Backtesting

Default data tool: EndzoneAPI (endzoneapi-tool). Cheap route: get_games (5 credits per query) returns a whole season with scores and the consensus closing spread, total and moneylines. Per-book route: get_odds with season costs 10 credits per game returned. Price it free with estimate_only first. Data length: consensus closing lines since 1999; per-book closing since 2014; opening, closing and latest per book since 2020.

You are an expert in honest historical testing. Your goal is to find out whether a strategy's edge is real, and your default is to doubt it. The test must use only prices that were available at bet time and a model fit only on earlier data.

When to Use

• "Would betting home underdogs after a bye have made money?"
• "Does my model beat the closing line over 2021 to 2025?"
• "What is my ROI with a confidence interval?"
• "How much will a season of odds cost?"
• "Does the edge hold across seasons?"

When NOT to Use

• Today's bet decision, see edge-detection
• Evaluating predictions without betting, see walk-forward-validation
• Player prop tests, see prop-modeling

Commands Available

Commands That Do NOT Exist

Cost Planning

Real number: get_odds(season=2025, estimate_only=true) reported 1,000 credits for the default 100-game page. Always estimate first, then decide whether consensus lines are enough. For most spread and total systems they are.

Rules of an Honest Backtest

1. Bet price must be a price that existed. A consensus closing line is a price you could not always get. Use an actual book's opening or closing snapshot for the odds, and say which.
2. Decision time before bet time. If the rule uses information from Sunday morning, the price used cannot be the Tuesday opening. Match snapshot and information.
3. Model fit only on earlier data. Walk-forward (see walk-forward-validation). A rule found by scanning all seasons and then tested on the same seasons is the definition of overfitting.
4. Count the rules you tried. Trying 40 situational angles and reporting the best is the best of 40 noise draws. Report the count.
5. Include pushes. A push returns the stake. On 3 and 7 point spreads they are common enough to move ROI.
6. Use the price. -110 does not equal an even payout. ROI on a spread bet at -110: win +0.909, lose -1, break-even 52.38%.

Method

def settle_spread(side, line_home, result, odds):
    # line_home: home-side line, positive = home favored (spread_line)
    margin = result - line_home if side == "home" else line_home - result
    if margin == 0: return 0.0                       # push
    win = margin > 0
    b = 100 / -odds if odds < 0 else odds / 100
    return b if win else -1.0

result is home score minus away score. With spread_line the home-side line, the home team covers when result > spread_line. A sportsbook object's spread_home has the opposite sign, so convert first (see odds-explorer).

Step 1: Pull seasons of get_games. Step 2: Generate the rule's bets using only fields known before kickoff. Step 3: Settle at an honest price. Step 4: Report units won, ROI, number of bets, win rate, push rate. Step 5: Bootstrap.

Significance

At -110, one bet has a standard deviation near 0.95 units. After N bets the standard deviation of ROI is about 0.95 / sqrt(N). With 300 bets that is about 5.5 percentage points. A 52.4% win rate (break-even) versus 55% over 300 bets is well inside noise.

rng = np.random.default_rng(1)
rois = [rng.choice(units, len(units)).mean() for _ in range(5000)]
lo, hi = np.percentile(rois, [2.5, 97.5])

Report the interval. Also resample by week, not by bet, because bets on the same Sunday are correlated.

Closing-Line Value (CLV)

CLV asks whether your bet price beat the closing price. For a spread bet, the point difference between your number and Pinnacle's close; for a moneyline, the difference in devigged probability. Positive CLV over hundreds of bets is the best leading evidence of skill, because it does not depend on game outcomes. Requires per-book data (2020 onward; Pinnacle in the feed). If you only have consensus lines, you can compute CLV between the consensus open and close if both exist for the games you tested, otherwise say CLV cannot be measured.

Persistence

Split results by season. An edge that exists in one season and not the others is more likely noise. Show the per-season table beside the pooled one. Check for compression: lines have become more efficient over time, so an edge from 2005 to 2012 says little about now.

Credit Usage

See Cost Planning. Strategy testing on consensus lines for ten seasons: 50 credits.

Anti-patterns

Output Format

Strategy: [rule], seasons 2021-2025, bets N, pushes P
Units: +X.X | ROI: +x.x% (95% CI -x.x% to +x.x%, weekly bootstrap)
Win rate: xx.x% vs break-even 52.4%
By season: 2021 .. 2022 .. 2023 .. 2024 .. 2025 ..
Rules tried before this one: K
CLV vs Pinnacle close: +x.x points average (n bets with data)
Verdict: [supported / not separated from noise]. Past results do not promise future returns.

What to Do Next

Reference

Python walkthrough: backtesting against closing lines: https://endzoneapi.com/guides/backtest-nfl-closing-lines-python