Monte Carlo Standard Errors and Run Length

A political poll asking 1,000 people who they'll vote for doesn't perfectly reveal the whole electorate's preference — it has a margin of error. If you wanted to cut that margin of error in half, you might guess you need twice as many respondents. You'd be wrong: you need four times as many, because sampling error shrinks only with the square root of the sample size, not the sample size itself. A Monte Carlo simulation — pricing an option, estimating a portfolio's VaR, projecting expected P&L across thousands of simulated paths — is exactly the same kind of poll, just polling the space of possible future outcomes instead of voters. Every Monte Carlo estimate comes with its own standard error, obeying the identical square-root law, and treating a simulation output as an exact number rather than an estimate with a margin of error is one of the most common quiet mistakes in quantitative finance.

The formula, one symbol at a time

Run $N$ independent simulated paths, each producing a payoff (or any quantity of interest) $X_1, \dots, X_N$. The Monte Carlo estimate is simply their average, $\bar X = \frac{1}{N}\sum_i X_i$, and its standard error is:

$$SE(\bar X) = \frac{s}{\sqrt N}$$

In words: the standard error of a Monte Carlo mean equals the sample standard deviation $s$ of the individual simulated payoffs, divided by the square root of the number of paths — this is the exact same formula as the standard error of any ordinary sample mean, because that's precisely what a Monte Carlo estimate is. It follows immediately that halving $SE$ requires quadrupling $N$: if $SE \propto 1/\sqrt{N}$, then to get $SE$ down to half its current value you need $N$ to go up by a factor of $2^2 = 4$. Getting one more decimal digit of precision (roughly a 10x reduction in $SE$) costs roughly $10^2 = 100$ times more simulated paths.

Watch the running average in the explorer above settle inside a shrinking envelope band as more simulated draws accumulate — that shrinking band, narrowing like $1/\sqrt{n}$, is the standard-error-shrinkage law made visible; the point estimate wobbles early on and steadily tightens, but never stops wobbling entirely, it just wobbles less.

Worked example 1: sizing a simulation to a target precision

A binary payoff pays $100 with probability 0.5 and $0 otherwise. Its standard deviation is $s = \sqrt{0.5\times(100-50)^2 + 0.5\times(0-50)^2} = \sqrt{0.5\times2500+0.5\times2500} = \sqrt{2500} = 50$ dollars. At $N=1{,}000$ simulated paths, the Monte Carlo standard error is $SE = 50/\sqrt{1000} = 50/31.62 \approx 1.58$, i.e. about $1.58. To bring that down to $0.50, solve $50/\sqrt N = 0.50 \Rightarrow \sqrt N = 100 \Rightarrow N=10{,}000$. Going from $N=1{,}000$ to $N=10{,}000$ — a 10x increase in simulated paths — buys a reduction in standard error from $1.58 to $0.50, a factor of about $3.16$ ($=\sqrt{10}$), not $10$: the payoff to more computation genuinely diminishes, exactly as $1/\sqrt N$ predicts.

Worked example 2: a fatter-tailed payoff needs far more paths for the same precision

Now consider an option-like payoff: $0 with probability 0.98, and $2{,}000 with probability 0.02 (same mean as before, $0.98\times0+0.02\times2000=40$... to keep the comparison cleaner, adjust to mean $50: payoff is $0 with probability 0.975 and $2{,}000 with probability 0.025, mean $=0.025\times2000=50$). Its standard deviation: $s=\sqrt{0.975\times(0-50)^2+0.025\times(2000-50)^2} = \sqrt{0.975\times2500+0.025\times3{,}802{,}500} = \sqrt{2437.5+95{,}062.5}=\sqrt{97{,}500}\approx312.25$. At the same $N=1{,}000$ paths, this payoff's Monte Carlo standard error is $312.25/\sqrt{1000}\approx9.87$, i.e. about $9.87 — more than 6x the standard error of the first payoff at the identical path count, purely because this payoff's outcomes are far more spread out (a small chance of a huge $2,000 result versus a much more common $0). To reach the same $0.50 target precision as before would require $N = (312.25/0.50)^2 \approx 390{,}000$ paths — roughly 39x more simulation than the first, well-behaved payoff needed.

What this means in practice

• Always report the achieved standard error, not just the point estimate. A Monte Carlo option price or VaR figure without its standard error is an unverifiable number — the same output could be trustworthy to four decimal places or barely trustworthy to the first digit, and there's no way to tell without computing $s/\sqrt N$.
• Variance reduction techniques exist because $N$ is expensive but $s$ isn't fixed. Control variates, antithetic variates, and importance sampling all work by reducing the effective $s$ in the numerator, which is often far cheaper than brute-force increasing $N$ to shrink the same standard error.
• Tail-risk and VaR estimation need disproportionately more paths. As worked example 2 shows, payoffs with rare, large outcomes have large sample variance relative to their mean, and a VaR or expected-shortfall calculation that only rarely samples the relevant tail event effectively has a much smaller usable sample size than its raw path count suggests.