定量方法（Quantitative Methods）— 抽样与估计 · 第 4 课
一、引言：回顾与过渡
前情回顾（L127）：标准误 SE = σ/√n，衡量样本均值作为估计量的精确度。标准误越小，估计越精确。
有了标准误，我们就可以做两件事： 1. 给出总体参数的最佳单一猜测值（点估计） 2. 给出总体参数可能落在的区间范围（区间估计）
二、点估计（Point Estimate）
2.1 定义
点估计（Point Estimate）： 用一个单一的样本统计量值来估计总体参数。
$$\mu \approx \bar{x}, \quad \sigma^2 \approx s^2, \quad p \approx \hat{p}$$
2.2 常用点估计对照表
| 总体参数 | 符号 | 点估计量 | 符号 |
|---|---|---|---|
| 总体均值 | μ | 样本均值 | $\bar{x}$ |
| 总体方差 | σ² | 样本方差 | $s^2$ |
| 总体标准差 | σ | 样本标准差 | $s$ |
| 总体比例 | p | 样本比例 | $\hat{p}$ |
| 两总体均值差 | μ₁ - μ₂ | 样本均值差 | $\bar{x}_1 - \bar{x}_2$ |
2.3 点估计的局限
案例： 分析师测算某 ETF 过去 60 个月的年化收益率 $\bar{x}$ = 8.2%，并直接用 8.2% 作为未来预期收益率的估计。
问题在哪？ 这个 8.2% 是一个精确到小数点的数字，但分析师心里很清楚：真实值可能高于 8.2%，也可能低于 8.2%。点估计没有告诉我们这个"可能偏离多少"。
💡 用一个数字去代表总体参数，就像用一根针去钉住一只飘动的气球——无法体现不确定性。
2.4 优秀估计量的三大性质
CFA 考试要求掌握判断点估计是否"好"的三个标准：
| 性质 | 英文 | 含义 | 直观理解 |
|---|---|---|---|
| 无偏性 | Unbiasedness | 估计量的期望值 = 真值：$E(\hat{\theta}) = \theta$ | 长期反复抽样，"平均而言"打中靶心 |
| 有效性 | Efficiency | 在所有无偏估计量中方差最小 | 同样"打中靶心"，弹孔最集中 |
| 一致性 | Consistency | n → ∞ 时，估计量依概率收敛于真值 | 样本越大，离真值越近 |
🎯 靶纸类比： - 无偏但不有效 → 弹孔散布在靶心四周，但没有系统偏左/偏右 - 有效但偏倚 → 弹孔集中，但都偏移到靶心左侧 - 无偏、有效但不一致 → 小样本时就很好，大样本也没有继续改进
三、区间估计（Confidence Interval）
3.1 为什么需要区间估计？
承接上文：点估计给了我们一个具体数字，但没有告诉我们不确定性有多大。
区间估计（Interval Estimate）： 给出一个数值范围，并声明总体参数以一定置信水平落在此范围内。
$$\text{点估计} \pm (\text{可靠系数} \times \text{标准误})$$
3.2 置信区间的构造公式
场景一：总体方差 σ² 已知（或大样本 n ≥ 30）
$$\bar{x} \pm z_{\alpha/2} \times \frac{\sigma}{\sqrt{n}}$$
其中 $z_{\alpha/2}$ 为标准正态分布的临界值。
场景二：总体方差 σ² 未知，小样本（n < 30）
$$\bar{x} \pm t_{\alpha/2, \, n-1} \times \frac{s}{\sqrt{n}}$$
其中 $t_{\alpha/2, \, n-1}$ 为 t 分布的临界值，自由度 = n-1。
3.3 常用置信水平对应的临界值
| 置信水平 | $1-\alpha$ | $\alpha/2$ | $z_{\alpha/2}$ | 通俗说法 |
|---|---|---|---|---|
| 90% | 0.90 | 0.05 | 1.645 | "10 次里有 9 次" |
| 95% | 0.95 | 0.025 | 1.96 | "20 次里有 19 次" |
