agents24 量化交易插件中的风险指标计算VaR、CVaR、回撤与压力测试的完整 Python 实现指南【免费下载链接】agentsMulti-harness agentic plugin marketplace for Claude Code, Codex, Cursor, OpenCode, GitHub Copilot, and Google Antigravity项目地址: https://gitcode.com/GitHub_Trending/agents24/agents本篇技术指南围绕 agents24 仓库中quantitative-trading插件的risk-metrics-calculation技能展开完整讲解单资产风险度量、组合风险分解、滚动风险监控与压力测试四大实现模式的 Python 代码细节。读完本文你将掌握基于 NumPy / pandas / SciPy 编写一套可用于组合管理、风险限额、风险看板与监管报送的可复用风险计算工具箱并理解每个指标的年化逻辑、边界条件与适用场景。本文的全部代码与配置均来自 details.md并与同插件的 SKILL.md、risk-manager.md 等仓库文档相互印证。一、技能定位这个风险指标库解决什么问题risk-metrics-calculation是 agents24 仓库中quantitative-trading插件下的一个 Skill其定位是组合管理的综合性风险度量工具箱覆盖 Value at RiskVaR、Expected ShortfallCVaR与回撤分析。对应 Skill 的元信息描述为Calculate portfolio risk metrics including VaR, CVaR, Sharpe, Sortino, and drawdown analysis. Use when measuring portfolio risk, implementing risk limits, or building risk monitoring systems.该技能在 SKILL.md 中明确了典型使用场景度量组合风险Measuring portfolio risk实施风险限额Implementing risk limits构建风险看板Building risk dashboards计算风险调整后收益Calculating risk-adjusted returns设定仓位大小Setting position sizes监管报送Regulatory reporting在 agent 层该插件提供两个角色配合使用quant-analyst.mdopus负责金融建模、交易策略与风险指标计算与 risk-manager.mdsonnet负责组合风险监控、R-multiple 跟踪与头寸限额。risk-manager 的职责清单中明确包含Value at Risk (VaR) calculationsCorrelation and beta analysisStress testing and scenario analysisRisk-adjusted performance metrics与本文要实现的指标一一对应。安装该插件的方式为见 docs/plugins.md/plugin install quantitative-trading下文将逐一展开 details.md 中四个实现模式核心风险指标、组合风险、滚动风险指标与压力测试最后给出快速使用参考与最佳实践清单。二、风险指标的分类框架与时间尺度继承 SKILL.md 核心概念在进入代码之前先建立分类框架。SKILL.md 将风险指标划分为四个类别类别Category指标Metrics用途Use Case波动率Volatility标准差Std Dev、Beta一般风险度量尾部风险Tail RiskVaR、CVaR极端亏损度量回撤Drawdown最大回撤Max DD、Calmar资本保全风险调整收益Risk-AdjustedSharpe、Sortino绩效评估此外指标的时间尺度应与投资决策周期匹配Intraday: Minute/hourly VaR for day traders日内分钟/小时级 VaR用于日内交易者 Daily: Standard risk reporting日频标准风险报告 Weekly: Rebalancing decisions周频再平衡决策 Monthly: Performance attribution月频业绩归因 Annual: Strategic allocation年频战略配置代码中默认的年化因子ann_factor 252即建立在日频收益、一年 252 个交易日的假设之上——这是整套实现最重要的前提假设涉及任何年化计算时都应显式确认。三、Pattern 1单资产核心风险指标RiskMetrics 类RiskMetrics是整套风险库的地基负责对单一收益序列计算波动率、VaR/CVaR、回撤与风险调整收益。构造函数如下import numpy as np import pandas as pd from scipy import stats from typing import Dict, Optional, Tuple class RiskMetrics: Core risk metric calculations. def __init__(self, returns: pd.Series, rf_rate: float 0.02): Args: returns: Series of periodic returns rf_rate: Annual risk-free rate self.returns returns self.rf_rate rf_rate self.ann_factor 252 # Trading days per year参数说明returns周期收益序列pandas Series默认按日频收益假设处理若传入周频或月频数据需同步调整ann_factor。rf_rate年化无风险利率默认 2%用于 Sharpe / Sortino 比率的超额收益计算。ann_factor年化交易日数默认 252是全部年化计算的基准。