CFA Level I · Quantitative LOS 7.

a ·
Reading 7 — Estimation and inference Methods · 2025 7.b · 7.c

SAMPLING METHODS (LOS 7.A) CENTRAL LIMIT THEOREM (LOS 7.B)

Probability sampling — every member has a known The most important theorem in statistics
chance of selection. for finance.

Method How it works Sampling The CLT states:

error For a population with mean μ and finite variance σ², if
you take random samples of size n ≥ 30, the
Simple Every member has equal Baseline —
distribution of SAMPLE MEANS will be approximately
random chance. Draw randomly moderate
normal — regardless of what shape the original
from full population.
population has.
Stratified Split into subgroups Lower than
random (strata), draw simple
proportionally from each. random 3 KEY PROPERTIES OF THE CLT
Used in bond indexing.
Sample means are normally distributed (if n ≥
Cluster Split into clusters, Higher — 1
30)
randomly select entire clusters
clusters. Each cluster may not 2 Mean of sample means = population mean μ
assumed representative. represent
population Variance of sample means = σ²/n (shrinks as n
3
grows)

Non-probability sampling — randomness not

guaranteed. STANDARD ERROR OF THE SAMPLE MEAN

σ known (rare):
Convenienc Use whatever data is High — not
e easily available. random
σ
σx̄ =
Judgmental Researcher hand-picks Can be low √n
observations based on if unbiased,
experience/judgment. high if
biased σ unknown (usual — use sample std dev s):

s
Stratified vs Cluster: Stratified = sample from EVERY sx̄ =
subgroup. Cluster = sample SOME subgroups entirely. √n
Key exam distinction.
Larger n → smaller standard error → sample mean is
closer to true population mean. This is WHY bigger
samples are better.

WORKED EXAMPLE

n=30, s=20%, mean=2% → sx̄ = 20/√30 = 3.6%

n=200, s=20%, mean=2% → sx̄ = 20/√200 = 1.4%

Same data, bigger sample: standard error drops from

3.6% to 1.4%. More data → more precision.
THE MAGIC OF THE CLT — WHY IT MATTERS RESAMPLING — JACKKNIFE & BOOTSTRAP (LOS 7.C)

The population can be ANY shape — skewed, bimodal, Used to estimate the standard error of a statistic when
weird. Doesn't matter. Once n ≥ 30, the distribution of the formula alone isn't reliable.
sample means becomes normal.

Jackknife Bootstrap
n ≥ 30 How Remove one obs at Draw repeated
Population n = 5 sample
sample a time. Calculate samples of size n
shape means
means statistic n times WITH replacement.
(each with one Calculate statistic
Still somewhat Normal bell
Skewed right removed). each time.
skewed curve
Result Std dev of those n Std dev of all
Getting more Normal bell means = std error sample means =
Uniform (flat)
bell-shaped curve estimate std error estimate
Bimodal (two Two humps Normal bell Strength Simple. Removes More accurate.
humps) fading curve bias. Good for small Works for median,
samples. complex stats,
non-analytic
Why this is powerful: We can use z-tests and t-tests distributions.
(which assume normality) on sample means — even if Limitatio Less accurate than Computationally
the underlying data is not normal — as long as n ≥ 30. n bootstrap. demanding (more
processing power
needed).
SAMPLING ERROR VS STANDARD ERROR

Sampling Difference between a sample statistic Exam tip: Bootstrap is MORE computationally
and the true population parameter. demanding than jackknife. Jackknife was created
error
Always exists — cannot be eliminated. when computing was expensive — it's the simpler,
cheaper tool.
Std dev of the sampling distribution of
Standard the mean. Measures HOW MUCH
error sampling error we expect on average.
Reduced by larger n.

EXAM MEMORY TRICKS

CLT magic n ≥ 30 → Standard error sx̄ = s/√n. Stratified vs Cluster Jackknife vs Bootstrap
sample means are Bigger n → smaller std Stratified = sample from Bootstrap is more
normal, no matter the error → more precise EVERY subgroup. Cluster accurate but more
population shape. estimate. = randomly pick SOME computationally
subgroups. demanding than
jackknife.