定量方法（Quantitative Methods）— 概率论模块 · 第三课
一、本课定位
L116 学习了期望值、方差和协方差——用来描述随机变量"中心在哪""波动多大"和"怎么联动"。本课进入概率论的核心思维工具：如何在新信息到来时更新我们的判断。
| 项目 | 说明 |
|---|---|
| 模块 | 2.4 概率论 |
| 前置知识 | L115 概率基础、L116 期望与方差 |
| 后续衔接 | L118 计数原理（排列组合） |
| 难度 | ★★★★☆ |
| 考试权重 | 高（概念题 + 计算题，每年必考） |
| 阅读时间 | 约 15 分钟 |
二、核心概念
1. 条件概率（Conditional Probability）
直觉引入：
假设一家公司面试了 100 个候选人。其中 20 人通过了技术面，15 人最终拿到 offer。但拿到 offer 的人中，有 12 人通过了技术面。
问：已知一个人通过了技术面，他拿到 offer 的概率是多少？
你的直觉会自然缩小"样本空间"——只看那 20 个通过技术面的人，然后看里面有多少拿到了 offer：
$$P(\text{Offer} \mid \text{技术面通过}) = \frac{12}{20} = 60\%$$
这就是条件概率的本质：有了新信息之后，重新划定范围，重新算比例。
正式定义：
$$P(A \mid B) = \frac{P(A \cap B)}{P(B)}, \quad P(B) > 0$$
其中： - P(A|B)：在事件 B 已经发生的条件下，事件 A 发生的概率 - P(A ∩ B)：A 和 B 同时发生的联合概率 - P(B)：事件 B 发生的概率（新的"全集"）
🧠 读法："给定 B 时 A 的概率"，竖线右边的是已知条件。
案例 1：行业认证考试
CFA 一级考生中，60% 来自金融背景（F），40% 来自非金融背景。通过率分别为 70% 和 50%。
| 背景 | 占比 | 通过率 |
|---|---|---|
| 金融背景 (F) | 60% | 70% |
| 非金融背景 (Fᶜ) | 40% | 50% |
问题：随机抽一个通过的人，他是金融背景的概率？
先算联合概率： - P(F ∩ 通过) = 0.6 × 0.7 = 0.42 - P(Fᶜ ∩ 通过) = 0.4 × 0.5 = 0.20 - P(通过) = 0.42 + 0.20 = 0.62（总通过率）
$$P(F \mid \text{通过}) = \frac{P(F \cap \text{通过})}{P(\text{通过})} = \frac{0.42}{0.62} \approx 67.7\%$$
📊 直觉检查： 金融背景通过率高（70% vs 50%），所以通过者中金融背景占比（67.7%）确实高于整体中的金融背景占比（60%）。方向对了。
2. 乘法法则（Multiplication Rule）
从条件概率定义变形得到：
$$P(A \cap B) = P(A \mid B) \cdot P(B) = P(B \mid A) \cdot P(A)$$
推广到三个事件：
$$P(A \cap B \cap C) = P(A) \cdot P(B \mid A) \cdot P(C \mid A \cap B)$$
📊 实战应用： 结构化产品违约分析中，常用乘法法则层层分解联合违约概率。
案例 2：连续违约风险
某投资组合中： - P(公司A违约) = 5% - 若A违约，P(B违约｜A违约) = 40% - 若A、B都违约，P(C违约｜A∩B) = 60%
三家同时违约的概率：
$$P(A \cap B \cap C) = 5\% \times 40\% \times 60\% = 1.2\%$$
看上去每家违约概率不高，但链条叠加后仍有 1.2%——这就是系统性风险传染的数学本质。
3. 独立性的正式定义
L115 直觉上讲了"一件事不影响另一件事"。现在用条件概率精确表达：
事件 A 和 B 独立，当且仅当：
$$P(A \mid B) = P(A) \quad \text{或等价地} \quad P(A \cap B) = P(A) \cdot P(B)$$
🧠 B 的发生有没有改变 A 的概率？没有 → 两者独立。
注意区分： - 互斥（Mutually Exclusive）： P(A ∩ B) = 0，两个事件不能同时发生 - 独立（Independent）： P(A ∩ B) = P(A)·P(B)，一个发生不影响另一个