3.1 波动率类指标volatility、downside_deviation、beta# Volatility Metrics def volatility(self, annualized: bool True) - float: Standard deviation of returns. vol self.returns.std() if annualized: vol * np.sqrt(self.ann_factor) return vol def downside_deviation(self, threshold: float 0, annualized: bool True) - float: Standard deviation of returns below threshold. downside self.returns[self.returns threshold] if len(downside) 0: return 0.0 dd downside.std() if annualized: dd * np.sqrt(self.ann_factor) return dd def beta(self, market_returns: pd.Series) - float: Beta relative to market. aligned pd.concat([self.returns, market_returns], axis1).dropna() if len(aligned) 2: return np.nan cov np.cov(aligned.iloc[:, 0], aligned.iloc[:, 1]) return cov[0, 1] / cov[1, 1] if cov[1, 1] ! 0 else 0实现要点volatility使用returns.std()样本标准差ddof1年化时乘sqrt(252)即日频波动率平方根法则放大。若不年化可传annualizedFalse得到周期波动率。downside_deviation只取低于阈值的样本计算标准差是 Sortino 比率的输入。注意边界处理当没有任何收益低于阈值时返回0.0避免 NaN 污染后续比率。beta通过pd.concat对齐两序列并dropna()用np.cov计算协方差矩阵beta Cov(p, m) / Var(m)。边界处理样本不足 2 个返回np.nan市场方差为 0 时返回 0。3.2 尾部风险三种 VaR 与 CVaR# Value at Risk def var_historical(self, confidence: float 0.95) - float: Historical VaR at confidence level. return -np.percentile(self.returns, (1 - confidence) * 100) def var_parametric(self, confidence: float 0.95) - float: Parametric VaR assuming normal distribution. z_score stats.norm.ppf(confidence) return self.returns.mean() - z_score * self.returns.std() def var_cornish_fisher(self, confidence: float 0.95) - float: VaR with Cornish-Fisher expansion for non-normality. z stats.norm.ppf(confidence) s stats.skew(self.returns) # Skewness k stats.kurtosis(self.returns) # Excess kurtosis # Cornish-Fisher expansion z_cf (z (z**2 - 1) * s / 6 (z**3 - 3*z) * k / 24 - (2*z**3 - 5*z) * s**2 / 36) return -(self.returns.mean() z_cf * self.returns.std()) # Conditional VaR (Expected Shortfall) def cvar(self, confidence: float 0.95) - float: Expected Shortfall / CVaR / Average VaR. var self.var_historical(confidence) return -self.returns[self.returns -var].mean()三种 VaR 的差异与适用场景历史模拟法historical直接取收益分布的分位数取负np.percentile(returns, (1-confidence)*100)即左侧尾部下分位点。无分布假设稳健直观是默认推荐。参数法parametric假设收益服从正态分布用stats.norm.ppf(confidence)取标准正态分位数 zVaR mean - z * std。计算快但对肥尾分布会系统性低估风险。Cornish-Fisher 展开在正态分位数基础上用偏度s和超额峰度k修正z_cf公式为z (z²-1)s/6 (z³-3z)k/24 - (2z³-5z)s²/36。这是文档与代码注释反复强调的关键点——金融收益具有肥尾特征不要默认假设正态性见 SKILL.md 的 Donts 清单。CVaR / 期望损失Expected Shortfall取所有亏损超过历史 VaR 阈值的样本均值再取负衡量最坏 α 尾部内的平均损失。它是比 VaR 更完备的尾部风险度量——VaR 只回答阈值在哪CVaR 回答一旦突破阈值平均亏多少。3.3 回撤分析drawdowns、max_drawdown、avg_drawdown、drawdown_duration# Drawdown Analysis def drawdowns(self) - pd.Series: Calculate drawdown series. cumulative (1 self.returns).cumprod() running_max cumulative.cummax() return (cumulative - running_max) / running_max def max_drawdown(self) - float: Maximum drawdown. return self.drawdowns().min() def avg_drawdown(self) - float: Average drawdown. dd self.drawdowns() return dd[dd 0].mean() if (dd 0).any() else 0 def drawdown_duration(self) - Dict[str, int]: Drawdown duration statistics. dd