| 99% | 0.99 | 0.005 | 2.576 | "100 次里有 99 次" |
📌 考试必记： 95% → 1.96，90% → 1.645，99% → 2.576
四、置信区间的正确解读 ⚠️ 高频考点
4.1 正确解释
对于 95% 置信区间
[a, b]，正确的解读是："如果我们从总体中重复抽取大量容量为 n 的样本，并对每个样本构造一个 95% 置信区间，那么约 95% 的区间会包含真实的总体参数。"
4.2 错误解释（CFA 考试常见陷阱）
| 错误表述 | 为什么错了？ |
|---|---|
| ❌ "参数 μ 有 95% 的概率落在 [a, b] 内" | μ 是固定常数，不是随机变量，不能说"概率" |
| ❌ "95% 的样本数据落在 [a, b] 内" | 混淆了置信区间与预测区间 |
| ❌ "该区间包含 μ 的概率是 95%" | 区间是随机的，μ 是固定的——一旦计算出来，区间要么包含 μ，要么不包含 |
🔥 正确表述的模板： "We are 95% confident that the interval [a, b] contains the true population parameter μ."
4.3 可视化理解
假设真实 μ = 100，我们重复抽取 20 个样本，每个构造 95% CI：
样本 1: [98, 104] ✅ 包含 μ
样本 2: [96, 102] ✅ 包含 μ
...
样本 19: [101, 107] ✅ 包含 μ
样本 20: [105, 111] ❌ 没包含 μ ← 约 1/20 = 5%
预期：20 个区间中约 19 个包含 μ（95%）
五、影响置信区间宽度的因素
$$\text{CI 宽度} = 2 \times \text{临界值} \times \frac{\sigma}{\sqrt{n}}$$
| 因素 | 变化方向 | 对区间宽度的影响 | 逻辑 |
|---|---|---|---|
| 样本量 n ↑ | 增大 | 变窄 | 信息更多，估计更精确 |
| 置信水平 (1-α) ↑ | 增大 | 变宽 | 要有更高把握，就得更保守 |
| 总体波动 σ ↑ | 增大 | 变宽 | 数据本身不确定性大 |
| 从 z 改用 t | — | 变宽 | t 分布尾部更厚（自由度较小时明显） |
💡 实用结论： 提高精度（缩窄 CI）⇄ 提高置信度（更宽的区间）是一对矛盾。实践中通常固定置信度（如 95%），通过增加样本量来缩窄区间。
六、实战案例
案例一：基金业绩评估
场景： 某对冲基金过去 36 个月的月度超额收益（相对于基准）均值 $\bar{x}$ = 0.35%，标准差 s = 1.8%。请构造该基金月度 alpha 的 95% 置信区间。
$$\text{SE} = \frac{1.8\%}{\sqrt{36}} = 0.30\%$$
$$95\% \text{ CI} = 0.35\% \pm 1.96 \times 0.30\% = 0.35\% \pm 0.588\% = [-0.238\%, 0.938\%]$$
🟡 解读： 区间跨过零，说明在 95% 置信水平下，不能排除真实 alpha 为负的可能。36 个月的数据可能不够。
案例二：样本量对 CI 宽度的实际影响
接上例，如果该基金有 144 个月（12 年）的业绩数据，s 仍为 1.8%：
$$\text{SE} = \frac{1.8\%}{\sqrt{144}} = 0.15\%$$
$$95\% \text{ CI} = 0.35\% \pm 1.96 \times 0.15\% = 0.35\% \pm 0.294\% = [0.056\%, 0.644\%]$$
🟢 解读： CI 不跨零，区间整体为正——有 95% 的置信度认定 alpha 为正。这正是机构投资者要求长业绩记录的数学基础。
七、点估计 vs 区间估计：对决
| 维度 | 点估计 | 区间估计 |
|---|---|---|
| 输出形式 | 单一数字 | 一个范围 [L, U] |