🔴 CFA 高频陷阱： 互斥 ≠ 独立！如果 A、B 互斥且概率都大于 0，则它们一定不独立——因为已知 A 发生，B 必不发生；P(B|A) = 0 ≠ P(B)。
4. 全概率公式（Total Probability Rule）
动机： 有时候我们直接算 P(A) 很难，但通过"分情况"算 P(A|Bᵢ) 很容易。
公式： 如果 B₁, B₂, ..., Bₙ 构成样本空间的一个划分（互斥且穷尽），则：
$$P(A) = \sum_{i=1}^{n} P(A \mid B_i) \cdot P(B_i)$$
通俗理解： 把 A 的概率按"场景"拆开，每个场景加权求和。
案例 3：全概率公式——分析师评级
某券商分析师对股票给出三种评级：
| 评级 | 出现概率 | 该评级下股票上涨概率 |
|---|---|---|
| 买入 (B) | 40% | 75% |
| 持有 (H) | 35% | 50% |
| 卖出 (S) | 25% | 20% |
问题： 不考虑评级时，该股票上涨的总概率是多少？
$$P(\text{上涨}) = P(\text{涨} \mid B) \cdot P(B) + P(\text{涨} \mid H) \cdot P(H) + P(\text{涨} \mid S) \cdot P(S)$$
$$= 0.75 \times 0.40 + 0.50 \times 0.35 + 0.20 \times 0.25$$
$$= 0.30 + 0.175 + 0.05 = 0.525 = 52.5\%$$
📊 全概率公式就是把"加权平均"的思想从收益率搬到了概率上——场景概率 × 条件概率，全部加起来。
5. 贝叶斯公式（Bayes' Formula）—— 本课核心
这是 CFA 量化方法中最优雅、最强大的工具之一。
公式：
$$P(B_i \mid A) = \frac{P(A \mid B_i) \cdot P(B_i)}{\sum_{j=1}^{n} P(A \mid B_j) \cdot P(B_j)}$$
结构的含义：
| 部分 | 名称 | 含义 |
|---|---|---|
| P(Bᵢ) | 先验概率（Prior） | 看到新信息之前，我们对 Bᵢ 的判断 |
| P(A|Bᵢ) | 似然（Likelihood） | 如果 Bᵢ 为真，观察到 A 的可能性 |
| 分母 | 归一化常数 | 在所有可能情况下观察到 A 的总概率 |
| P(Bᵢ|A) | 后验概率（Posterior） | 看到 A 之后，更新后的判断 |
$$Posterior = \frac{Likelihood \times Prior}{Total\;Probability}$$
🧠 一句话：贝叶斯公式告诉你怎么用新证据更新旧信念。
案例 4：贝叶斯——基金经理能力评估（CFA 经典）
某基金经理声称自己能持续战胜市场。
- 先验判断：行业中只有 10% 的基金经理真正有能力（Genuine, G），90% 靠运气（Luck, L）
- 有能力的经理某年跑赢市场的概率 = 80%
- 靠运气的经理某年跑赢市场的概率 = 50%（市场随机）
问题：如果某经理今年跑赢了市场，他真正有能力的概率是多少？
步骤 1：整理已知信息 - P(G) = 0.10（先验：他真的行） - P(L) = 0.90（先验：他靠运气） - P(赢 ∣ G) = 0.80 - P(赢 ∣ L) = 0.50
步骤 2：计算全概率（分母） $$P(\text{赢}) = 0.80 \times 0.10 + 0.50 \times 0.90$$ $$= 0.08 + 0.45 = 0.53$$
步骤 3：应用贝叶斯公式 $$P(G \mid \text{赢}) = \frac{0.80 \times 0.10}{0.53} = \frac{0.08}{0.53} \approx 15.1\%$$
📊 惊人发现： 即使跑赢了市场，真正有能力的概率也只从 10% 提升到 15.1%！这是因为靠运气赢的概率（50%）也不低，而且运气型经理太多（90%）。
案例 5：贝叶斯——内幕交易检测
SEC 调查某交易员是否从事内幕交易（Insider, I）。
- P(I) = 5%（先验：仅有 5% 交易员在暗中从事内幕交易）
- 如果做内幕交易，异常盈利的概率 = 95%
- 如果没做内幕交易，异常盈利的概率 = 2%（罕见事件）
