Backtest Explanation
The NHL model results are built to answer a live-test question: if the Three Check Bets NHL process had posted every qualified moneyline and enriched team-total play in 2024 and 2025 at $100 per bet, what would the record, ROI, and closing-line value have been?
Walk-Forward Method
The NHL package uses strict chronological walk-forward testing. The model trains only on information available before the test window, scores future games, compares the model price to the available market price, and grades only plays that pass the model and market filters.
This matters because hockey is noisy. A process can look good after it already knows goalie form, team shooting regression, and market movement. The walk-forward structure keeps the test closer to the decisions that would have existed before puck drop.
Three Check Process
The model must estimate a fair price or team-total probability that is better than the sportsbook number.
The bet must have enough line-shopping value, enough sportsbook coverage, and a price that still makes sense at the decision point.
The current moneyline lane uses goalie and schedule guardrails; the next live build must log confirmed goalie status at bet time.
Moneyline Model
The moneyline model estimates full-game win probability using MoneyPuck team metrics, official NHL results, rest and schedule context, goalie starter-proxy data, and historical h2h market prices. The model is not trying to pick only winners. It is trying to find prices where the fair probability is better than the market.
Team Totals
The enriched team-total lane models an individual team's scoring market, not the full-game total. The current candidate is home-team unders only. It uses a market-residual model with context features for selected-team offense, opponent defense, goalie workload/form, rest, congestion, special teams, and matchup aggregates.
CLV
Closing-line value measures whether the final market moved toward the model's bet. For NHL moneylines and team totals here, CLV is shown in decimal-odds terms rather than spread points or total points. Positive CLV does not guarantee one bet wins, but across a large sample it is a useful sign that the market later agreed the price had value.
The combined NHL live-test card had 75.85% positive CLV across 621 bets, with average decimal CLV of +0.0624.
Limitations
The NHL model is promising enough to structure a 2026-27 live test, but it should not be framed as a finished paid-customer proof yet.