| 包含不确定信息？ | ❌ 无 | ✅ 有（通过宽度体现） |
| 样本量信息 | ❌ 不体现 | ✅ 通过 SE 体现 |
| 可用于比较 | ✅ 方便排序 | ⚠️ 重叠区间不能简单比较 |
| 实战应用 | 初始筛查、排序 | 检验显著性、风控 |
🔥 实战铁律： 永远不要只看点估计！没有标准误/置信区间的点估计就像没有量纲的数字——看起来精确，实则危险。
八、本节要点总结
| # | 要点 |
|---|---|
| 1 | 点估计 = 单一统计量估计参数；区间估计 = 范围 + 置信水平 |
| 2 | 好估计量的三性：无偏性（长期打中靶心）、有效性（最集中）、一致性（大样本收敛） |
| 3 | CI 公式：$\text{点估计} \pm \text{临界值} \times \text{标准误}$ |
| 4 | 正确解释 CI：重复抽样的长期频率，不是单一区间包含参数的概率 |
| 5 | 影响 CI 宽度的因素：n ↑ → 变窄；置信度 ↑ → 变宽；σ ↑ → 变宽 |
| 6 | σ 已知用 z，σ 未知小样本用 t。大样本（n ≥ 30）可近似用 z |
| 7 | CI 跨过零 → 统计上不显著；CI 全部为正/负 → 有方向性结论 |
九、测试题
Q1： 分析师以样本均值 15 作为总体均值的估计。这是一个：
A. 区间估计 B. 点估计 C. 无偏估计
Q2： 以下关于 95% 置信区间的描述，哪一项是正确的？
A. 该区间有 95% 的概率包含总体参数 B. 如果反复抽样并构造区间，约 95% 的区间会包含总体参数 C. 95% 的总体数据落在该区间内
Q3： 某样本容量为 49，样本均值 = 100，总体标准差 σ = 14。总体均值的 95% 置信区间为：
A. [96.08, 103.92] B. [97.00, 103.00] C. [94.00, 106.00]
Q4： 在不改变置信水平的前提下，以下哪种方法能最有效地缩窄置信区间？
A. 增大样本量 B. 使用 t 临界值代替 z 临界值 C. 降低总体标准差
Q5： 以下关于点估计性质的描述，哪一项是正确的？
A. 一个估计量可以是有偏但一致的 B. 无偏性意味着用任一次样本计算的估计值一定等于真值 C. 有效性意味着估计量的期望值等于真值
Q6（判断）： 将置信水平从 95% 提升到 99%，置信区间会变窄。
答案与解析
Q1：B 用"样本均值 15"这一个数字来估计总体参数，是典型的点估计。仅凭这个信息无法判断是否无偏（选项 C 需要额外信息）。
Q2：B 经典 CFA 考点。A 是把参数当随机变量（错误）；C 混淆了置信区间和容许区间。正确解释围绕"重复抽样"展开。
Q3：A $$\text{SE} = 14/\sqrt{49} = 2$$ $$95\%\text{ CI} = 100 \pm 1.96 \times 2 = 100 \pm 3.92 = [96.08, 103.92]$$
Q4：A 只有增大样本量是可控的。使用 t 临界值会使区间变宽（B 错误）；降低 σ 是总体固有属性，不可控（C 不现实）。
Q5：A 经典考点。有偏但一致：小样本时有偏差，但样本增大后收敛到真值（如：用 $\frac{\sum(X_i-\bar{x})^2}{n}$ 算方差）。B 混淆了无偏性（期望相等）和每次结果相等。C 混淆了有效性和无偏性。
Q6：错误。 置信水平越高 → z 值越大 → CI 越宽。要在更高把握下涵盖真值，区间必须更保守。
🎯 下一课预告：L129 · 假设检验（一）— 零假设与备择假设的设定
Quantitative Methods — Sampling and Estimation · Lesson 4
1. Introduction: Review and Transition
Recap (L127): Standard Error SE = σ/√n measures the precision of the sample mean as an estimator. The smaller the SE, the more precise the estimate.