该交易员出现异常盈利。他做内幕交易的后验概率：
$$P(I \mid \text{异常}) = \frac{0.95 \times 0.05}{0.95 \times 0.05 + 0.02 \times 0.95}$$
$$= \frac{0.0475}{0.0475 + 0.019} = \frac{0.0475}{0.0665} \approx 71.4\%$$
📊 从 5% 飙升到 71.4%——因为"没干内幕交易却异常盈利"非常罕见（2%），所以观察到异常盈利时，后验概率大幅跃升。这是贝叶斯更新力量最强的场景：似然比（95%/2% = 47.5 倍）越大，更新越剧烈。
6. 贝叶斯公式的"自然频率"理解
人类大脑不擅长概率，但擅长数人头。
把案例 4 翻译成自然频率（假设 1000 个基金经理）：
| 类型 | 人数 | 赢市场的比例 | 赢市场的人数 |
|---|---|---|---|
| 有能力 (G) | 100 | 80% | 80 |
| 靠运气 (L) | 900 | 50% | 450 |
| 总计 | 1000 | — | 530 |
💡 530 个人赢了市场。其中 80 个是真有能力的。所以： $$P(G \mid \text{赢}) = \frac{80}{530} \approx 15.1\%$$
和贝叶斯公式完全一致。算贝叶斯题的最佳心法：画 2×2 表格！
三、重要公式汇总
| 概念 | 公式 | 关键词 |
|---|---|---|
| 条件概率 | P(A|B) = P(A∩B) / P(B) | 缩小样本空间 |
| 乘法法则 | P(A∩B) = P(A|B)·P(B) | 链式分解 |
| 独立性检验 | P(A|B) = P(A) 或 P(A∩B)=P(A)·P(B) | B 不改变 A |
| 全概率公式 | P(A) = Σ P(A|Bᵢ)·P(Bᵢ) | 按场景加权 |
| 贝叶斯公式 | P(Bᵢ|A) = P(A|Bᵢ)·P(Bᵢ) / P(A) | 先验 → 后验 |
四、贝叶斯更新的 CFA 应用场景
| 领域 | 应用 |
|---|---|
| 投资分析 | 收到新财报 → 更新对公司质量的判断 |
| 风险管理 | 观察到违约信号 → 更新违约概率 |
| 技术分析 | 看到突破形态 → 更新趋势判断的概率 |
| 行为金融 | 过度反应 vs 保守偏差——锚定先验不够更新 |
| 资产定价 | 多阶段信息到达时的动态信念更新 |
五、常见易错点
| 易错 | 正确认识 |
|---|---|
| P(A|B) = P(B|A) ❌ | 条件概率不能随便交换！两者一般不等 |
| 互斥 = 独立 ❌ | 互斥意味着 P(A∩B)=0，而独立需要 P(A∩B)=P(A)·P(B) |
| 贝叶斯分母忘了全概率 ❌ | 分母是 P(A) = Σ P(A|Bⱼ)·P(Bⱼ)，不是随便一个数 |
| P(A|B) > P(A) 说明 A 和 B 正相关 ❌ | 这是常见误解。"给定 B 时 A 概率变大"不等于"正相关"——需要看定义 |
| 先验不重要 ❌ | 先验选择影响后验，尤其在数据稀少时。CFA 的行为金融部分专门讨论这个 |
六、CFA 考试应考指南
概念题（高频）： - 条件概率 vs 无条件概率的区别 - 独立性的两种等价定义 - 全概率公式的使用场景 - 贝叶斯公式各部分的名称和含义（先验/似然/后验） - 互斥与独立的区别
计算题（必考）： - 给定的概率表格，求条件概率 - 用全概率公式求总概率 - 贝叶斯公式完整计算（三步：先验 → 全概率 → 后验） - 用 2×2 表格法/树形图法快速求解
七、课后测试题
概念题
Q1： 如果事件 A 和事件 B 互斥（mutually exclusive），以下哪项一定正确？ A. P(A|B) = P(A) B. P(A|B) = 0 C. P(A|B) = P(B)
Q2： 关于贝叶斯公式，分母 P(A) 的作用是： A. 表示事件 A 的先验概率 B. 确保后验概率之和为 1（归一化） C. 是 A 的似然函数
Q3： P(A|B) = P(A) 意味着： A. A 和 B 互斥 B. A 和 B 独立 C. A 是 B 的子集
计算题
Q4： 某公司发行债券。经济好的概率为 60%，经济差为 40%。经济好时违约概率 5%，经济差时违约概率 20%。该债券违约的总概率是多少？ A. 8% B. 11% C. 15%
