Become an Actuary › Exams › Flashcards › Exam MAS-I › Statistics & Simulation
Exam MAS-I · ~15-20%

Exam MAS-I: Statistics & Simulation Flashcards

Statistical inference and simulation for CAS Exam MAS-I: estimator properties (bias, variance, MSE, consistency, efficiency), maximum likelihood and the method of moments with closed-form solutions, Fisher information and asymptotic normality, confidence intervals for means and proportions, hypothesis testing with Type I/II errors, power, p-values, and z/t/chi-square tests, plus the inverse-transform method, Monte Carlo run sizing, and the bootstrap — with fully worked numeric examples.

44 cards6 topicsFree · fact-checked · LaTeX math
Tap card or press Space to flip
Answer

Every deck is built into the Willys app

All of these decks, including the full practice problem banks, come built into Willys AI Flashcards & Quizzes for iPhone, iPad & Mac, with FSRS + SM-2 spaced repetition, streaks, and exam-date cram mode. 14-day free trial, then $14.99. To load a deck in the app: Settings → Library → Browse, then pick your exam and deck.

More Exam MAS-I decks:

Statistics & Simulation Practice Survival & Reliability Models Survival & Reliability Models Practice Credibility Credibility Practice Frequency, Severity & Aggregate Models Frequency, Severity & Aggregate Models Practice Markov Chains

← All Exam MAS-I decks

Browse all 44 cards as a list
  1. Point estimation
    Define the **bias** of an estimator $\hat\theta$ of a parameter $\theta$, and state what **unbiased** means.
    Bias is the systematic error: $\text{bias}(\hat\theta)=E[\hat\theta]-\theta$. An estimator is **unbiased** when $\text{bias}(\hat\theta)=0$, i.e. $E[\hat\theta]=\theta$ for every value of $\theta$. Bias measures whether the estimator is centered on the target on average; it says nothing about spread.
  2. Bias, variance & MSE
    State the relationship between **mean squared error**, variance, and bias.
    $\text{MSE}(\hat\theta)=E\!\left[(\hat\theta-\theta)^{2}\right]=\text{Var}(\hat\theta)+\big(\text{bias}(\hat\theta)\big)^{2}$. MSE bundles both error sources: how spread out the estimator is (variance) and how off-center it is (bias squared). For an **unbiased** estimator the bias term vanishes, so $\text{MSE}=\text{Var}(\hat\theta)$.
  3. Point estimation
    Define a **consistent** estimator.
    $\hat\theta_{n}$ is **consistent** for $\theta$ if it converges in probability to $\theta$ as the sample size grows: for every $\varepsilon>0$, $P\!\left(|\hat\theta_{n}-\theta|>\varepsilon\right)\to 0$ as $n\to\infty$. A convenient sufficient condition: if $\text{bias}(\hat\theta_{n})\to 0$ and $\text{Var}(\hat\theta_{n})\to 0$, then $\text{MSE}\to 0$ and $\hat\theta_{n}$ is consistent (mean-square convergence implies convergence in probability).
  4. Point estimation
    What does it mean for one unbiased estimator to be more **efficient** than another, and define **relative efficiency**?
    Among unbiased estimators, the more **efficient** one has the smaller variance. The **relative efficiency** of $\hat\theta_{1}$ to $\hat\theta_{2}$ is $\frac{\text{Var}(\hat\theta_{2})}{\text{Var}(\hat\theta_{1})}$. An estimator whose variance attains the **Cramér–Rao lower bound** $\frac{1}{I(\theta)}$ is called **efficient** (a UMVUE). Smaller variance $\Rightarrow$ tighter, more reliable estimates.
  5. Point estimation
    For an i.i.d. sample, why is the sample mean $\bar X$ unbiased for $\mu$, and what is its variance?
    $E[\bar X]=\frac{1}{n}\sum E[X_{i}]=\frac{1}{n}(n\mu)=\mu$, so $\bar X$ is **unbiased**. $\text{Var}(\bar X)=\frac{1}{n^{2}}\sum \text{Var}(X_{i})=\frac{n\sigma^{2}}{n^{2}}=\frac{\sigma^{2}}{n}$. The variance $\frac{\sigma^{2}}{n}\to 0$ as $n\to\infty$, so $\bar X$ is also **consistent** for $\mu$.