self.drawdowns() in_drawdown dd 0 # Find drawdown periods drawdown_starts in_drawdown ~in_drawdown.shift(1).fillna(False) drawdown_ends ~in_drawdown in_drawdown.shift(1).fillna(False) durations [] current_duration 0 for i in range(len(dd)): if in_drawdown.iloc[i]: current_duration 1 elif current_duration 0: durations.append(current_duration) current_duration 0 if current_duration 0: durations.append(current_duration) return { max_duration: max(durations) if durations else 0, avg_duration: np.mean(durations) if durations else 0, current_duration: current_duration }实现要点drawdowns先计算累计净值cumprod再用cummax得到历史峰值序列回撤 (净值 - 峰值) / 峰值。结果为 ≤ 0 的序列。max_drawdown取回撤序列最小值最深的负回撤。avg_drawdown只对回撤为负的区间求均值若从未出现回撤返回 0避免mean()对空切片告警。drawdown_duration通过逐日扫描统计回撤持续期返回max_duration最长回撤天数、avg_duration平均回撤天数与current_duration当前正处于回撤中的天数。该统计对资本保全与心理预期管理非常实用——很多策略回撤深度可控但回撤时间过长。3.4 风险调整收益Sharpe、Sortino、Calmar、Omega# Risk-Adjusted Returns def sharpe_ratio(self) - float: Annualized Sharpe ratio. excess_return self.returns.mean() * self.ann_factor - self.rf_rate vol self.volatility(annualizedTrue) return excess_return / vol if vol 0 else 0 def sortino_ratio(self) - float: Sortino ratio using downside deviation. excess_return self.returns.mean() * self.ann_factor - self.rf_rate dd self.downside_deviation(threshold0, annualizedTrue) return excess_return / dd if dd 0 else 0 def calmar_ratio(self) - float: Calmar ratio (return / max drawdown). annual_return (1 self.returns).prod() ** (self.ann_factor / len(self.returns)) - 1 max_dd abs(self.max_drawdown()) return annual_return / max_dd if max_dd 0 else 0 def omega_ratio(self, threshold: float 0) - float: Omega ratio. returns_above self.returns[self.returns threshold] - threshold returns_below threshold - self.returns[self.returns threshold] if returns_below.sum() 0: return np.inf return returns_above.sum() / returns_below.sum()四个比率的口径差异Sharpe超额收益年化平均收益减无风险利率÷ 年化总波动率。波动率用全样本标准差对上行波动与下行波动一视同仁。Sortino分子与 Sharpe 相同但分母换成下行偏差downside_deviation(threshold0)只惩罚亏损波动更适合收益分布不对称的策略。Calmar年化收益 ÷ |最大回撤|衡量每承受 1 单位最深回撤能换来多少年化收益是趋势型/CTA 策略常用指标。注意其年化收益使用复利几何平均(1r).prod() ** (ann_factor/n) - 1与 Sharpe 中使用的算术年化mean * ann_factor口径不同。Omega阈值之上收益总和 ÷ 阈值之下亏损总和不依赖收益分布形态。当没有下行收益时返回np.inf纯赚不亏的理想情形。3.5 信息比率与综合摘要# Information Ratio def information_ratio(self, benchmark_returns: pd.Series) - float: Information ratio vs benchmark. active_returns self.returns - benchmark_returns tracking_error active_returns.std() * np.sqrt(self.ann_factor) active_return active_returns.mean() * self.ann_factor return active_return / tracking_error if tracking_error 0 else 0 # Summary def summary(self) - Dict[str, float]: Generate comprehensive risk summary. dd_stats self.drawdown_duration() return { # Returns total_return: (1 self.returns).prod() - 1, annual_return: (1 self.returns).prod() ** (self.ann_factor / len(self.returns)) - 1, # Volatility