概率重尾稳定分布

Prob Heavy Tail Stable Distribution

专题: Probability / 概率
难度: L3
来源: MyntBit

个人笔记

题目详情

团队使用的模型假设资产回报遵循稳定分布，特征指数 α < 2。稳定分布由四个参数表征：位置、尺度、偏度和特征指数。

任务：当 α < 2 时，分布的方差（二阶矩）不存在（无限），这与正态分布（α=2）形成鲜明对比。这意味着尾部衰减足够慢，使得 E[X²] 发散。解释这对风险管理的影响：使用标准差衡量风险会低估极端事件的概率。

英文原题

Your team employs a model where asset returns are assumed to follow a stable distribution with index $\alpha < 2$ . Recall that stable distributions are characterized by four parameters: location, scale, skewness and index (also called characteristic exponent). What fundamental assumption of Modern Portfolio Theory (MPT) is most significantly violated when asset returns follow such a distribution?

解析

问题分析

金融收益率呈现明显的厚尾特征——极端事件发生频率远高于正态分布预测。稳定分布（如 Levy α-stable）能更好地拟合厚尾数据，但除少数特例外没有闭合形式的 PDF。

核心特性

特征函数：φ(t) = exp(itμ - |γt|^α * (1 - iβ sign(t) tan(πα/2)))
α ∈ (0,2]：尾部指数，越小尾部越厚（α=2 即正态）
β ∈ [-1,1]：偏度参数
只有 α=2（正态）、α=1,β=0（柯西）、α=0.5,β=1（Lévy）有闭合 PDF

double mean(const std::vector<double>& v) { return accumulate(v.begin(),v.end(),0.0)/v.size(); }
double stddev(const std::vector<double>& v, double m) {
    double sq=0; for(double x:v) sq+=(x-m)*(x-m); return sqrt(sq/v.size()); }

复杂度与边界

时间复杂度：模拟通过 Chambers-Mallows-Stuck 方法，每次 O(1)
空间复杂度：O(1)
边界条件：(1) α=1 时使用特殊公式避免除零 (2) 方差在 α<2 时无穷大 (3) 均值在 α≤1 时不存在

英文解析

Analysis

Financial returns exhibit pronounced heavy-tail characteristics - extreme events occur far more frequently than normal distribution predictions. Stable distributions (such as Levy alpha-stable) better fit heavy-tailed data, but except for a few special cases, they have no closed-form PDF.

Solution

// Chambers-Mallows-Stuck method for generating alpha-stable random variables
double generateStable(double alpha, double beta, double gamma, double delta) {
    double V = M_PI * (uniform_random() - 0.5);  // Uniform on (-pi/2, pi/2)
    double W = -std::log(uniform_random());       // Exponential
    if (alpha == 1.0) {
        // Cauchy-like special case
        return delta + gamma * (2.0/M_PI) * ((1.0 + beta * V) * std::tan(V));
    }
    double S = std::pow(std::cos((1-alpha)*V) / W, (1-alpha)/alpha)
               * std::sin(alpha * V) / std::pow(std::cos(V), 1.0/alpha);
    double B = std::atan(beta * std::tan(alpha * M_PI / 2.0));
    return delta + gamma * S;
}

Complexity & Edge Cases

Time complexity: O(1) per sample via Chambers-Mallows-Stuck method
Space complexity: O(1)
Edge cases: (1) alpha=1 uses special formula to avoid division by zero (2) Variance is infinite when alpha < 2 (3) Mean does not exist when alpha <= 1

Verification

Generate samples with alpha=2 (normal), verify mean and variance match. Test alpha=1, beta=0 (Cauchy) - verify heavy tails. Benchmark sample kurtosis against normal for alpha=1.5.

Key Considerations

Alpha-stable distributions are the theoretical model for financial return distributions. The tail index alpha < 2 means variance is undefined - standard risk metrics like standard deviation and Sharpe ratio are statistically meaningless for heavy-tailed returns. In practice, alpha for equity returns is estimated around 1.5-1.7, meaning extreme events are 10-100x more likely than normal distribution predicts.