Q5： 沿用 Q4 数据。如果已知该债券已违约，经济好的概率是多少？（即 P(经济好|违约)） A. 27.3% B. 50.0% C. 60.0%
Q6： 某疾病在人群中的发病率为 1%。检测准确率（真有病时检出阳性）为 95%，假阳性率（没病时检出阳性）为 2%。随机一人检测呈阳性，他真正患病的概率最接近： A. 32% B. 50% C. 95%
Q7： 设 P(A) = 0.4，P(B) = 0.5，P(A∩B) = 0.2。P(A|B) 等于： A. 0.4 B. 0.5 C. 0.8
Q8： 在 Q7 中，A 和 B 是否独立？ A. 独立，因为 P(A|B) = P(A) B. 不独立，因为 P(A∩B) ≠ 0 C. 无法判断
八、答案与解析
A1：B — 如果 A、B 互斥，意味着它们不能同时发生。给定 B 的发生意味着 A 不能发生，所以 P(A|B) = 0。A 错（那是独立的定义），C 错（P(B) 一般不等于 0）。
A2：B — 分母 P(A) = Σ P(A|Bⱼ)·P(Bⱼ) 是全概率，确保 Σ P(Bᵢ|A) = 1（所有后验概率加起来是 1）。A 错（P(A) 是总概率不是先验），C 错（似然是 P(A|Bᵢ)，不是分母）。
A3：B — P(A|B) = P(A) 正是独立的定义：B 的发生没有改变 A 的概率。
A4：B — P(违约) = P(违约|好)×P(好) + P(违约|差)×P(差) = 0.05×0.60 + 0.20×0.40 = 0.03 + 0.08 = 0.11 = 11%
A5：A — 用贝叶斯公式： P(好|违约) = P(违约|好)×P(好) / P(违约) = (0.05×0.60) / 0.11 = 0.03/0.11 ≈ 27.3%
💡 原来的 P(好) = 60%，但看到了违约 → 后验骤降到 27.3%。违约信号很强。
A6：A — 经典贝叶斯： - P(病) = 0.01，P(没病) = 0.99 - P(阳|病) = 0.95，P(阳|没病) = 0.02 - P(阳) = 0.95×0.01 + 0.02×0.99 = 0.0095 + 0.0198 = 0.0293 - P(病|阳) = 0.0095 / 0.0293 ≈ 32.4%
🔴 CFA 经典题： 即使检测出阳性，真患病概率也只有 32%——因为疾病本身极其罕见（1%），假阳性的绝对人数远多于真阳性。
A7：A — P(A|B) = P(A∩B) / P(B) = 0.2 / 0.5 = 0.4。注意 0.4 = P(A)，这正好说明了……
A8：A — P(A|B) = 0.4 = P(A) → 独立。注意 P(A∩B) = 0.2 = 0.4×0.5 = P(A)×P(B)，也验证了独立。
九、今日小结
| 概念 | 一句话 |
|---|---|
| 条件概率 P(A|B) | 在 B 的世界里，A 占多大比例 |
| 乘法法则 | 把联合概率拆成"一步接一步" |
| 独立 | B 发生了，A 的概率不变 |
| 全概率公式 | 把总概率按场景拆开再加权 |
| 贝叶斯公式 | 先验 × 似然 → 归一化 → 后验 |
🧠 三个核心洞察： 1. 条件概率的本质是缩小样本空间——从这个直觉出发，一切公式都是自然的 2. 贝叶斯 = 用证据更新信念——这是投资分析的科学方法 3. 人类直觉不擅长条件概率——画 2×2 表格，把概率变成人数，世界就清晰了
明天 L118 将进入计数原理（排列组合）——这是计算等可能概率场景的基础工具。
本内容仅供学习参考，不构成投资建议。
Quantitative Methods — Probability Module · Lesson 3
1. Lesson Positioning
L116 covered expected value, variance, and covariance — tools to describe where a random variable is centered, how much it fluctuates, and how two variables move together. This lesson introduces the core thinking tool of probability theory: how to update our judgments when new information arrives.
