numbers = relative attention weight assigned by the model (0ā100, redder is higher). Shown for the NBA, MLB, NHL and NFL only; every other league has its own full weighting.
| Feature | NBA | MLB | NHL | NFL |
|---|---|---|---|---|
|
Key player status
Key Personnel
|
95 | 95 | 92 | 96 |
|
Schedule fatigue
Fatigue / B2B
|
88 | 45 | 65 | 60 |
|
Environment & weather
Environment
|
5 | 82 | 10 | 94 |
|
Money flow
Money Flow
|
75 | 60 | 55 | 85 |
|
Team momentum
Momentum
|
60 | 30 | 50 | 70 |
|
Home advantage
Home Advantage
|
85 | 40 | 65 | 80 |
|
Referee / umpire tendencies
Ref / Umpire
|
70 | 90 | 30 | 60 |
|
Advanced efficiency metrics
Advanced Metrics
|
92 | 95 | 96 | 98 |
Three-Tier Plan Comparison: Basic v4.2 ā Standard v4.5 ā Pro v4.8
A model should be judged on hit rate, max drawdown and the profit/loss ratio together. A smaller drawdown means a smoother bankroll; a 3.50 profit/loss ratio means that for every 1 unit lost, the long-run average returns 3.5 units of profit.
| Key metric | Pro v4.8 | Standard v4.5 | Basic v4.2 |
|---|---|---|---|
| Feature coverage | + full feature set and three special factors | + advanced metrics (Four Factors / wOBA etc.) | Basic data (form, hit rate) |
| Model architecture | Ensemble + deep learning (LSTM / TabNet / GNN) | GBM ensemble | Gradient boosted trees (GBM) |
| Special factor analysis | All three special factors | Partial (1 factor) | ā (not included) |
| Leagues covered | 25 leagues | 21 leagues | 9 major leagues |
| Bet types | All markets | All markets | Moneyline |
| Return on investment (ROI) | 28.0% | 18.5% | 10.5% |
| Max drawdown | -4.8% | -10.2% | -18.5% |
| Average odds | 1.92 | 1.85 | 1.65 |
| Profit/loss ratio | 3.50 | 1.88 | 1.25 |
Basic v4.2 focuses on strength assessment in the major leagues; Standard v4.5 unlocks every market and excels in regional leagues where lines adjust more slowly; Pro v4.8 integrates the full feature set, the three special factors, probability calibration and ensemble learning, and is built to select high-confidence games in high-variance, low-scoring sports such as MLB and the NHL.
By League: Hit Rate Across Confidence Thresholds, Laid Bare
Moneyline backtests for 2022ā2025 show that as the confidence threshold rises, the number of qualifying games usually falls and the hit rate usually rises. Every league is listed in full, including Serie A's slight dip at ā„70%; nothing was removed after the fact.
| League | ā„50% | ā„55% | ā„60% | ā„65% | ā„70% |
|---|---|---|---|---|---|
| NBA | 67.5%(5,439) | 69.9%(4,672) | 72.5%(3,795) | 75.5%(2,912) | 79.1%(2,151) |
| MLB | 57.0%(10,721) | 59.8%(7,067) | 64.5%(3,483) | 70.4%(1,432) | 73.7%(433) |
| NHL | 61.3%(5,980) | 63.7%(4,447) | 67.5%(2,824) | 71.4%(1,618) | 74.7%(823) |
| Premier League | 66.1%(834) | 69.5%(663) | 73.7%(460) | 76.4%(331) | 80.7%(223) |
| La Liga | 67.0%(698) | 70.8%(518) | 77.2%(347) | 81.1%(244) | 83.9%(143) |
| Serie A | 63.8%(743) | 66.9%(537) | 70.4%(361) | 76.5%(230) | 72.7%(128) |
| Bundesliga | 64.9%(589) | 68.2%(453) | 71.5%(337) | 75.8%(231) | 76.9%(156) |
| Ligue 1 | 65.1%(653) | 69.4%(480) | 71.4%(336) | 74.9%(219) | 76.4%(123) |
| MLS | 59.3%(859) | 65.0%(509) | 68.0%(231) | 76.1%(88) | 85.7%(28) |
| Overall | 61.3%(26,516) | 64.5%(19,346) | 69.0%(12,174) | 73.9%(7,305) | 77.6%(4,208) |
ā» Numbers = moneyline hit rate for that league at that confidence threshold (2022ā2025 walk-forward validation); the figure in brackets is the sample size for that cell, matching the downloadable backtest report below.