annual_volatility: self.volatility(), downside_deviation: self.downside_deviation(), # VaR CVaR var_95_historical: self.var_historical(0.95), var_99_historical: self.var_historical(0.99), cvar_95: self.cvar(0.95), # Drawdowns max_drawdown: self.max_drawdown(), avg_drawdown: self.avg_drawdown(), max_drawdown_duration: dd_stats[max_duration], # Risk-Adjusted sharpe_ratio: self.sharpe_ratio(), sortino_ratio: self.sortino_ratio(), calmar_ratio: self.calmar_ratio(), omega_ratio: self.omega_ratio(), # Distribution skewness: stats.skew(self.returns), kurtosis: stats.kurtosis(self.returns), }information_ratio计算相对基准的超额收益与跟踪误差之比主动收益 组合收益 − 基准收益跟踪误差 主动收益标准差 × √252。这是衡量主动管理能力的标准指标。summary()一次性输出 14 项指标涵盖收益、波动、VaR/CVaR、回撤、风险调整收益与分布形态偏度、超额峰度可直接作为风险看板的数据源。注意 Cornish-Fisher 修正所需的偏度与峰度也被一并输出方便交叉核验尾部假设是否成立。四、Pattern 2组合层风险分解PortfolioRisk 类单资产指标无法回答多资产组合整体风险多大、每只资产贡献多少的问题PortfolioRisk负责组合层计算。class PortfolioRisk: Portfolio-level risk calculations. def __init__( self, returns: pd.DataFrame, weights: Optional[pd.Series] None ): Args: returns: DataFrame with asset returns (columns assets) weights: Portfolio weights (default: equal weight) self.returns returns self.weights weights if weights is not None else \ pd.Series(1/len(returns.columns), indexreturns.columns) self.ann_factor 252 def portfolio_return(self) - float: Weighted portfolio return. return (self.returns self.weights).mean() * self.ann_factor def portfolio_volatility(self) - float: Portfolio volatility. cov_matrix self.returns.cov() * self.ann_factor port_var self.weights cov_matrix self.weights return np.sqrt(port_var) def marginal_risk_contribution(self) - pd.Series: Marginal contribution to risk by asset. cov_matrix self.returns.cov() * self.ann_factor port_vol self.portfolio_volatility() # Marginal contribution mrc (cov_matrix self.weights) / port_vol return mrc def component_risk(self) - pd.Series: Component contribution to total risk. mrc self.marginal_risk_contribution() return self.weights * mrc def risk_parity_weights(self, target_vol: float None) - pd.Series: Calculate risk parity weights. from scipy.optimize import minimize n len(self.returns.columns) cov_matrix self.returns.cov() * self.ann_factor def risk_budget_objective(weights): port_vol np.sqrt(weights cov_matrix weights) mrc (cov_matrix weights) / port_vol rc weights * mrc target_rc port_vol / n # Equal risk contribution return np.sum((rc - target_rc) ** 2) constraints [ {type: eq, fun: lambda w: np.sum(w) - 1}, # Weights sum to 1 ] bounds [(0.01, 1.0) for _ in range(n)] # Min 1%, max 100% x0 np.array([1/n] * n) result minimize( risk_budget_objective, x0, methodSLSQP, boundsbounds, constraintsconstraints ) return pd.Series(result.x, indexself.returns.columns) def correlation_matrix(self) - pd.DataFrame: Asset correlation matrix. return self.returns.corr() def diversification_ratio(self) - float: