US Sports Probability Engine: NFL, NBA, MLB & NHL

🏈🏀⚾🏒 US Sports Probability Engine: NFL, NBA, MLB & NHL

Statistical win probabilities for every game in the four major US leagues. For each game you get win probability, expected score, over/under and spread prices, the most likely scores and fair odds. When a betting line is published, you also get the model's edge versus the market.

Pick the league with one filter: NFL, NBA, MLB, NHL or all. Between them, the four leagues cover the whole year:

League	Season
🏒 NHL	October → June
🏀 NBA	October → June
🏈 NFL	September → February
⚾ MLB	March → October

---

Why this one

• A real model, not tips. Goals and runs use a Poisson or Negative Binomial model. Points use a Gaussian model. Each team gets an attack and a defense rating, adjusted for home advantage, back-to-back fatigue (NBA/NHL) and time decay. Ratings are shrunk toward the league average, so a hot two-week streak doesn't produce silly numbers.
• Overtime handled properly. In the NHL the model prices regulation (home / tie / away) separately, then resolves overtime and shootouts. In MLB it resolves extra innings the same way.
• Proven calibration. The calibration mode runs a walk-forward backtest. Each week the model is refit using only past games, then it predicts the next week. You get accuracy, Brier score, log loss and calibration bins. In a well-calibrated model, games priced at 60% are won about 60% of the time, and you can check it yourself.
• Market comparison. When a line is published, the output shows the no-vig market probability, the bookmaker margin and the model's edge.
• Cheap and fast. Game history is cached between runs, so a typical run only downloads the latest results.

---

Modes

Mode	What you get
predictions (default)	Every game today, tomorrow or both (US Eastern, the leagues' own calendar): win %, expected score, totals ladder, spread, team totals, most likely scores, fair odds, market edge
matchup	Any two teams, home/away or neutral site, with your own total and spread lines
power_rankings	All teams ranked by net rating, with offense/defense per game and P(beat an average team)
team	One team: rating and rank, last 10 results, next 5 games with win probabilities
calibration	Walk-forward backtest: accuracy, Brier score, log loss, calibration bins and model vs market

Every field is optional. Empty fields get sensible examples (today's games, top-ranked teams), and the run tells you what it assumed in defaults_used.

---

Example inputs

Tonight's NHL games:

{ "mode": "predictions", "league": "nhl" }

Tomorrow's NBA games:

{ "mode": "predictions", "league": "nba", "day": "tomorrow" }

A hypothetical NBA Finals game on neutral ground, priced at a 224.5 total:

{ "mode": "matchup", "league": "nba", "home_team": "BOS", "away_team": "Lakers", "neutral_site": true, "total_line": 224.5 }

How good is the model? Backtest the last 90 days of MLB:

{ "mode": "calibration", "league": "mlb", "backtest_days": 90 }

---

Sample output (predictions, NHL)

{
  "league": "NHL",
  "date": "2026-11-03",
  "matchup": "TOR @ BOS",
  "prob_home_win": 0.5712,
  "prob_away_win": 0.4288,
  "prob_regulation": { "home": 0.4471, "tie": 0.2236, "away": 0.3293 },
  "expected_score": { "home": 3.21, "away": 2.84 },
  "expected_total": 6.05,
  "most_likely_scores": [ { "score": "3-2", "prob": 0.0812 } ],
  "totals": [ { "line": 6.5, "over": 0.4213, "under": 0.5787 } ],
  "spread": { "home_line": -1.5, "home_covers": 0.3102, "away_covers": 0.6898 },
  "team_totals": { "home": { "3+": 0.6011 }, "away": { "3+": 0.5402 } },
  "fair_odds": { "home_american": -133, "away_american": 133 },
  "market": { "market_prob_home_no_vig": 0.5904, "edge_home": -0.0192 }
}

(Illustrative values.)

---

🤖 Use with AI agents (MCP)

This actor is available through the Apify MCP server (mcp.apify.com), so Claude, ChatGPT and other agents can call it directly:

• "What are the win probabilities for tonight's NBA games?"
• "Price Yankees vs Red Sox, over 8.5 runs."
• "Which NHL team has the best defense right now?"

Set "summary_only": true for a compact answer that fits an agent's context.

---

Backtest (out of sample)

The model was refit every week using only past games, then it predicted the following week. Parameters were chosen in one season and tested on the next, so the test numbers below are out of sample. Brier skill is the improvement over always predicting the league's home win rate (0% = no skill).

League	Test period	Games	Accuracy	Brier	Brier skill
🏀 NBA	2025-26 season	1,322	67.8%	0.207	+16%
🏈 NFL	2024 season to Oct 2026	676	65.1%	0.225	+10%
⚾ MLB	2025 season	2,422	56.1%	0.243	+2%
🏒 NHL	2025-26 season	1,394	53.9%	0.247	+1%

Across two NBA seasons, games where the model gave 70% or more were won 82% of the time. Totals are well calibrated in the NBA.

What this means in practice:

• NBA and NFL: strong, well-calibrated signal.
• MLB and NHL: small edge only. Single games are close to coin flips in both sports, and the biggest single factor (the starting pitcher or goalie) is not in the model. Treat these numbers as a calibrated baseline and add your own pitcher or goalie view.
• Against the betting market (NFL closing lines), the market is more accurate (Brier 0.213 vs 0.226). Betting every "edge" the model found lost money over the period. Use the market comparison to spot disagreements worth investigating, not as automatic picks.

Run "mode": "calibration" to check the live model yourself at any time.

Methodology and limits

• Inputs: final scores, schedule, home/neutral site and rest days. Player-level statistics are not used, so injuries and lineup news only show up once they affect results. Factor them in yourself when they matter (a star out, a backup goalie or pitcher).
• Time decay: half-life of 50 days (NBA), 120 days (NHL), 150 days (MLB) or 200 days (NFL), so early-season numbers lean on last season and adapt fast. These values were picked in one season and validated on the next.
• NFL: with only 17 games per team, NFL ratings are the noisiest. Key numbers (3, 7) are not modeled explicitly.
• This is a statistical model for analysis and research. It is not betting advice, and no model wins every game.

---

Pricing

You pay per result (one game, team or report). Runs that return nothing, such as an off-season date or a team not found, cost nothing beyond the platform start fee. The reason appears at the top of the run.

---

Made by Jenko Systems: refined public data, modeled.