How does bootstrapping estimate the standard error of a statistic without relying on distributional assumptions?

Bootstrapping estimates standard errors by treating your observed sample as a proxy for the entire population and repeatedly resampling from it with replacement. The fundamental insight is that the variability across bootstrap samples approximates the variability you would observe if you could repeatedly sample from the true population.\n\nThe Bootstrap Algorithm:\n\n1. Start with original sample of size n\n2. Draw B bootstrap samples, each of size n, with replacement\n3. Compute the statistic of interest (mean, median, regression coefficient) for each bootstrap sample\n4. The standard deviation of those B statistics is the bootstrap standard error\n\n`mermaid\ngraph TD\n A[\"Original Sample
n = 50 observations\"] --> B[\"Bootstrap Sample 1
Draw 50 with replacement\"]\n A --> C[\"Bootstrap Sample 2
Draw 50 with replacement\"]\n A --> D[\"...
Repeat B = 5000 times\"]\n A --> E[\"Bootstrap Sample 5000
Draw 50 with replacement\"]\n B --> F[\"Compute Statistic θ₁\"]\n C --> G[\"Compute Statistic θ₂\"]\n D --> H[\"...\"]\n E --> I[\"Compute Statistic θ₅₀₀₀\"]\n F --> J[\"Bootstrap SE = StdDev(θ₁, θ₂, ..., θ₅₀₀₀)\"]\n G --> J\n H --> J\n I --> J\n`\n\nWorked Example:\n\nAnalyst Priya at Fenwick Capital wants to estimate the standard error of the median monthly return for a hedge fund with only 24 months of data: {1.2%, -0.8%, 3.1%, -1.5%, 2.7%, 0.4%, ...}.\n\nParametric methods for median standard errors require assumptions about the underlying distribution. Instead, Priya runs B = 10,000 bootstrap iterations.\n\nEach iteration draws 24 returns with replacement from the original 24. Some months appear twice, others not at all. She computes the median for each bootstrap sample.\n\nResults across 10,000 bootstrap medians:\n- Mean of bootstrap medians: 0.95%\n- Standard deviation of bootstrap medians: 0.38%\n\nThe bootstrap standard error of the median is 0.38%, enabling her to construct a 95% confidence interval: 0.95% +/- 1.96 x 0.38% = (0.21%, 1.69%).\n\nWhy It Works:\n\nThe plug-in principle states that the empirical distribution function (EDF) of the sample is a consistent estimator of the true distribution. As n grows, the EDF converges to the population CDF, so resampling from the EDF mimics resampling from the population.\n\nLimitations:\n- Requires the sample to be representative (garbage in, garbage out)\n- Performs poorly with very small samples (n < 15)\n- Dependent data (time series) requires block bootstrapping to preserve autocorrelation structure\n- Computationally intensive, though modern hardware handles millions of iterations easily\n\nPractice bootstrap applications in our CFA Quantitative Methods question bank.