  6. Bias, variance & MSE
    Why is the divisor $n-1$ used in the sample variance $S^{2}=\frac{1}{n-1}\sum(X_{i}-\bar X)^{2}$?
    Dividing by $n-1$ makes $S^{2}$ **unbiased**: $E[S^{2}]=\sigma^{2}$. Using $n$ instead gives $\frac{n-1}{n}\sigma^{2}$, which underestimates $\sigma^{2}$ because the deviations are taken about the estimated mean $\bar X$ rather than the true $\mu$. The loss of one degree of freedom (the constraint $\sum(X_{i}-\bar X)=0$) is exactly corrected by the $n-1$ divisor.
  7. Bias, variance & MSE
    An estimator $\hat\theta$ has $E[\hat\theta]=0.9\theta$ and $\text{Var}(\hat\theta)=0.04\theta^{2}$. Find its bias and MSE at $\theta=10$.
    Bias $=E[\hat\theta]-\theta = 0.9\theta-\theta=-0.1\theta$. At $\theta=10$: bias $=-1$. Variance at $\theta=10$: $0.04(10)^{2}=4$. $\text{MSE}=\text{Var}+\text{bias}^{2}=4+(-1)^{2}=4+1=5$.
  8. Point estimation
    Two unbiased estimators of $\mu$ have variances $\text{Var}(\hat\theta_{1})=\frac{\sigma^{2}}{n}$ and $\text{Var}(\hat\theta_{2})=\frac{2\sigma^{2}}{n+1}$. Which is more efficient for $n=9$, and by how much?
    Compare variances (drop the common $\sigma^{2}$): $\text{Var}(\hat\theta_{1})=\frac{1}{9}\approx 0.1111\,\sigma^{2}$ and $\text{Var}(\hat\theta_{2})=\frac{2}{10}=0.2\,\sigma^{2}$. Since $0.1111<0.2$, $\hat\theta_{1}$ is more efficient. Relative efficiency of $\hat\theta_{1}$ to $\hat\theta_{2}$ $=\frac{\text{Var}(\hat\theta_{2})}{\text{Var}(\hat\theta_{1})}=\frac{0.2}{0.1111}\approx 1.80$, i.e. $\hat\theta_{2}$ has $1.80$ times the variance of $\hat\theta_{1}$ (equivalently $\hat\theta_{1}$'s variance is about $44\%$ smaller).
  9. Bias, variance & MSE
    Consider estimating $\sigma^{2}$ by $T=\frac{1}{n}\sum(X_{i}-\bar X)^{2}$ (divisor $n$). Find its bias for normal data.
    We know $E\!\left[\sum(X_{i}-\bar X)^{2}\right]=(n-1)\sigma^{2}$. So $E[T]=\frac{(n-1)\sigma^{2}}{n}=\left(1-\tfrac{1}{n}\right)\sigma^{2}$. Bias $=E[T]-\sigma^{2}=-\frac{\sigma^{2}}{n}<0$ (it underestimates). The bias $\to 0$ as $n\to\infty$, so $T$ is **asymptotically unbiased** and consistent.
  10. Maximum likelihood
    Write the general **likelihood** $L(\theta)$ and **log-likelihood** $\ell(\theta)$ for an i.i.d. sample, and state the MLE recipe.
    $L(\theta)=\prod_{i=1}^{n} f(x_{i};\theta)$ and $\ell(\theta)=\ln L(\theta)=\sum_{i=1}^{n}\ln f(x_{i};\theta)$. The **MLE** $\hat\theta$ maximizes $L$ (equivalently $\ell$). Solve the score equation $\frac{d}{d\theta}\ell(\theta)=0$ and verify it is a maximum (e.g. $\ell''<0$). Maximizing $\ell$ is easier than $L$ because the product becomes a sum.
