Risk of Ruin in Sports Betting: Monte Carlo Simulation and Bankroll Survival

1. Introduction: The Fundamental Law of Gambler's Ruin

Three different events need separate definitions: absolute ruin reaches zero or cannot fund the next required stake; a fixed lower barrier is measured from initial wealth; drawdown is measured from the running peak. This article keeps the classical zero-barrier model in Section 2, but the simulation table measures whether wealth touches half its initial value during 2,000 bets. That event is neither zero wealth nor a 50% fall from the running peak.

In quantitative probability theory and stochastic wealth modeling, the concept of Risk of Ruin (RoR) represents the supreme existential constraint of any wagering enterprise. Formally established by Christian Huygens (1657) and Pierre Rémond de Montmort (1708) in the classical gambler's ruin problem, Risk of Ruin is defined as the mathematical probability that an investor's capital trajectory hits an absorbing lower barrier—either absolute bankruptcy ($W=0$) or an unrecoverable drawdown threshold ($W\le W_{\text{ruin}}$)—prior to reaching an arbitrary upper wealth target.

Among recreational sports bettors, an erroneous intuition prevails: many believe that possessing a positive mathematical edge ($\text{EV}>0$) guarantees eventual financial triumph. In reality, a positive expectation is merely a necessary condition for long-term growth, never a sufficient one. Due to standard binomial variance, path-dependency, and random clustering of negative outcomes, an edge-bearing bettor with an aggressive or uncalibrated staking policy can go broke long before the Law of Large Numbers manifests.

While elementary textbook formulas provide closed-form solutions for simplified, constant-stake scenarios, real-world sports trading involves dynamic bet sizing, variable odds (from 1.25 heavy favorites to 8.50 longshots), non-stationary edges, and model estimation error. To quantify true survival probabilities under realistic trading conditions, quantitative syndicates deploy Monte Carlo simulation engines.

This technical paper establishes the mathematical foundations of Risk of Ruin, compares analytical formulas against empirical simulation, demonstrates the dramatic tail-risk reduction achieved via Fractional Kelly staking, and provides a vectorised Python simulation of 10,000 bankroll paths (full script: /papers/risk_of_ruin_simulation.py).

2. Analytical Derivations: The Classical Gambler's Ruin Formula

The classical model uses a fixed monetary stake $U$ on each independent even-money bet, with win probability $p$ and loss probability $q=1-p$. Initial money is $B_0$, equivalent to $N=B_0/U$ stake units. Let $T>N$ be the upper target in those same stake units. The walk starts at $N$ units and stops at zero or $T$. Thus state $k$ below counts stake units, not currency.

The Difference Equation and Closed-Form Solution

Let $R(k)$ denote the probability of ruin given an initial capital of $k$ units. By conditioning on the first wager, $R(k)$ satisfies the homogeneous second-order linear recurrence relation:

Subject to the boundary conditions $R(0)=1$ (ruin has occurred) and $R(T)=0$ (the target has been achieved). The characteristic equation $p\cdot r^2-r+q=0$ factors into $(r-1)(p\cdot r-q)=0$, yielding roots $r_1=1$ and $r_2=q/p$.

For an infinite wealth target ($T\to \infty$) and a favorable game ($p>q$):

Where $N=B_0/U$ represents the bankroll expressed in total betting units.

Uneven Odds and The General Continuous Brownian Approximation

For unequal odds, a fixed-cash-stake walk can be approximated by Brownian motion with positive monetary drift $\mu$ and monetary variance per bet ${\sigma }^2$. If $R$ is profit per unit stake, then $\mu =UE[R]$ and ${\sigma }^2=U^2\mathrm{Var}(R)$. Consequently the equivalent unit-stake formula uses bankroll depth $N=B_0/U$. This is a continuous, unlimited-horizon, zero-barrier approximation with no upper stopping target; it is not the finite proportional simulation below.

At fixed positive drift and initial wealth, this approximation gives a larger barrier-hit probability when variance rises. Its cash units, stake size and stopping event must stay consistent when comparing formulas.

3. Discrete vs. Proportional Staking: The Kelly Miracle

With ideal proportional staking, $0<f<1$ and finite decimal odds, each wealth multiplier is positive. Wealth therefore stays positive after every finite number of bets. This does not imply a positive limiting wealth: the long-run behaviour depends on the exact expected log growth and assumptions. Minimum stakes, estimation errors and fixed cash withdrawals are additional features that this proportional model does not include.

For a single proportional bet with return $X$, $\mathrm{SD}(fX)=f\mathrm{SD}(X)$ and $\mathrm{Var}(fX)=f^2\mathrm{Var}(X)$. Halving the fraction halves this standard deviation and quarters this variance. These identities do not specify a finite-horizon drawdown probability. For example, wealth $100\to 200\to 120$ has a 40% fall from its peak without touching half the initial bank.