| Item | Description |
|---|---|
| Module | 2.4 Probability Theory |
| Prerequisites | L115 Probability Basics, L116 Expectation & Variance |
| Next Up | L118 Counting Principles (Permutations & Combinations) |
| Difficulty | ★★★★☆ |
| Exam Weight | High (conceptual + computational, tested every year) |
| Reading Time | ~15 minutes |
2. Core Concepts
2.1 Conditional Probability
Intuitive Introduction:
Suppose a company interviews 100 candidates. 20 pass the technical round, and 15 ultimately receive an offer. Among those who received offers, 12 had passed the technical round.
Question: Given that a candidate passed the technical round, what is the probability they received an offer?
Your intuition naturally shrinks the sample space — look only at the 20 who passed the technical round, then see how many of those got offers:
$$P(\text{Offer} \mid \text{Tech Pass}) = \frac{12}{20} = 60\%$$
This is the essence of conditional probability: with new information, redefine the scope, recalculate the proportion.
Formal Definition:
$$P(A \mid B) = \frac{P(A \cap B)}{P(B)}, \quad P(B) > 0$$
Where: - P(A|B): probability of event A occurring, given that event B has already occurred - P(A ∩ B): joint probability of A and B both occurring - P(B): probability of event B (the new "universe")
🧠 Reading: "Probability of A given B" — the condition is on the right of the vertical bar.
Example 1: Professional Certification Exam
Among CFA Level 1 candidates, 60% come from finance backgrounds (F) and 40% from non-finance backgrounds. Pass rates are 70% and 50%, respectively.
| Background | Proportion | Pass Rate |
|---|---|---|
| Finance (F) | 60% | 70% |
| Non-Finance (Fᶜ) | 40% | 50% |
Question: Randomly select someone who passed. What's the probability they have a finance background?
First, compute joint probabilities: - P(F ∩ Pass) = 0.6 × 0.7 = 0.42 - P(Fᶜ ∩ Pass) = 0.4 × 0.5 = 0.20 - P(Pass) = 0.42 + 0.20 = 0.62 (overall pass rate)
$$P(F \mid \text{Pass}) = \frac{P(F \cap \text{Pass})}{P(\text{Pass})} = \frac{0.42}{0.62} \approx 67.7\%$$
📊 Intuition check: Finance candidates have a higher pass rate (70% vs. 50%), so the proportion of finance backgrounds among passers (67.7%) is indeed higher than their proportion in the overall population (60%). Direction checks out.
2.2 Multiplication Rule
Rearranging the conditional probability definition:
$$P(A \cap B) = P(A \mid B) \cdot P(B) = P(B \mid A) \cdot P(A)$$
Extension to three events:
$$P(A \cap B \cap C) = P(A) \cdot P(B \mid A) \cdot P(C \mid A \cap B)$$
📊 Practical application: In structured product default analysis, the multiplication rule is used to decompose joint default probabilities layer by layer.
Example 2: Chain Default Risk
In a portfolio: - P(Firm A defaults) = 5% - If A defaults, P(B defaults | A defaults) = 40% - If both A and B default, P(C defaults | A ∩ B) = 60%
Probability all three default simultaneously:
$$P(A \cap B \cap C) = 5\% \times 40\% \times 60\% = 1.2\%$$
Each individual default probability looks modest, but chain-linked, the joint probability still reaches 1.2% — this is the mathematical essence of systemic risk contagion.
2.3 Formal Definition of Independence
L115 introduced independence intuitively. Now we express it precisely with conditional probability:
Events A and B are independent if and only if:
$$P(A \mid B) = P(A) \quad \text{or equivalently} \quad P(A \cap B) = P(A) \cdot P(B)$$
🧠 Did B change the probability of A? No → the two are independent.