  11. Maximum likelihood
    Derive the **MLE of the mean $\theta$** of an exponential distribution $f(x)=\frac{1}{\theta}e^{-x/\theta}$ from an i.i.d. sample.
    $\ell(\theta)=\sum\left(-\ln\theta-\frac{x_{i}}{\theta}\right)=-n\ln\theta-\frac{1}{\theta}\sum x_{i}$. $\ell'(\theta)=-\frac{n}{\theta}+\frac{1}{\theta^{2}}\sum x_{i}=0 \Rightarrow \frac{\sum x_{i}}{\theta^{2}}=\frac{n}{\theta}\Rightarrow \theta=\frac{\sum x_{i}}{n}$. So $\hat\theta=\bar X$, the sample mean.
  12. Maximum likelihood
    Derive the **MLE of the Poisson rate $\lambda$** from an i.i.d. sample $x_{1},\dots,x_{n}$.
    $f(x;\lambda)=\frac{e^{-\lambda}\lambda^{x}}{x!}$, so $\ell(\lambda)=\sum\big(-\lambda + x_{i}\ln\lambda - \ln x_{i}!\big)=-n\lambda+\ln\lambda\sum x_{i}-\sum\ln x_{i}!$. $\ell'(\lambda)=-n+\frac{\sum x_{i}}{\lambda}=0\Rightarrow \lambda=\frac{\sum x_{i}}{n}$. So $\hat\lambda=\bar X$.
  13. Maximum likelihood
    Derive the **MLE of the Bernoulli/binomial success probability $p$** from $n$ i.i.d. trials with $\sum x_{i}$ successes.
    For Bernoulli, $f(x;p)=p^{x}(1-p)^{1-x}$, so $\ell(p)=\left(\sum x_{i}\right)\ln p + \left(n-\sum x_{i}\right)\ln(1-p)$. $\ell'(p)=\frac{\sum x_{i}}{p}-\frac{n-\sum x_{i}}{1-p}=0$. Cross-multiplying: $(1-p)\sum x_{i}=p\,(n-\sum x_{i})\Rightarrow \sum x_{i}=np\Rightarrow \hat p=\frac{\sum x_{i}}{n}=\bar X$.
  14. Maximum likelihood
    A sample of claim counts is $2,0,3,1,4,2$ from a Poisson distribution. Compute the MLE $\hat\lambda$.
    The Poisson MLE is the sample mean $\hat\lambda=\bar X$. $\sum x_{i}=2+0+3+1+4+2=12$, with $n=6$. $\hat\lambda=\frac{12}{6}=2.0$.
  15. Maximum likelihood
    Exponential lifetimes (mean $\theta$) are observed as $4,7,11,2,6$. Compute the MLE $\hat\theta$.
    The exponential MLE for the mean is $\hat\theta=\bar X$. $\sum x_{i}=4+7+11+2+6=30$, with $n=5$. $\hat\theta=\frac{30}{5}=6.0$.
  16. Maximum likelihood
    Define the **Fisher information** $I(\theta)$ and state how it gives the asymptotic variance of the MLE.
    For one observation, $I(\theta)=-E\!\left[\frac{d^{2}}{d\theta^{2}}\ln f(X;\theta)\right]=E\!\left[\left(\frac{d}{d\theta}\ln f(X;\theta)\right)^{2}\right]$. For an i.i.d. sample of size $n$ the total information is $nI(\theta)$, and the MLE is asymptotically normal: $\hat\theta \;\dot\sim\; N\!\left(\theta,\;\frac{1}{nI(\theta)}\right)$. Larger information $\Rightarrow$ smaller asymptotic variance.
  17. Maximum likelihood
    Find the **Fisher information** for the Poisson rate $\lambda$ (per observation) and the asymptotic variance of $\hat\lambda$ for a sample of size $n$.
