EV, Variance, and Sample Size: How Many Bets to Prove Real Skill?

1. Introduction: The Epistemological Fallacy of Short-Run Betting Records

In quantitative finance and statistical decision theory, realized investment performance is understood to be a noisy mixture of two distinct components: underlying systematic skill (Expected Value, $+\text{EV}$) and irreducible stochastic noise (Variance). In retail sports betting, however, participants almost universally conflate retrospective profitability with predictive competence. A recreational bettor who achieves a +18% Return on Investment (ROI) across 150 football wagers is heralded as a genius, while a quantitative model with an authenticated +4% edge that suffers a -6% drawdown over 300 bets is discarded as obsolete.

This fundamental cognitive distortion is known in behavioral economics as outcome bias, compounded by the Law of Small Numbers (Tversky & Kahneman, 1971): the mistaken belief that small samples reliably reflect the statistical properties of the parent population. In high-variance wagering environments, short-run financial trajectories provide virtually zero econometric signal regarding long-term mathematical expectancy.

How many bets are required before an analyst can reject the null hypothesis of luck with 95% or 99% statistical confidence? How does wager volatility scale with decimal odds? How does one compute rigorous confidence intervals around empirical yield?

This technical treatise formulates the mathematical mechanics of sports betting sample size determination. We derive the single-bet variance formula, apply the Central Limit Theorem to calculate exact confidence bounds, formalize the required sample size equation ($N^{\ast }$), evaluate the Brier Score for probability calibration, present a lookup table computed from the formula, and provide a full Python statistical significance testing script.

2. The Variance of a Single Sports Bet: The Volatility Multiplier

Consider a binary sports proposition resolved at decimal odds $O$ with true win probability $p$ and loss probability $q=1-p$. Let $R$ denote the random net payout per unit wagered ($S=1$):

Expected Value Formulation

The mathematical expectation of the return $𝔼[R]=\mu$ is:

Variance Formulation

The variance of the single-bet return $\text{Var}(R)={\sigma }^2$ is derived from the second central moment:

Expanding and simplifying the algebraic terms:

3. Central Limit Theorem: Portfolio Return Distribution across N Bets

Let an analyst place a sequence of $N$ independent wagers with identical return distributions $(\mu ,{\sigma }^2)$. By the Lindeberg-Lévy Central Limit Theorem (CLT), as $N\to \infty$, the sample mean return ${\bar{R}}_N=\frac{1}{N}{\sum }_{i=1}^N{R}_i$ converges in distribution to a Gaussian normal distribution:

The standard error of the sample mean return ($\text{SE}$) is:

This yields a fundamental statistical property: the uncertainty of your observed ROI decreases strictly with the square root of sample size ($\sqrt{N}$). To cut the margin of error in half, an analyst must quadruple their sample size.

4. Formulating the Hypothesis Test and Confidence Intervals

To determine whether an empirical betting record reflects genuine predictive edge or random luck, we formulate a one-sided statistical hypothesis test:

• Null Hypothesis ($H_0$): $\mu \le 0$ (The bettor has zero edge; outcomes are governed purely by chance and bookmaker margins).
• Alternative Hypothesis ($H_1$): $\mu >0$ (The bettor possesses statistically authenticated $+\text{EV}$ skill).

The Test Statistic (Z-Score)

Under the null hypothesis, the standardized Z-score is:

For a standard 95% confidence level (significance level $\alpha =0.05$ in a one-tailed test), the critical value is $z_{0.05}=1.645$. For a rigorous 99% confidence level, $z_{0.01}=2.326$.

Two-Sided 95% Confidence Interval for True Yield

If an analyst records an observed yield of ${\bar{R}}_N=+5.0\%$ over $N=400$ bets at odds of 2.00 ($\sigma =1.00$), the 95% confidence interval is:

Because the confidence interval spans across zero, the null hypothesis cannot be rejected. Despite showing a healthy +5% ROI, the bettor cannot statistically distinguish their performance from pure coin-flipping luck.

5. Analytical Sample Size Derivation: The $N^{\ast }$ Formula

To determine how many bets ($N^{\ast }$) are required to reject the null hypothesis of no edge with statistical power $1-\beta$ and significance level $\alpha$, we equate the expected Z-score to the critical boundary:

For standard testing parameters ($\alpha =0.05⟹z_{\alpha }=1.645$, and power $80\%⟹z_{\beta }=0.842$, so $z_{\alpha }+z_{\beta }\approx 2.487$):

This equation proves that the required sample size scales inversely with the square of your edge ($1/{\text{EV}}^2$). Halving your expected edge demands four times as many wagers to achieve statistical validation.