Definitions: What the Seven Terms on This Page Mean
Seven terms recur throughout this page; here they are defined once. The usage below is identical to every other section of the page.
ConfidenceConfidence
The strength of the model's signal for a single pick, expressed as a percentage. It reflects how clear the conditions and how consistent the features are for that game; it is not a guarantee that the bet wins. Every tiered figure on this page uses it as the threshold: ā„70% defines the Pro plan's high-confidence games, and the higher the threshold, the fewer games qualify and the higher the historical hit rate.
Expected valueExpected Value, EV
The long-run average profit or loss from combining hit probability with odds. The formula is "hit probability Ć decimal odds ā 1": above zero means a positive expected return over time, zero is break-even. Every game on the AI sports picks page shows this value. A high hit rate alone does not make a bet worth taking; at low enough odds the expected value can still be negative.
UnitUnit
A relative measure of stake size: 1 unit = your own base stake (for example 1% of bankroll). Expressing profit and loss in units rather than currency makes performance comparable across bankroll sizes. "+1,240 units" on this page means that staking a flat 1 unit per bet produced a season's net profit of 1,240 base stakes.
Max drawdownMax Drawdown
The largest peak-to-trough decline of the equity curve, expressed as a percentage. Max drawdown and hit rate measure different dimensions of risk; a high-hit-rate model can still draw down. The drawdown of each version is listed on this page (4.8% for v4.8, 18.5% for v4.2) so you can assess downside risk when sizing your bankroll.
Profit/loss ratioProfit Factor
Gross profit divided by gross loss. Above 1 means total profit exceeds total loss, and the higher the number, the better profits cover losses. The profit/loss ratio complements hit rate: hit rate counts the share of correct games, while the profit/loss ratio weighs how much each game won or lost, so a low hit rate at high odds can still produce a positive ratio.
Calibration curveCalibration Curve
A chart plotting the model's predicted probability against the actual hit rate, used to check probability calibration. Ideally every point sits near the diagonal: games predicted at 70% should hit close to 70%. The further the points stray, the larger the calibration error in the confidence values.
Walk-forward validationWalk-forward Validation
A backtesting method that trains only on data available up to a given point, predicts the next period, then rolls forward and repeats. Its purpose is to rule out data leakage, where future data is used to predict the past, so the resulting hit rate is closer to live performance than a single train/test split. Every 2022ā2025 backtest figure on this page was produced this way.
Model Limits and Risk Disclosure: AI Is Not 100% Accurate
The Statistical Limits of Historical Data
The figures on this page use different definitions: the 79% hit rate and 28% ROI are backtest values for high-confidence games on the latest NBA model (v4.8 Pro plan); +1,240 units is last year's live result across all leagues; 13,000+ games is the 2025-26 season's live monitoring; and the downloadable moneyline record covers 2022ā2025. Past performance does not guarantee future results. Random events such as in-game injuries and officiating errors cannot be fully predicted, and the output is a decision aid only.
Even a High-Hit-Rate Model Has Losing Streaks
Every probabilistic model goes through short-term swings, and 79% is no exception. A higher hit rate usually reduces the likelihood and depth of losing streaks (v4.8 max drawdown -4.8%, far better than v4.2's -18.5%). Set a stake cap and judge the model over a long sample.
Bankroll Management and the Nature of This Service
Manage your bankroll with the Kelly criterion or flat staking (the betting calculators can help), and only bet with disposable income. This service provides sports data analysis only; it does not offer any form of betting, and does not accept or place bets on anyone's behalf. Legal and financial risks of betting on third-party platforms rest with the user. See the FAQ.
FAQ and Model Limitations
Why does the Basic plan support only 9 major leagues?
How does the "no profit, free renewal" guarantee work?
This is the Mysports.AI commitment to model performance: if you do not profit during a subscription period, it is extended by one month for free, under these conditions:
- Subscription plan: Pro plan only.
- Bet type: Moneyline markets only.