    $\ln f = -\lambda + x\ln\lambda - \ln x!$, so $\frac{d}{d\lambda}\ln f = -1+\frac{x}{\lambda}$ and $\frac{d^{2}}{d\lambda^{2}}\ln f = -\frac{x}{\lambda^{2}}$. $I(\lambda)=-E\!\left[-\frac{X}{\lambda^{2}}\right]=\frac{E[X]}{\lambda^{2}}=\frac{\lambda}{\lambda^{2}}=\frac{1}{\lambda}$. So $\text{Var}(\hat\lambda)\approx\frac{1}{nI(\lambda)}=\frac{\lambda}{n}$ — matching $\text{Var}(\bar X)$ exactly since $\hat\lambda=\bar X$.
  18. Maximum likelihood
    State the **method of moments** and contrast it with maximum likelihood.
    Equate the lowest sample moments to the corresponding population moments and solve for the parameters. For one parameter, set $\bar X = E[X]=g(\theta)$ and solve for $\hat\theta$; for two parameters use $\bar X$ and $\frac{1}{n}\sum X_{i}^{2}$. Method-of-moments estimators are simple and need no optimization, but are generally **less efficient** than MLEs and need not be unbiased. MLEs are typically preferred for their asymptotic efficiency.
  19. Maximum likelihood
    A gamma distribution has mean $\alpha\theta$ and variance $\alpha\theta^{2}$. A sample has $\bar X=10$ and sample second moment $\frac{1}{n}\sum X_{i}^{2}=140$. Find the method-of-moments $\hat\alpha,\hat\theta$.
    Sample variance (about the mean) $=\frac{1}{n}\sum X_{i}^{2}-\bar X^{2}=140-10^{2}=40$. Match moments: $\alpha\theta=10$ and $\alpha\theta^{2}=40$. Divide: $\frac{\alpha\theta^{2}}{\alpha\theta}=\theta=\frac{40}{10}=4$, so $\hat\theta=4$. Then $\hat\alpha=\frac{10}{\hat\theta}=\frac{10}{4}=2.5$.
  20. Confidence intervals
    State the **large-sample (z) confidence interval** for a population mean $\mu$ when $\sigma$ is known, and for unknown $\sigma$.
    Known $\sigma$: $\bar x \pm z_{\alpha/2}\,\frac{\sigma}{\sqrt n}$, where $z_{\alpha/2}$ is the standard-normal critical value ($z_{0.025}=1.96$ for $95\%$). Unknown $\sigma$ (use the sample $s$): $\bar x \pm t_{\alpha/2,\,n-1}\,\frac{s}{\sqrt n}$, using the $t$ distribution with $n-1$ degrees of freedom. For large $n$ the $t$ critical value $\to z$.
  21. Confidence intervals
    Interpret a **$95\%$ confidence interval** correctly.
    It means the *procedure* captures the true parameter $95\%$ of the time: if we repeated the sampling and interval construction many times, about $95\%$ of the resulting intervals would contain the fixed true $\theta$. It does **not** mean there is a $95\%$ probability that $\theta$ lies in this particular computed interval — once computed, the interval either contains $\theta$ or it doesn't. The randomness is in the interval, not in $\theta$.
  22. Confidence intervals
    A sample of $n=64$ losses has $\bar x=520$ and known $\sigma=160$. Build a $95\%$ confidence interval for the mean loss.
    Standard error $=\frac{\sigma}{\sqrt n}=\frac{160}{\sqrt{64}}=\frac{160}{8}=20$. Margin $=z_{0.025}\cdot 20 = 1.96(20)=39.2$. CI $=520\pm 39.2 = (480.8,\;559.2)$.
  23. Confidence intervals
    A sample of $n=25$ claims has $\bar x=4{,}200$ and sample sd $s=900$. Build a $95\%$ confidence interval for the mean (use $t_{0.025,24}=2.064$).
    Standard error $=\frac{s}{\sqrt n}=\frac{900}{\sqrt{25}}=\frac{900}{5}=180$. Margin $=t_{0.025,24}\cdot 180 = 2.064(180)=371.52$. CI $=4200\pm 371.52 = (3{,}828.48,\;4{,}571.52)$.
  24. Confidence intervals
    State the large-sample (Wald) **confidence interval for a proportion** $p$.