Important distinction: - Mutually Exclusive: P(A ∩ B) = 0 — the two events cannot occur simultaneously - Independent: P(A ∩ B) = P(A) · P(B) — one event occurring does not affect the other
🔴 High-frequency CFA trap: Mutually exclusive ≠ Independent! If A and B are mutually exclusive and both have non-zero probability, they cannot be independent — because given A occurs, B definitely does not occur: P(B|A) = 0 ≠ P(B).
2.4 Total Probability Rule
Motivation: Sometimes computing P(A) directly is hard, but computing P(A|Bᵢ) by "splitting cases" is easy.
Formula: If B₁, B₂, ..., Bₙ form a partition of the sample space (mutually exclusive and exhaustive), then:
$$P(A) = \sum_{i=1}^{n} P(A \mid B_i) \cdot P(B_i)$$
In plain terms: Break the probability of A down by scenarios, then take a weighted sum.
Example 3: Total Probability — Analyst Ratings
An analyst assigns three ratings to a stock:
| Rating | Probability | P(Stock Rises | Rating) | |--------|------------|-------------------| | Buy (B) | 40% | 75% | | Hold (H) | 35% | 50% | | Sell (S) | 25% | 20% |
Question: Regardless of rating, what is the overall probability the stock rises?
$$P(\text{Rise}) = P(\text{Rise} \mid B) \cdot P(B) + P(\text{Rise} \mid H) \cdot P(H) + P(\text{Rise} \mid S) \cdot P(S)$$
$$= 0.75 \times 0.40 + 0.50 \times 0.35 + 0.20 \times 0.25$$
$$= 0.30 + 0.175 + 0.05 = 0.525 = 52.5\%$$
📊 The total probability rule takes the "weighted average" idea from returns and applies it to probabilities — scenario probability × conditional probability, sum it all up.
2.5 Bayes' Formula — The Core of This Lesson
One of the most elegant and powerful tools in the CFA quantitative methods toolkit.
Formula:
$$P(B_i \mid A) = \frac{P(A \mid B_i) \cdot P(B_i)}{\sum_{j=1}^{n} P(A \mid B_j) \cdot P(B_j)}$$
Breaking down the structure:
| Part | Name | Meaning |
|---|---|---|
| P(Bᵢ) | Prior Probability | Our judgment about Bᵢ before seeing new evidence |
| P(A|Bᵢ) | Likelihood | How likely we are to observe A if Bᵢ is true |
| Denominator | Normalizing Constant | Total probability of observing A across all possibilities |
| P(Bᵢ|A) | Posterior Probability | Our updated judgment about Bᵢ after seeing A |
$$Posterior = \frac{Likelihood \times Prior}{Total\;Probability}$$
🧠 In one sentence: Bayes' formula tells you how to update old beliefs with new evidence.
Example 4: Bayes — Fund Manager Skill Assessment (CFA Classic)
A fund manager claims they can consistently beat the market.
- Prior judgment: only 10% of managers in the industry are truly skilled (Genuine, G); 90% rely on luck (Luck, L)
- Skilled manager beats the market in a given year: 80% probability
- Lucky manager beats the market in a given year: 50% probability (random market)
Question: If a manager beat the market this year, what is the probability they are genuinely skilled?
Step 1: Organize known information - P(G) = 0.10 (prior: truly skilled) - P(L) = 0.90 (prior: just lucky) - P(Win | G) = 0.80 - P(Win | L) = 0.50
Step 2: Compute total probability (denominator) $$P(\text{Win}) = 0.80 \times 0.10 + 0.50 \times 0.90$$ $$= 0.08 + 0.45 = 0.53$$
Step 3: Apply Bayes' formula $$P(G \mid \text{Win}) = \frac{0.80 \times 0.10}{0.53} = \frac{0.08}{0.53} \approx 15.1\%$$
📊 Striking insight: Even after beating the market, the probability of genuine skill only rises from 10% to 15.1%! This is because the probability of winning by luck (50%) is also quite high, and lucky managers are vastly more numerous (90%).
Example 5: Bayes — Insider Trading Detection
The SEC investigates whether a trader engaged in insider trading (Insider, I).