With standard error in hand, we can do two things: 1. Provide the best single guess for a population parameter (point estimate) 2. Provide a range within which the population parameter likely falls (interval estimate)
2. Point Estimate
2.1 Definition
Point Estimate: A single sample statistic value used to estimate a population parameter.
$$\mu \approx \bar{x}, \quad \sigma^2 \approx s^2, \quad p \approx \hat{p}$$
2.2 Common Point Estimates
| Population Parameter | Symbol | Point Estimator | Symbol |
|---|---|---|---|
| Population mean | μ | Sample mean | $\bar{x}$ |
| Population variance | σ² | Sample variance | $s^2$ |
| Population standard deviation | σ | Sample standard deviation | $s$ |
| Population proportion | p | Sample proportion | $\hat{p}$ |
| Difference between two means | μ₁ − μ₂ | Difference of sample means | $\bar{x}_1 - \bar{x}_2$ |
2.3 Limitations of Point Estimates
Example: An analyst calculates that an ETF has an annualized return of $\bar{x}$ = 8.2% over the past 60 months and uses 8.2% directly as the expected future return.
What's wrong? This 8.2% is a precise number to two decimal places, but the analyst knows deep down: the true value could be higher or lower. A point estimate alone reveals nothing about how far off it could be.
💡 Using a single number to represent a population parameter is like pinning a drifting balloon with a needle — it conveys zero information about uncertainty.
2.4 Three Properties of a Good Estimator
The CFA exam requires you to know three criteria for judging whether a point estimator is "good":
| Property | Meaning | Intuitive Interpretation |
|---|---|---|
| Unbiasedness | Expected value of the estimator equals the true value: $E(\hat{\theta}) = \theta$ | Over repeated sampling, "on average" hits the bullseye |
| Efficiency | Among all unbiased estimators, has the smallest variance | Hits the bullseye with the tightest spread |
| Consistency | As n → ∞, the estimator converges in probability to the true value | Larger samples → closer to the truth |
🎯 Dartboard analogy: - Unbiased but inefficient → shots are spread around the bullseye, but no systematic left/right bias - Efficient but biased → shots are tightly clustered, but offset to one side of the bullseye - Unbiased, efficient but inconsistent → great even with small samples; no further improvement with large samples
3. Confidence Interval (Interval Estimate)
3.1 Why Do We Need Interval Estimation?
Point estimates give us a specific number, but tell us nothing about how much uncertainty surrounds that number.
Confidence Interval (Interval Estimate): A range of values, constructed from sample data, that states the population parameter lies within this range with a given level of confidence.
$$\text{Point Estimate} \pm (\text{Reliability Factor} \times \text{Standard Error})$$
3.2 Confidence Interval Formulas
Case 1: Population variance σ² known (or large sample n ≥ 30)
$$\bar{x} \pm z_{\alpha/2} \times \frac{\sigma}{\sqrt{n}}$$
Where $z_{\alpha/2}$ is the critical value from the standard normal distribution.
Case 2: Population variance σ² unknown, small sample (n < 30)
$$\bar{x} \pm t_{\alpha/2, \, n-1} \times \frac{s}{\sqrt{n}}$$
Where $t_{\alpha/2, \, n-1}$ is the critical value from the t-distribution with df = n−1.