Diversification ratio (higher more diversified). asset_vols self.returns.std() * np.sqrt(self.ann_factor) weighted_vol (self.weights * asset_vols).sum() port_vol self.portfolio_volatility() return weighted_vol / port_vol if port_vol 0 else 1 def tracking_error(self, benchmark_returns: pd.Series) - float: Tracking error vs benchmark. port_returns self.returns self.weights active_returns port_returns - benchmark_returns return active_returns.std() * np.sqrt(self.ann_factor) def conditional_correlation( self, threshold_percentile: float 10 ) - pd.DataFrame: Correlation during stress periods. port_returns self.returns self.weights threshold np.percentile(port_returns, threshold_percentile) stress_mask port_returns threshold return self.returns[stress_mask].corr()实现要点逐项拆解组合收益用矩阵乘法returns weights得到组合日收益序列再取均值年化。权重缺省时等权分配1/n。组合波动率先算年化协方差矩阵cov * 252再用二次型wᵀΣw求方差、开方得波动率。这是 Markowitz 组合理论的核心公式。边际风险贡献MRCΣw / σp表示每增加 1 单位某资产权重引起的组合波动率边际变化。成分风险Component Riskwᵢ × MRCᵢ各资产成分风险之和恰好等于组合波动率欧拉分解可直接回答组合风险中每只资产占了多少。风险平价权重Risk Parity通过scipy.optimize.minimize以 SLSQP 求解目标函数为最小化各资产风险贡献与等风险贡献目标σp/n的平方偏差约束权重和为 1边界为每只资产 1%100%。这是每个资产贡献相同风险的经典风险预算方案与 risk-manager 关注的相关性集中风险Monitor correlations to avoid concentration直接相关。分散化比率Diversification Ratio加权单资产波动率之和 ÷ 组合波动率比值越高说明分散化越有效当组合波动率为 0 时返回 1 兜底。条件相关性Conditional Correlation取组合收益最低 10% 分位的压力期样本单独计算相关性矩阵。这对应 SKILL.md 中Dont ignore correlation - Increases in stress的警示——正常市况下相关性低但压力期相关性会飙升用压力期相关性做风险管理比全样本相关性更保守。五、Pattern 3滚动窗口风险监控RollingRiskMetrics 类风险是时变的静态全样本指标会掩盖 regime波动率状态切换。RollingRiskMetrics用滚动窗口把风险指标变成时间序列。class RollingRiskMetrics: Rolling window risk calculations. def __init__(self, returns: pd.Series, window: int 63): Args: returns: Return series window: Rolling window size (default: 63 ~3 months) self.returns returns self.window window def rolling_volatility(self, annualized: bool True) - pd.Series: Rolling volatility. vol self.returns.rolling(self.window).std() if annualized: vol * np.sqrt(252) return vol def rolling_sharpe(self, rf_rate: float 0.02) - pd.Series: Rolling Sharpe ratio. rolling_return self.returns.rolling(self.window).mean() * 252 rolling_vol self.rolling_volatility() return (rolling_return - rf_rate) / rolling_vol def rolling_var(self, confidence: float 0.95) - pd.Series: Rolling historical VaR. return self.returns.rolling(self.window).apply( lambda x: -np.percentile(x, (1 - confidence) * 100), rawTrue ) def rolling_max_drawdown(self) - pd.Series: Rolling maximum drawdown. def max_dd(returns): cumulative (1 returns).cumprod() running_max cumulative.cummax() drawdowns (cumulative - running_max) / running_max return drawdowns.min() return self.returns.rolling(self.window).apply(max_dd, rawFalse) def rolling_beta(self, market_returns: pd.Series) - pd.Series: Rolling beta vs market. def calc_beta(window_data): port_ret window_data.iloc[:, 0] mkt_ret window_data.iloc[:, 1] cov np.cov(port_ret, mkt_ret) return cov[0, 1] / cov[1, 1] if cov[1, 1] ! 