    With $\hat p=\frac{x}{n}$, the standard error is $\sqrt{\frac{\hat p(1-\hat p)}{n}}$, and the interval is $\hat p \pm z_{\alpha/2}\sqrt{\frac{\hat p(1-\hat p)}{n}}$. This normal approximation is reliable when both $n\hat p$ and $n(1-\hat p)$ are at least about $5$ to $10$.
  25. Confidence intervals
    Out of $400$ policyholders, $60$ filed a claim. Build a $95\%$ confidence interval for the claim probability $p$.
    $\hat p=\frac{60}{400}=0.15$. Standard error $=\sqrt{\frac{0.15(0.85)}{400}}=\sqrt{\frac{0.1275}{400}}=\sqrt{0.00031875}\approx 0.017854$. Margin $=1.96(0.017854)\approx 0.0350$. CI $=0.15\pm 0.0350 = (0.1150,\;0.1850)$.
  26. Confidence intervals
    How many observations $n$ are needed so a $95\%$ confidence interval for a mean has margin of error at most $5$, given $\sigma=40$?
    Require $z_{0.025}\frac{\sigma}{\sqrt n}\le 5$, i.e. $1.96\cdot\frac{40}{\sqrt n}\le 5$. Solve: $\sqrt n \ge \frac{1.96(40)}{5}=\frac{78.4}{5}=15.68$, so $n\ge 15.68^{2}=245.86$. Round up: $n=246$.
  27. Hypothesis testing
    Define the **null hypothesis** $H_0$, the **alternative** $H_1$, and the two error types in hypothesis testing.
    $H_0$ is the default claim to be tested; $H_1$ is what we conclude if we reject $H_0$. **Type I error** (probability $\alpha$): rejecting $H_0$ when it is true — a false positive. $\alpha$ is the chosen significance level. **Type II error** (probability $\beta$): failing to reject $H_0$ when $H_1$ is true — a false negative.
  28. Hypothesis testing
    Define the **power** of a test and how it relates to $\beta$.
    Power $=1-\beta=P(\text{reject }H_0 \mid H_1 \text{ true})$ — the probability of correctly detecting a true effect. Power rises with a larger sample size, a larger true effect (further from $H_0$), a larger significance level $\alpha$, and smaller variance. There is a tradeoff: lowering $\alpha$ to reduce Type I error raises $\beta$ and lowers power, all else equal.
  29. Hypothesis testing
    Define the **$p$-value** and state the decision rule.
    The $p$-value is the probability, assuming $H_0$ is true, of observing a test statistic at least as extreme as the one actually observed (in the direction of $H_1$). Decision rule: **reject $H_0$ if $p\le\alpha$**; otherwise fail to reject. A small $p$-value means the data would be unlikely under $H_0$, i.e. evidence against it. The $p$-value is *not* the probability that $H_0$ is true.
  30. Hypothesis testing
    Distinguish a **one-tailed** from a **two-tailed** test and how the critical value changes.
    A **two-tailed** test ($H_1:\mu\neq\mu_0$) splits $\alpha$ across both tails, using $z_{\alpha/2}$ ($1.96$ for $\alpha=0.05$). A **one-tailed** test ($H_1:\mu>\mu_0$ or $\mu<\mu_0$) puts all of $\alpha$ in one tail, using $z_{\alpha}$ ($1.645$ for $\alpha=0.05$). The one-tailed test has a less extreme critical value, so it has more power against the specified direction but cannot detect effects in the other direction.
  31. Hypothesis testing
    Test $H_0:\mu=100$ vs $H_1:\mu\neq 100$ at $\alpha=0.05$ given $n=36$, $\bar x=106$, known $\sigma=18$. State the conclusion.
    Test statistic $z=\frac{\bar x-\mu_0}{\sigma/\sqrt n}=\frac{106-100}{18/\sqrt{36}}=\frac{6}{18/6}=\frac{6}{3}=2.0$. Two-tailed critical value is $z_{0.025}=1.96$. Since $|2.0|>1.96$, **reject $H_0$**. The two-sided $p$-value $=2\,P(Z>2.0)=2(0.0228)=0.0455<0.05$, consistent with rejection.