3.3 Critical Values for Common Confidence Levels
| Confidence Level | $1-\alpha$ | $\alpha/2$ | $z_{\alpha/2}$ | Plain English |
|---|---|---|---|---|
| 90% | 0.90 | 0.05 | 1.645 | "9 out of 10 times" |
| 95% | 0.95 | 0.025 | 1.96 | "19 out of 20 times" |
| 99% | 0.99 | 0.005 | 2.576 | "99 out of 100 times" |
📌 Must memorize for the exam: 95% → 1.96, 90% → 1.645, 99% → 2.576
4. Interpreting Confidence Intervals Correctly ⚠️ High-Frequency Exam Topic
4.1 Correct Interpretation
For a 95% confidence interval
[a, b], the correct interpretation is:"If we repeatedly draw many samples of size n from the population and construct a 95% CI for each sample, then approximately 95% of these intervals will contain the true population parameter."
4.2 Incorrect Interpretations (Common CFA Traps)
| Incorrect Statement | Why It's Wrong |
|---|---|
| ❌ "There is a 95% probability that μ lies within [a, b]" | μ is a fixed constant, not a random variable — probability language is incorrect |
| ❌ "95% of the sample data falls within [a, b]" | Confuses confidence interval with prediction interval |
| ❌ "The probability that this interval contains μ is 95%" | The interval is random, μ is fixed — once computed, the interval either contains μ or it doesn't |
🔥 Correct phrasing template: "We are 95% confident that the interval [a, b] contains the true population parameter μ."
4.3 Visual Illustration
Suppose the true μ = 100. We draw 20 independent samples, each yielding a 95% CI:
Sample 1: [98, 104] ✅ contains μ
Sample 2: [96, 102] ✅ contains μ
...
Sample 19: [101, 107] ✅ contains μ
Sample 20: [105, 111] ❌ does NOT contain μ ← roughly 1/20 = 5%
Expected: approximately 19 out of 20 intervals contain μ (95%)
5. Factors Affecting Confidence Interval Width
$$\text{CI Width} = 2 \times \text{Critical Value} \times \frac{\sigma}{\sqrt{n}}$$
| Factor | Change | Effect on CI Width | Rationale |
|---|---|---|---|
| Sample size n ↑ | Increase | Narrows | More information → more precise estimate |
| Confidence level (1−α) ↑ | Increase | Widens | Higher confidence requires a more conservative range |
| Population variability σ ↑ | Increase | Widens | Greater inherent data uncertainty |
| Switch from z to t | — | Widens | t-distribution has fatter tails (notable when df is small) |
💡 Practical takeaway: Improving precision (narrower CI) ⇄ Raising confidence level (wider CI) is a trade-off. In practice, we fix the confidence level (e.g., 95%) and increase sample size to narrow the interval.
6. Real-World Applications
Case 1: Fund Performance Evaluation
Scenario: A hedge fund has 36 months of monthly excess returns (vs. benchmark) with mean $\bar{x}$ = 0.35% and standard deviation s = 1.8%. Construct a 95% confidence interval for the fund's monthly alpha.
$$\text{SE} = \frac{1.8\%}{\sqrt{36}} = 0.30\%$$
$$95\% \text{ CI} = 0.35\% \pm 1.96 \times 0.30\% = 0.35\% \pm 0.588\% = [-0.238\%, 0.938\%]$$
🟡 Interpretation: The interval spans zero, meaning that at the 95% confidence level, we cannot rule out the possibility that the true alpha is negative. 36 months of data may not be sufficient.
Case 2: The Impact of Sample Size on CI Width
Continuing from the previous example: suppose the fund has 144 months (12 years) of track record, with s still 1.8%:
$$\text{SE} = \frac{1.8\%}{\sqrt{144}} = 0.15\%$$
$$95\% \text{ CI} = 0.35\% \pm 1.96 \times 0.15\% = 0.35\% \pm 0.294\% = [0.056\%, 0.644\%]$$
🟢 Interpretation: The CI is entirely positive — we are 95% confident the alpha is positive. This is the mathematical basis for why institutional investors demand long track records.
