核心思想

极端风险管理关注的是概率极低但影响极大的事件（肥尾、黑天鹅）。传统 VaR 假设正态分布时会严重低估这类风险。

肥尾建模

金融收益率通常服从幂律尾部（Pareto distribution）：

$P(|R| > x) \sim C \cdot x^{-\alpha}$

其中尾部指数 $\alpha$ 通常在 2~5 之间（正态分布 $\alpha = \infty$）。

GPD — Generalized Pareto Distribution

对超过阈值的超额损失建模：

$F(y) = 1 - \left(1 + \xi \frac{y}{\sigma}\right)^{-1/\xi}$

其中 $\xi$ 为形状参数，$\sigma$ 为尺度参数。$\xi > 0$ 表示厚尾。

import numpy as np
from scipy import stats

def gpd_fitting(returns, threshold=0.95):
    """超额损失拟合 GPD"""
    excesses = returns[returns < -threshold] - (-threshold)
    
    # MLE 估计 GPD 参数
    def neg_log_likelihood(params):
        mu, sigma, xi = params
        if sigma <= 0 or (1 + xi * excesses / sigma) <= 0:
            return np.inf
        ll = -np.sum(np.log(1/sigma) * (1 + xi * excesses / sigma)**(-1/xi - 1))
        return ll
    
    from scipy.optimize import minimize
    result = minimize(neg_log_likelihood, x0=[0, 0.05, 0.3], 
                      bounds=[(None, None), (1e-6, None), (-0.5, 0.5)])
    
    mu_hat, sigma_hat, xi_hat = result.x
    
    # VaR 估计
    def gpd_var(prob):
        if xi_hat != 0:
            return -mu_hat + (sigma_hat/xi_hat) * ((prob / n_excess)**(-xi_hat) - 1)
        else:
            return -mu_hat + sigma_hat * np.log(prob / n_excess)
    
    return {
        'mu': mu_hat, 'sigma': sigma_hat, 'xi': xi_hat,
        'VaR_99': gpd_var(0.01),
        'CVaR_99': gpd_var(0.005),
    }

# 峰值过阈 (Peaks Over Threshold)
def POT_analysis(returns, percentile=90):
    """POT 分析"""
    threshold = np.percentile(returns, percentile)
    excesses = returns[returns < -threshold] + threshold  # 负值表示超额损失
    return excesses

系统性风险度量

CoVaR — Conditional VaR

Adrian & Brunnermeier (2016) 定义 CoVaR 为：当某机构陷入困境时，整个金融系统的 VaR。

$\text{CoVaR}_\alpha = \text{VaR}_\alpha(R_{\text{system}} \mid R_i = \text{VaR}_{\beta}(R_i))$

$\Delta$CoVaR — 边际贡献

$\Delta\text{CoVaR} = \text{CoVaR}_{\text{institution}_i} - \text{CoVaR}_{\text{median}(R_i)}$

def delta_covar(X_system, X_institution, alpha=0.05):
    """计算单个机构对系统性风险的边际贡献"""
    # 排序 institutional returns
    sorted_idx = np.argsort(X_institution)
    
    # CoVaR when institution is at VaR quantile
    var_idx = int(alpha * len(X_institution))
    co_var_at_crash = np.percentile(X_system[sorted_idx[:var_idx]], alpha)
    
    # CoVaR at median institutional performance
    co_var_at_median = np.percentile(X_system, alpha)
    
    return co_var_at_crash - co_var_at_median

压力测试框架

def stress_test(portfolio, scenarios):
    """多维度压力测试"""
    results = {}
    for name, params in scenarios.items():
        # 冲击因子
        price_shock = params.get('price_shock', 0)
        vol_multiplier = params.get('vol_multiplier', 1)
        corr_shift = params.get('corr_shift', 0)
        
        # 计算冲击后组合价值
        stressed_returns = portfolio.returns * price_shock + \
                          np.random.randn(*portfolio.returns.shape) * vol_multiplier
        
        results[name] = {
            'pnl': stressed_returns.sum(),
            'var_95': -np.percentile(stressed_returns.sum(axis=1), 5),
            'max_drawdown': max_drawdown(stressed_returns.cumsum()),
        }
    return results

局限

• ⚠️ GPD 阈值选择高度主观
• ⚠️ CoVaR 估计需要大量高频数据
• ⚠️ 黑天鹅本质上是不可预测的

变体

参考文献

1. McNeil, A.J. & Frey, R., 2000. “Estimation of Tail-Related Risk Measures for Heteroskedastic Financial Series.”
2. Adrian, T. & Brunnermeier, M.K., 2016. “CoVaR.” American Economic Review.
3. Embrechts, P., Kluppelberg, C. & Mikosch, T., 1997. Modelling Extremal Events for Insurance and Finance.