6. Table: Required Sample Size across Odds and Edges

The table below gives the minimum sample size $N^{\ast }$ required to show statistical significance at the 95% confidence level ($p<0.05$, 80% power) across odds and edge levels, computed from the formula above with $p=(1+\text{EV})/O$:

For an edge of 2.5% to 3.0% at odds of 1.95, an analyst needs about 6,500 to 9,400 bets before realized profit alone can rule out luck at this level.

7. Probability Calibration and The Brier Score

Because profit converges so slowly, analysts also check probability calibration, which uses the forecast probability of every bet rather than only its result. The main scoring rule is the Brier Score (Brier, 1950):

Where $p_i$ is the model's estimated win probability and $y_i\in \{0,1\}$ is the actual outcome. The Brier score decomposes into Reliability (Calibration), Resolution, and Uncertainty. A clearly miscalibrated model can show up in the reliability component before the profit record says anything definite.

8. Python Implementation: Statistical Significance Tester

Below is a Python script that evaluates any betting history, calculates exact Z-scores, constructs 95% confidence intervals, and determines the remaining sample size needed to confirm edge.

# betting_significance_tester.py
import numpy as np
from scipy import stats

def evaluate_betting_record(odds_list, outcomes_list, alpha=0.05):
    # Evaluates statistical significance of a betting track record.
    odds = np.array(odds_list, dtype=np.float64)
    outcomes = np.array(outcomes_list, dtype=np.float64) # 1 for win, 0 for loss
    n = len(odds)
    
    # Net returns per unit staked: (O - 1) if win else -1
    returns = np.where(outcomes == 1, odds - 1.0, -1.0)
    
    sample_mean_roi = np.mean(returns)
    sample_std = np.std(returns, ddof=1)
    standard_error = sample_std / np.sqrt(n)
    
    # Z-test against null hypothesis of ROI <= 0
    z_score = sample_mean_roi / standard_error
    p_value_one_tailed = 1.0 - stats.norm.cdf(z_score)
    
    # 95% Confidence Interval for ROI
    z_crit = stats.norm.ppf(1.0 - alpha / 2.0)
    ci_lower = sample_mean_roi - z_crit * standard_error
    ci_upper = sample_mean_roi + z_crit * standard_error
    
    # Required sample size for 80% power at current observed edge
    if sample_mean_roi > 0:
        n_required = int(np.ceil(((1.645 + 0.842) * sample_std / sample_mean_roi)**2))
    else:
        n_required = np.nan
        
    return {
        'num_bets': n,
        'observed_roi_percent': sample_mean_roi * 100.0,
        'z_score': z_score,
        'p_value': p_value_one_tailed,
        'statistically_significant': p_value_one_tailed < alpha,
        'ci_95_roi_percent': (ci_lower * 100.0, ci_upper * 100.0),
        'bets_needed_for_significance': n_required
    }

if __name__ == '__main__':
    # 500 bets at 1.95 odds with 54% win rate (+5.3% ROI)
    np.random.seed(42)
    synthetic_odds = [1.95] * 500
    synthetic_outcomes = np.random.binomial(1, 0.54, 500)
    
    results = evaluate_betting_record(synthetic_odds, synthetic_outcomes)
    for k, v in results.items():
        print(f"{k}: {v}")

9. Survivorship Bias and the "Winning Tipster" Illusion

Suppose 1,000 independent recreational tipsters each place 500 random wagers with 0% edge. By normal distribution properties, approximately 130 of these tipsters will show a +5% ROI or higher purely by chance (13.2% each, as in the callout above). These lucky outliers will launch paid subscription services, write social media posts claiming analytical superiority, and attract naive capital.

Without understanding sample size and variance, investors fall prey to survivorship bias: observing the visible winners while ignoring the roughly 870 others who disappeared. Evaluating tipsters strictly through Closing Line Value (CLV) and sample size requirements is the most reliable protection against survivorship bias.

10. Skewness and Longshots

When applying the Central Limit Theorem to betting returns, the higher moments matter: skewness and kurtosis. For skewed returns, such as longshot wagers with a low win probability and a high payout, convergence to the normal distribution is slower; the Berry-Esseen bound on the approximation error grows with the third moment of the return. For bets at odds above 4.00, the long right tail means that symmetric normal confidence intervals understate the sample size actually needed. Cornish-Fisher expansions can adjust the confidence limits for this skewness.