7. Point Estimate vs Confidence Interval: Head-to-Head
| Dimension | Point Estimate | Confidence Interval |
|---|---|---|
| Output form | Single number | A range [L, U] |
| Contains uncertainty info? | ❌ No | ✅ Yes (via width) |
| Sample size info reflected? | ❌ Not reflected | ✅ Reflected through SE |
| Useful for ranking/sorting? | ✅ Easy to rank | ⚠️ Overlapping CIs complicate comparisons |
| Practical use | Initial screening, ranking | Significance testing, risk management |
🔥 Golden rule: Never rely solely on a point estimate! A point estimate without a standard error or confidence interval is like a number without units — it looks precise but can be dangerously misleading.
8. Key Takeaways
| # | Takeaway |
|---|---|
| 1 | Point estimate = single statistic for a parameter; Interval estimate = range + confidence level |
| 2 | Three properties of good estimators: unbiasedness (long-run on target), efficiency (tightest spread), consistency (converges with large samples) |
| 3 | CI formula: $\text{Point Estimate} \pm \text{Critical Value} \times \text{Standard Error}$ |
| 4 | Correct CI interpretation: long-run frequency of repeated sampling, NOT the probability that a single interval contains the parameter |
| 5 | Factors affecting CI width: n ↑ → narrower; Confidence level ↑ → wider; σ ↑ → wider |
| 6 | σ known: use z; σ unknown, small sample: use t. Large samples (n ≥ 30): z approximation is acceptable |
| 7 | CI spans zero → statistically insignificant; CI entirely positive/negative → directional conclusion |
9. Practice Questions
Q1: An analyst uses a sample mean of 15 to estimate the population mean. This is an example of:
A. An interval estimate B. A point estimate C. An unbiased estimate
Q2: Which of the following is a correct description of a 95% confidence interval?
A. There is a 95% probability that the interval contains the population parameter B. If we repeatedly sample and construct intervals, approximately 95% of the intervals will contain the population parameter C. 95% of the population data falls within the interval
Q3: A sample of size n = 49 yields a sample mean of 100 with a known population standard deviation σ = 14. The 95% confidence interval for the population mean is:
A. [96.08, 103.92] B. [97.00, 103.00] C. [94.00, 106.00]
Q4: Without changing the confidence level, which method most effectively narrows the confidence interval?
A. Increase the sample size B. Use a t critical value instead of a z critical value C. Reduce the population standard deviation
Q5: Which of the following statements about the properties of point estimators is correct?
A. An estimator can be biased but consistent B. Unbiasedness means that an estimate from any single sample must equal the true value C. Efficiency means that the expected value of the estimator equals the true value
Q6 (True/False): Increasing the confidence level from 95% to 99% will narrow the confidence interval.
Answers and Explanations
Q1: B Using a single number ("sample mean of 15") to estimate a population parameter is a classic point estimate. There is not enough information to determine whether it is unbiased (option C requires additional information).
Q2: B A classic CFA exam trap. Option A incorrectly treats the parameter as a random variable. Option C confuses confidence intervals with tolerance intervals. The correct interpretation centers on "repeated sampling."
Q3: A $$\text{SE} = 14/\sqrt{49} = 2$$ $$95\%\text{ CI} = 100 \pm 1.96 \times 2 = 100 \pm 3.92 = [96.08, 103.92]$$
Q4: A Only increasing the sample size is controllable. Using a t critical value would widen the interval (B is wrong). Reducing σ is a property of the population and not under our control (C is unrealistic).
Q5: A A classic exam point. Biased but consistent: an estimator may have a small-sample bias but converge to the true value as the sample grows (e.g., using $\frac{\sum(X_i-\bar{x})^2}{n}$ for variance). Option B confuses unbiasedness (expectation equality) with individual outcomes. Option C confuses efficiency with unbiasedness.
Q6: False. Higher confidence level → larger z value → wider CI. To cover the true value with greater confidence, the interval must be more conservative.
🎯 Next lesson: L129 · Hypothesis Testing (I) — Setting up the Null and Alternative Hypotheses