  32. Hypothesis testing
    Test $H_0:\mu=50$ vs $H_1:\mu>50$ at $\alpha=0.05$ with $n=16$, $\bar x=53.5$, sample $s=8$ (use $t_{0.05,15}=1.753$).
    Test statistic $t=\frac{\bar x-\mu_0}{s/\sqrt n}=\frac{53.5-50}{8/\sqrt{16}}=\frac{3.5}{8/4}=\frac{3.5}{2}=1.75$. One-tailed critical value $t_{0.05,15}=1.753$. Since $1.75<1.753$, **fail to reject $H_0$** (just barely). The evidence for $\mu>50$ is not quite significant at the $5\%$ level.
  33. Hypothesis testing
    Observed counts in 4 categories are $O=(22,18,20,40)$; the model predicts $E=(25,25,25,25)$. Compute the chi-square goodness-of-fit statistic and decide at $\alpha=0.05$ (critical $\chi^{2}_{0.05,3}=7.815$).
    Statistic $\chi^{2}=\sum_{j}\frac{(O_{j}-E_{j})^{2}}{E_{j}}$: $\chi^{2}=\frac{(22-25)^{2}}{25}+\frac{(18-25)^{2}}{25}+\frac{(20-25)^{2}}{25}+\frac{(40-25)^{2}}{25}$ $=\frac{9}{25}+\frac{49}{25}+\frac{25}{25}+\frac{225}{25}=\frac{308}{25}=12.32$. Degrees of freedom $=(\text{cells})-1=4-1=3$. Since $12.32>7.815$, **reject $H_0$** — the model fits poorly (the last category is far over-observed).
  34. Hypothesis testing
    Compute the **power** of the test $H_0:\mu=0$ vs $H_1:\mu>0$, $\alpha=0.05$, $n=25$, known $\sigma=10$, when the true mean is $\mu=4$.
    Reject when $\bar X > 0 + z_{0.05}\frac{\sigma}{\sqrt n}=1.645\cdot\frac{10}{5}=3.29$. Under the true $\mu=4$, $\bar X\sim N\!\left(4,\,\left(\tfrac{10}{5}\right)^{2}=4\right)$, so $\text{sd}(\bar X)=2$. Power $=P(\bar X>3.29\mid\mu=4)=P\!\left(Z>\frac{3.29-4}{2}\right)=P(Z>-0.355)\approx 0.639$. So the power is about $64\%$.
  35. Monte Carlo & bootstrap
    State the **inverse-transform method** for generating a random variate with cdf $F$.
    If $U\sim\text{Unif}(0,1)$, then $X=F^{-1}(U)$ has cdf $F$. Procedure: draw $u$ from $\text{Unif}(0,1)$, then solve $F(x)=u$ for $x$. This works because $P(X\le x)=P(F^{-1}(U)\le x)=P(U\le F(x))=F(x)$. It requires an invertible (or at least solvable) cdf and uses exactly one uniform per variate.
  36. Monte Carlo & bootstrap
    Derive the **inverse-transform formula for an exponential** with rate $\lambda$ (mean $\frac{1}{\lambda}$).
    The cdf is $F(x)=1-e^{-\lambda x}$. Set $F(x)=u$: $1-e^{-\lambda x}=u \Rightarrow e^{-\lambda x}=1-u \Rightarrow x=-\frac{1}{\lambda}\ln(1-u)$. So $X=-\frac{1}{\lambda}\ln(1-U)$. Since $1-U$ is also $\text{Unif}(0,1)$, the equivalent simplification $X=-\frac{1}{\lambda}\ln U$ is often used.
  37. Monte Carlo & bootstrap
    Using $X=-\frac{1}{\lambda}\ln(1-U)$ with $\lambda=0.5$ and the uniform draw $U=0.8$, generate an exponential variate.
    $X=-\frac{1}{0.5}\ln(1-0.8)=-2\ln(0.2)$. $\ln(0.2)\approx -1.609438$, so $X=-2(-1.609438)=3.218876$. $X\approx 3.219$.