- P(I) = 5% (prior: only 5% of traders engage in insider trading)
- If insider trading: P(abnormal profit) = 95%
- If no insider trading: P(abnormal profit) = 2% (rare event)
The trader shows abnormal profits. Posterior probability of insider trading:
$$P(I \mid \text{Abnormal}) = \frac{0.95 \times 0.05}{0.95 \times 0.05 + 0.02 \times 0.95}$$
$$= \frac{0.0475}{0.0475 + 0.019} = \frac{0.0475}{0.0665} \approx 71.4\%$$
📊 From 5% to 71.4% — because "abnormal profits without insider trading" is extremely rare (2%). Observing abnormal profits causes a dramatic jump in the posterior. This is where Bayes' updating is most powerful: the larger the likelihood ratio (95%/2% = 47.5×), the more dramatic the update.
2.6 Understanding Bayes Through "Natural Frequencies"
The human brain is bad at probabilities but good at counting heads.
Translate Example 4 into natural frequencies (assume 1,000 fund managers):
| Type | Count | % Who Beat Market | # Who Beat Market |
|---|---|---|---|
| Skilled (G) | 100 | 80% | 80 |
| Lucky (L) | 900 | 50% | 450 |
| Total | 1,000 | — | 530 |
💡 530 people beat the market. 80 of them are genuinely skilled. So: $$P(G \mid \text{Win}) = \frac{80}{530} \approx 15.1\%$$
Exactly matches Bayes' formula. The best mental shortcut for Bayes problems: draw a 2×2 table!
3. Key Formula Summary
| Concept | Formula | Keyword |
|---|---|---|
| Conditional Probability | P(A|B) = P(A∩B) / P(B) | Shrink sample space |
| Multiplication Rule | P(A∩B) = P(A|B)·P(B) | Chain decomposition |
| Independence Test | P(A|B) = P(A) or P(A∩B)=P(A)·P(B) | B doesn't change A |
| Total Probability | P(A) = Σ P(A|Bᵢ)·P(Bᵢ) | Weight by scenario |
| Bayes' Formula | P(Bᵢ|A) = P(A|Bᵢ)·P(Bᵢ) / P(A) | Prior → Posterior |
4. Bayes Updating: CFA Application Scenarios
| Domain | Application |
|---|---|
| Investment Analysis | New earnings report → update judgment on company quality |
| Risk Management | Observe default signal → update default probability |
| Technical Analysis | See breakout pattern → update trend probability |
| Behavioral Finance | Overreaction vs. conservatism bias — anchoring on priors, insufficient updating |
| Asset Pricing | Dynamic belief updating with multi-stage information arrival |
5. Common Pitfalls
| Pitfall | Correct Understanding |
|---|---|
| P(A|B) = P(B|A) ❌ | Conditional probabilities cannot be casually swapped! They are generally not equal. |
| Mutually exclusive = independent ❌ | Mutually exclusive means P(A∩B)=0; independence requires P(A∩B)=P(A)·P(B). They are incompatible when both P>0. |
| Forgetting total probability in Bayes denominator ❌ | The denominator is P(A) = Σ P(A|Bⱼ)·P(Bⱼ), not just any number. |
| P(A|B) > P(A) means A and B are positively correlated ❌ | Common misunderstanding. Need to examine the definitions carefully. |
| Priors don't matter ❌ | Prior choice affects the posterior, especially when data is scarce. CFA's behavioral finance section covers this specifically. |
6. CFA Exam Guide
Conceptual Questions (high frequency): - Conditional probability vs. unconditional probability - Two equivalent definitions of independence - When to use the total probability rule - Names and meanings of Bayes formula components (prior/likelihood/posterior) - Difference between mutually exclusive and independent
Computational Questions (must-know): - Find conditional probability from a given probability table - Use total probability rule to find overall probability - Complete Bayes' formula calculation (three steps: prior → total probability → posterior) - Use 2×2 table method / tree diagram method for quick solving
7. Practice Questions
Conceptual
Q1: If events A and B are mutually exclusive, which of the following must be true? A. P(A|B) = P(A) B. P(A|B) = 0 C. P(A|B) = P(B)
