From ML Model to +EV: Why We Pivoted

The Original Idea

When we started EdgeBets, the plan was straightforward: build a machine learning model that predicts NHL game outcomes better than sportsbook lines, then bet the difference. We used XGBoost trained on team stats, rest patterns, injuries, and recent form.

Here's how the pipeline worked:

1. Scrape daily stats from hockey-reference.com
2. Engineer features: rolling win rates, shot differentials, Corsi, special teams, rest days
3. Train an XGBoost classifier on recent games
4. Compare model probabilities to sportsbook implied probabilities
5. Publish picks where the model found a 4%+ edge

What We Found

After extensive backtesting on the 2025-26 NHL season, the results were clear:

• Model Brier score: 0.234 (how well probabilities match reality)
• Market Brier score: 0.220 (sportsbook implied probabilities)
• The market was more accurate than our model

We tried everything: calibration tuning, feature selection, ensemble methods, situational filters (back-to-backs, home underdogs), totals prediction, and 30+ game-type segments. Nothing produced a statistically significant edge.

The honest conclusion: sportsbook lines are set by teams of quantitative analysts with access to real-time data, injury intel, and millions of dollars of market-making experience. A solo ML model can't consistently beat that.

The Pivot: Math Over Models

Instead of trying to predict outcomes better than the market, we asked a different question: where do sportsbooks disagree with each other?

Sportsbooks set lines independently. When one book offers +150 and four others cluster around +130, the outlier is mispriced relative to the consensus. You don't need to know who will win — you just need to find the best price.

Our new approach:

1. Collect odds from 6 major books (FanDuel, DraftKings, BetMGM, Caesars, BetRivers, ESPN BET)
2. Calculate no-vig consensus fair odds by removing each book's margin and averaging
3. Flag +EV opportunities where any book offers better than fair value
4. Size bets with half-Kelly based on edge magnitude

Why This Works Better

The backtested results speak for themselves:

• NHL: 244 picks, +21.3% ROI (p=0.003)
• NBA: 444 picks, +33.7% ROI (p=0.002)

Both statistically significant. Both confirmed in out-of-sample testing.

The key difference: we're not making predictions anymore. We're doing arbitrage-adjacent math — finding where one book's price is out of line with the market consensus. This is a much more reliable edge than trying to out-predict the market.

Lessons Learned

1. Respect market efficiency. Sportsbook lines contain more information than any individual model can capture.
2. The edge is in the price, not the prediction. Finding the best available odds is more valuable than having a slightly better model.
3. Transparency matters. We could have quietly buried the failed ML model and only shown the +EV results. Instead, we're documenting the whole journey because honest track records build trust.
4. Data quality is everything. We discovered that ~10% of FanDuel's odds in the Action Network API had home/away reversed. Without catching this, our results would have been meaningless.

What's Next

We now cover both NHL and NBA with our +EV approach, running twice daily to catch the best odds before they move. Every pick is tracked publicly on our track record page.

The ML model taught us what doesn't work. The math-based approach shows us what does.

---

Read more about how +EV betting works or see today's picks.