0 else 0 combined pd.concat([self.returns, market_returns], axis1) return combined.rolling(self.window).apply( lambda x: calc_beta(x.to_frame()), rawFalse ).iloc[:, 0] def volatility_regime( self, low_threshold: float 0.10, high_threshold: float 0.20 ) - pd.Series: Classify volatility regime. vol self.rolling_volatility() def classify(v): if v low_threshold: return low elif v high_threshold: return high else: return normal return vol.apply(classify)实现要点默认窗口 63 天约 3 个月window可自行调整SKILL.md 的 Donts 明确提示Dont use short lookbacks - Miss regime changes过短的窗口会错过波动率状态切换。rolling_volatility用 pandas 原生rolling(window).std()实现比循环快几个数量级年化因子在此硬编码为 252。rolling_sharpe在窗口内用算术年化mean * 252计算滚动年化收益。rolling_var/rolling_max_drawdown通过rolling().apply()传入自定义函数rawTrue传 numpy 数组更快rawFalse传 pandas Series可访问索引等元数据。rolling_beta将组合与市场收益 concat 成两列后滚动回归式地计算协方差比。volatility_regime把年化滚动波动率分类为low10%、high20%、normal介于两者之间可直接用于驱动风控规则的切换——例如高波动 regime 下自动收紧仓位与止损对应 risk-manager 的Implementing risk limits职责。六、Pattern 4压力测试StressTester 类压力测试回答极端情景下组合会怎样分三类历史危机重演、假设冲击、Monte Carlo 模拟。class StressTester: Historical and hypothetical stress testing. # Historical crisis periods HISTORICAL_SCENARIOS { 2008_financial_crisis: (2008-09-01, 2009-03-31), 2020_covid_crash: (2020-02-19, 2020-03-23), 2022_rate_hikes: (2022-01-01, 2022-10-31), dot_com_bust: (2000-03-01, 2002-10-01), flash_crash_2010: (2010-05-06, 2010-05-06), } def __init__(self, returns: pd.Series, weights: pd.Series None): self.returns returns self.weights weights def historical_stress_test( self, scenario_name: str, historical_data: pd.DataFrame ) - Dict[str, float]: Test portfolio against historical crisis period. if scenario_name not in self.HISTORICAL_SCENARIOS: raise ValueError(fUnknown scenario: {scenario_name}) start, end self.HISTORICAL_SCENARIOS[scenario_name] # Get returns during crisis crisis_returns historical_data.loc[start:end] if self.weights is not None: port_returns (crisis_returns self.weights) else: port_returns crisis_returns total_return (1 port_returns).prod() - 1 max_dd self._calculate_max_dd(port_returns) worst_day port_returns.min() return { scenario: scenario_name, period: f{start} to {end}, total_return: total_return, max_drawdown: max_dd, worst_day: worst_day, volatility: port_returns.std() * np.sqrt(252) } def hypothetical_stress_test( self, shocks: Dict[str, float] ) - float: Test portfolio against hypothetical shocks. Args: shocks: Dict of {asset: shock_return} if self.weights is None: raise ValueError(Weights required for hypothetical stress test) total_impact 0 for asset, shock in shocks.items(): if asset in self.weights.index: total_impact self.weights[asset] * shock return total_impact def monte_carlo_stress( self, n_simulations: int 10000, horizon_days: int 21, vol_multiplier: float 2.0 ) - Dict[str, float]: Monte Carlo stress test with elevated volatility. mean self.returns.mean() vol self.returns.std() * vol_multiplier simulations np.random.normal( mean, vol, (n_simulations, horizon_days) ) total_returns (1 simulations).prod(axis1) - 1 return { expected_loss: -total_returns.mean(), var_95: -np.percentile(total_returns, 5), var_99: -np.percentile(total_returns, 1), worst_case: -total_returns.min(), prob_10pct_loss: (total_returns -0.10).mean() } def _calculate_max_dd(self, returns: pd.Series) - float: cumulative (1 returns).cumprod() running_max cumulative.cummax() drawdowns (cumulative - running_max) / running_max return drawdowns.min()实现要点历史情景内置 5 个经典危机区间——2008 金融危机、2020 疫情崩盘、2022 加息周期、2000 互联网泡沫、2010 闪崩。historical_stress_test用loc[start:end]截取危机区间收益输出区间总收益、最大回撤、最差单日、波动率。未知情景名会抛ValueError。假设冲击hypothetical_stress_test接收{资产: 冲击收益}字典用权重加权求和计算组合冲击影响未传权重时抛ValueError。适合做某资产跌 20% 会怎样的单点分析。Monte Carlo 压力模拟默认 10000 次模拟、21 天约一个月持有期、波动率放大 2 倍从Normal(mean, vol*2)抽样生成(10000, 21)矩阵逐行复利累乘得到期末总收益分布输出期望损失、VaR95/99、最坏情形与亏损超过 10% 的概率。注意该模拟假设正态分布在极端尾部可能低估风险——建议与历史 VaR / CVaR 交叉验证。七、快速参考把整套工具用起来details.md 提供了开箱即用的日常使用示例# Daily usage metrics RiskMetrics(returns) print(fSharpe: {metrics.sharpe_ratio():.2f}) print(fMax DD: {metrics.max_drawdown():.2%}) print(fVaR 95%: {metrics.var_historical(0.95):.2%}) # Full summary summary metrics.summary() for metric, value in summary.items(): print(f{metric}: {value:.4f})接入风险的完整链路可以这样组织结合同插件 backtesting-frameworks 的向量化回测器用回测器产出组合净值曲线与日收益序列对应 backtesting-frameworks/references/details.md 中的VectorizedBacktester将日收益喂给RiskMetrics输出 Sharpe、Sortino、Calmar、VaR、CVaR、最大回撤等单资产指标多资产场景用PortfolioRisk做组合波动率、成分风险分解与风险平价配置用RollingRiskMetrics监控波动率 regime 变化动态调整风控阈值上线前用StressTester跑历史危机与 Monte Carlo 压力测试完成风险预检。八、最佳实践Dos 与 Donts继承 SKILL.mdSKILL.md 给出了整套方法论的纪律性要求Dos应当遵守Use multiple metrics- 单一指标无法刻画全部风险多指标交叉印证Consider tail risk- 仅有 VaR 不够必须配合 CVaRRolling analysis- 风险随时间变化静态指标会失效Stress test- 历史情景与假设情景都要测Document assumptions- 明确记录分布假设、回看窗口等关键参数Donts必须避免Dont rely on VaR alone- VaR 会低估尾部风险正态假设下尤其明显Dont assume normality- 收益是肥尾的用 Cornish-Fisher 或历史法修正Dont ignore correlation- 压力期相关性会上升用conditional_correlation验证Dont use short lookbacks- 过短窗口会错过波动率 regime 切换Dont forget transaction costs- 交易成本会影响实际实现的风险收益九、小结从代码到风控体系的落地路径本文完整继承了 risk-metrics-calculation 技能的全部四个实现模式与快速参考并逐项注解了参数默认值、年化口径与边界处理。这套代码在 agents24 仓库中的定位是为 risk-manager 与 quant-analyst 两个 Agent 提供可复用的风险计算基础支撑其风险限额实施、R-multiple 跟踪、头寸保护与压力测试的完整职责链。需要记住的五个关键设计决策年化因子默认 252 交易日尾部风险以历史法 VaR CVaR Cornish-Fisher 修正三件套互相校验回撤分析覆盖深度与持续时间两个维度组合层用欧拉分解回答谁贡献了风险压力测试同时覆盖历史危机、假设冲击与 Monte Carlo 三种范式。据此即可搭建从日频监控、动态限额到压力预检的完整风控体系。【免费下载链接】agentsMulti-harness agentic plugin marketplace for Claude Code, Codex, Cursor, OpenCode, GitHub Copilot, and Google Antigravity项目地址: https://gitcode.com/GitHub_Trending/agents24/agents创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考