Q2: In Bayes' formula, the role of the denominator P(A) is to: A. Represent the prior probability of event A B. Ensure posterior probabilities sum to 1 (normalization) C. Serve as the likelihood function of A
Q3: P(A|B) = P(A) implies that: A. A and B are mutually exclusive B. A and B are independent C. A is a subset of B
Computational
Q4: A company issues bonds. P(Good economy) = 60%, P(Bad economy) = 40%. Default probability in good economy = 5%, in bad economy = 20%. What is the total probability of default? A. 8% B. 11% C. 15%
Q5: Using Q4 data. Given the bond has defaulted, what is P(Good economy | Default)? A. 27.3% B. 50.0% C. 60.0%
Q6: A disease has a 1% prevalence rate in the population. Test accuracy (positive given disease) = 95%. False positive rate (positive given no disease) = 2%. A random person tests positive. The probability they actually have the disease is closest to: A. 32% B. 50% C. 95%
Q7: Given P(A) = 0.4, P(B) = 0.5, P(A∩B) = 0.2. P(A|B) equals: A. 0.4 B. 0.5 C. 0.8
Q8: In Q7, are A and B independent? A. Independent, because P(A|B) = P(A) B. Not independent, because P(A∩B) ≠ 0 C. Cannot determine
8. Answers & Explanations
A1: B — If A and B are mutually exclusive, they cannot occur simultaneously. Given B has occurred, A cannot occur, so P(A|B) = 0. A is wrong (that's the definition of independence), C is wrong (P(B) is generally not 0).
A2: B — The denominator P(A) = Σ P(A|Bⱼ)·P(Bⱼ) is the total probability, ensuring Σ P(Bᵢ|A) = 1 (all posterior probabilities sum to 1). A is wrong (P(A) is total probability, not prior), C is wrong (likelihood is P(A|Bᵢ), not the denominator).
A3: B — P(A|B) = P(A) is precisely the definition of independence: B's occurrence does not change the probability of A.
A4: B — P(Default) = P(Default|Good)×P(Good) + P(Default|Bad)×P(Bad) = 0.05×0.60 + 0.20×0.40 = 0.03 + 0.08 = 0.11 = 11%
A5: A — Apply Bayes formula: P(Good|Default) = P(Default|Good)×P(Good) / P(Default) = (0.05×0.60) / 0.11 = 0.03/0.11 ≈ 27.3%
💡 Original P(Good) = 60%, but seeing a default → posterior drops to 27.3%. The default signal is strong.
A6: A — Classic Bayes: - P(Disease) = 0.01, P(No Disease) = 0.99 - P(Positive|Disease) = 0.95, P(Positive|No Disease) = 0.02 - P(Positive) = 0.95×0.01 + 0.02×0.99 = 0.0095 + 0.0198 = 0.0293 - P(Disease|Positive) = 0.0095 / 0.0293 ≈ 32.4%
🔴 CFA classic: Even with a positive test, the probability of actually having the disease is only ~32% — because the disease is extremely rare (1%) and false positives outnumber true positives in absolute terms.
A7: A — P(A|B) = P(A∩B) / P(B) = 0.2 / 0.5 = 0.4. Note that 0.4 = P(A), which conveniently leads to...
A8: A — P(A|B) = 0.4 = P(A) → independent. Verify: P(A∩B) = 0.2 = 0.4×0.5 = P(A)×P(B). ✓
9. Today's Summary
| Concept | One Line |
|---|---|
| Conditional Probability P(A|B) | In the world of B, what proportion is A? |
| Multiplication Rule | Break joint probability into step-by-step links |
| Independence | B happened, A's probability unchanged |
| Total Probability Rule | Split total probability by scenarios, then weighted sum |
| Bayes' Formula | Prior × Likelihood → Normalize → Posterior |
🧠 Three core insights: 1. The essence of conditional probability is shrinking the sample space — from this intuition, all formulas follow naturally 2. Bayes = updating beliefs with evidence — this is the scientific method of investment analysis 3. Human intuition is bad at conditional probability — draw a 2×2 table, turn probabilities into head counts, and the world becomes clear
Tomorrow L118 enters counting principles (permutations & combinations) — foundational tools for computing probabilities in equally-likely-outcome scenarios.
This content is for educational reference only and does not constitute investment advice.