4. The Monte Carlo Simulation Architecture

The table uses 10,000 independent paths, 2,000 bets, initial wealth $10,000 and seed 20261016. Each scenario fixes win probability, odds and staking rule. A path is counted once when its minimum wealth is at most $5,000. The generator continues after that threshold; the threshold is an observation, not a stop instruction. For flat stakes the public script uses an unrestricted additive walk; its terminal values do not model an inability to fund later stakes.

5. Empirical Simulation Results across Edge, Odds, and Fractions

The table reports the fixed starting-bank barrier event defined above, not peak drawdown or zero-bank ruin. The scenario generator and saved results are available as /papers/risk_of_ruin_simulation.py, /papers/risk_of_ruin_results.json.

The simulation shows the cost of aggressive sizing: Full Kelly led to a hit of half the initial bank in 39.5% to 46.8% of paths within 2,000 bets. Half Kelly cut that to 7.7% (3% edge) and 11.4% (5% edge), a four- to fivefold reduction, while the median final bankroll fell by 23% and 43% respectively. Quarter Kelly kept the risk below 0.5%. Flat staking at 2% of the starting bankroll had a 17.9% risk, because its stake does not shrink when the bankroll does.

6. Python Implementation: 10,000-Path Monte Carlo Engine

Below is a compact NumPy version of the simulator for one proportional-staking scenario. The full script that produced the table above is /papers/risk_of_ruin_simulation.py.

# monte_carlo_ror.py
import numpy as np

def run_half_bank_simulation(
    initial_bankroll=10000.0,
    odds=2.00,
    true_prob=0.53, # +6% EV
    stake_fraction=0.03, # 3% proportional
    num_bets=2000,
    num_simulations=10000,
    barrier_fraction=0.50
):
    # Fixed starting-bank barrier simulation
    lower_barrier = initial_bankroll * barrier_fraction
    b = odds - 1.0
    
    # Pre-generate uniform random variates for all paths (num_bets, num_simulations)
    np.random.seed(42)
    random_matrix = np.random.rand(num_bets, num_simulations)
    win_matrix = random_matrix < true_prob
    
    # Multiplier per step: (1 + f * b) if win else (1 - f)
    step_multipliers = np.where(win_matrix, 1.0 + stake_fraction * b, 1.0 - stake_fraction)
    
    # Cumulative wealth trajectory via cumulative product
    wealth_trajectories = initial_bankroll * np.cumprod(step_multipliers, axis=0)
    
    # Calculate minimum wealth to detect a fixed starting-bank barrier hit
    min_wealth_per_path = np.min(wealth_trajectories, axis=0)
    threshold_hits = min_wealth_per_path <= lower_barrier
    threshold_hit = np.mean(threshold_hits) * 100.0
    
    final_wealths = wealth_trajectories[-1, :]
    
    return {
        'threshold_hit_pct': threshold_hit,
        'median_final_wealth': np.median(final_wealths),
        'p5_minimum_wealth': np.percentile(min_wealth_per_path, 5),
        'p95_best_wealth': np.percentile(final_wealths, 95)
    }

if __name__ == '__main__':
    res = run_half_bank_simulation(stake_fraction=0.03)
    print(f"P(hit half initial bank within horizon): {res['threshold_hit_pct']:.2f}%")
    print(f"Median Final Bankroll: ${res['median_final_wealth']:,.2f}")
    print(f"5th Percentile Lowest Dip: ${res['p5_minimum_wealth']:,.2f}")

7. Institutional Risk Guidelines: The 1% Survival Threshold

A planning limit such as 1% must name its event and horizon. In the listed scenarios only Quarter Kelly kept the chance of touching half the initial bank below 1% during 2,000 bets. This does not bound the chance of a 50% fall from a later peak or certify a real strategy. Different odds, probabilities and execution constraints require a new calculation.

1. Fractional Cap: Treat Half Kelly ($\gamma =0.50$) as the upper limit and use Quarter Kelly ($\gamma \le 0.25$) when the edge estimate is uncertain or the market is high-variance (e.g., correct score).
2. Dynamic Re-baselining: Recalculate the absolute dollar unit stake after every session rather than fixing units at starting capital.
3. Stop-Loss Ratchet: If bankroll experiences a 25% drawdown, automatically reduce the Kelly fraction by 50% until capital crosses the previous high-water mark.

The Psychology of Asymmetric Drawdowns

Recovery arithmetic is separate from event probability: a loss of 40% requires a 66.7% gain to recover, and a loss of 50% requires a 100% gain. Neither identity determines how often a path reaches a lower barrier or a drawdown from a running peak. Reduced stakes can lower those risks in a specified model; they do not make recovery certain.

8. Bankroll Depth in Units

The difference equation in Section 2 shows why bankroll depth matters. When the win probability exceeds the loss probability, the ratio $q/p$ is below one, so the probability of ruin decays exponentially as the bankroll measured in betting units grows. Raising the unit stake lowers that depth linearly and raises the Risk of Ruin exponentially: at even money with $p=0.525$, a bankroll of 20 units has a 13.5% chance of ruin on an unlimited horizon, while 40 units bring it down to 1.8%.