  38. Monte Carlo & bootstrap
    Explain how to generate a **discrete** random variate (e.g. a claim count) by inversion using a uniform $U$.
    Build the cumulative probabilities $F(0)\le F(1)\le\cdots$ and partition $[0,1)$ into intervals $\big[F(k-1),F(k)\big)$. Draw $U\sim\text{Unif}(0,1)$ and return the smallest $x$ with $F(x)\ge U$. Equivalently, return $x=k$ when $F(k-1)\le U < F(k)$. Each value's interval width equals its probability, so the output has the target pmf.
  39. Monte Carlo & bootstrap
    A discrete loss has $P(X=0)=0.5$, $P(X=1)=0.3$, $P(X=2)=0.2$. Using $U=0.65$ and the inversion rule, what value is generated?
    Cumulative cutoffs: $F(0)=0.5$, $F(1)=0.8$, $F(2)=1.0$, giving intervals $[0,0.5)\to 0$, $[0.5,0.8)\to 1$, $[0.8,1.0)\to 2$. $U=0.65$ falls in $[0.5,0.8)$, so the generated value is $X=1$.
  40. Monte Carlo & bootstrap
    In **Monte Carlo estimation**, how is a quantity $\mu=E[g(X)]$ estimated and what is the standard error?
    Simulate $N$ independent draws and average: $\hat\mu=\frac{1}{N}\sum_{i=1}^{N} g(X_{i})$. By the LLN $\hat\mu\to\mu$, and the standard error is $\text{SE}=\frac{s}{\sqrt N}$ where $s$ is the sample sd of the $g(X_{i})$ values. Accuracy improves like $\frac{1}{\sqrt N}$, so cutting the error in half requires $4\times$ the runs.
  41. Monte Carlo & bootstrap
    A Monte Carlo simulation of a portfolio loss has per-run standard deviation $s=2{,}000$. How many runs $N$ are needed for a standard error of at most $25$?
    Require $\frac{s}{\sqrt N}\le 25$, i.e. $\frac{2000}{\sqrt N}\le 25$. Solve: $\sqrt N\ge\frac{2000}{25}=80$, so $N\ge 80^{2}=6400$. Thus at least $N=6{,}400$ runs.
  42. Monte Carlo & bootstrap
    Describe the **bootstrap** and what it estimates.
    The (nonparametric) bootstrap resamples the observed data **with replacement** to approximate the sampling distribution of a statistic. Draw $B$ resamples each of size $n$ from the original sample, recompute the statistic $\hat\theta^{*}_{b}$ on each, and use the spread of the $\hat\theta^{*}$ values. The **bootstrap standard error** is the sample sd of the $\hat\theta^{*}_{b}$, and a percentile interval uses the empirical quantiles of those replicates. It is useful when an analytic standard error is hard to derive.
  43. Monte Carlo & bootstrap
    Three bootstrap resamples of a sample give statistic values $\hat\theta^{*}=12,\,16,\,14$. Estimate the bootstrap standard error.
    Mean of replicates $=\frac{12+16+14}{3}=\frac{42}{3}=14$. Squared deviations: $(12-14)^{2}=4$, $(16-14)^{2}=4$, $(14-14)^{2}=0$; sum $=8$. Bootstrap SE (using divisor $B-1=2$) $=\sqrt{\frac{8}{2}}=\sqrt{4}=2$.
  44. Monte Carlo & bootstrap
    Explain how the bootstrap can estimate the **bias** of an estimator $\hat\theta$.
    Compute $\hat\theta$ on the original data, then on each bootstrap resample obtain $\hat\theta^{*}_{b}$. The bootstrap bias estimate is $\widehat{\text{bias}}=\bar{\theta}^{*}-\hat\theta$, where $\bar{\theta}^{*}=\frac{1}{B}\sum_{b}\hat\theta^{*}_{b}$. The resampling treats $\hat\theta$ as the "true" parameter and the resamples as new samples, so the average over-/under-shoot mimics the estimator's bias. A bias-corrected estimate is $\hat\theta-\widehat{\text{bias}}=2\hat\theta-\bar\theta^{*}$.