- League coverage: only the leagues marked "š Crown Picks" on the AI sports picks page.
- Pick-following: the AI-recommended games and units must be followed exactly.
How to apply: if there is no profit when the month closes, contact support to verify your Bet Tracker records; once confirmed, the subscription is extended by one month.
What are features, and do more features make the model more accurate?
What are the model's limitations, and is the AI 100% accurate?
Verify It Yourself: We Publish Every Record, Losing Games Included
This section publishes the complete moneyline prediction record of more than 26,000 games from 2022ā2025, with every winning and losing result kept. The latest daily past predictions and results are on the Mysports.AI AI sports picks page, game by game; the full backtest report (2022ā2025) can be downloaded below for cross-checking.
Historical Record Field Reference
- League : The competition (e.g. NBA, MLB, EPL)
- Date / Time (ET) : Game date and time in US Eastern Time (not local time)
- Home / Away Team : Home and away teams
- Home / Away Score : Final score for both teams
- Pick : AI selection, the specific bet the system recommended
- Result : Prediction outcome, the game graded as a Win or a Loss
- Strategy : Strategy model tied to the subscription tier (e.g. Pro)
- Confidence : Raw AI confidence score; the higher the value, the higher the historical hit rate
- 1 Download the full backtest report Download the complete Excel backtest report (2022ā2025) from this page. It keeps the game-by-game moneyline record, winning and losing games included.
- 2 Sample 30 games at random Draw 30 games at random across different leagues and months, avoiding a selection of winning games only, so the sample represents the whole.
- 3 Check scores and results on a public results site Compare the sampled games against the scores on any public results site, then confirm that the Result column (Win / Loss) in the report matches the actual outcome.
- 4 Log your own bets in Bet Tracker Use Mysports.AI Bet Tracker to log every bet you actually follow, then evaluate the results against your own equity curve.
Backtest Hit Rate by Confidence Threshold (2022ā2025)
Across the NBA / MLB / NHL, the six major football leagues and MLS Ā· figures taken from the downloadable file below, verifiable game by game.
| Confidence threshold | Hit rate | Sample |
|---|---|---|
| ā„50% | 61.3% | 26,516 |
| ā„55% | 64.5% | 19,346 |
| ā„60% | 69.0% | 12,174 |
| ā„65% | 73.9% | 7,305 |
| ā„70% | 77.6% | 4,208 |
ā» "Hit rate" = correct predictions Ć· qualifying games at that threshold (all moneyline). The overall high-confidence (ā„70%) hit rate across the 9 backtested leagues is 77.6%, with the NBA the highest single league (ā„70%: 1,702 / 2,151). Higher confidence means fewer qualifying games but a higher hit rate, the normal trade-off of a calibrated model.
2022ā2025 Ā· covers the NBA / MLB / NHL, the six major football leagues and MLS: hit rate by confidence threshold plus game-by-game detail for your own cross-checking.
Methodology and Sources: Standing on the Shoulders of Public Research
The feature engineering and rating methods are built on publicly verifiable statistical research and data sources, which you can check and cross-reference below:
- Attention mechanism (used by TabNet) : Vaswani et al. (2017), Attention Is All You Need
- Elo rating system : Arpad Elo (1978), strength ratings that adjust dynamically to opponent quality
- Football goal modelling : Dixon & Coles (1997), Modelling Association Football Scores
- Poisson distribution : the probabilistic basis for modelling scores and goal counts
- Player Efficiency Rating (PER) : John Hollinger's composite efficiency metric
- NBA historical data sources : Basketball-Reference Ā· stats.nba.com
Note: the above are the public research and data sources underlying this model's feature engineering and rating methods; the model architecture, weights and implementation are developed in-house by Mysports.AI.
Related research: Mysports.AI Research; our study comparing AI predictions with pre-game odds.
The Full Prediction Record Is Public: The Data Is Ready, Now It's Your Turn to Verify
Now that you know where the 79% comes from, the next step is building your own record: log every bet for free and let the equity curve tell you how accurate the model really is.
Boost your betting analysis with Mysports.AI Pro plan. Unlock today's AI picks, read the data with a more scientific approach and make better-grounded decisions while you watch. Enter promo code "BERTWIN" at checkout for 12% off, and if you have any questions, feel free to contact support at any time.
