aria-code 0.167.0__py3-none-win_amd64.whl

aria_code/packages/quant_engine/stochastic/__init__.py

@@ -0,0 +1,31 @@
+"""
+Arthera Stochastic Calculus Module
+随机微积分模块
+
+Components:
+  - ito_calculus.py      : 伊藤引理 / 伊藤积分 / 随机微分方程
+  - gbm_enhanced.py      : 增强几何布朗运动（多资产/跳跃扩散/随机波动率）
+  - stochastic_processes : OU / CIR / Vasicek / Hull-White 过程
+  - monte_carlo_advanced : 方差缩减蒙特卡罗（Antithetic/Control/Quasi-MC）
+  - kelly_criterion      : 凯利公式（连续时间 / 多资产 / Robust版本）
+"""
+
+from .ito_calculus import ItoCalculus, ItoProcess, apply_ito_lemma
+from .gbm_enhanced import EnhancedGBM
+from .stochastic_processes import (
+    OrnsteinUhlenbeck, CIRProcess, VasicekModel, HullWhiteModel
+)
+from .monte_carlo_advanced import MonteCarloEngine, VarianceReduction
+from .kelly_criterion import KellyCriterion
+
+__all__ = [
+    "ItoCalculus", "ItoProcess", "apply_ito_lemma",
+    "EnhancedGBM",
+    "OrnsteinUhlenbeck", "CIRProcess", "VasicekModel", "HullWhiteModel",
+    "MonteCarloEngine", "VarianceReduction",
+    "KellyCriterion",
+]
+
+# Optional explicit backend modules never override this distribution.
+from .._namespace import extend as _extend_namespace
+_extend_namespace(__path__, "stochastic")

aria_code/packages/quant_engine/stochastic/gbm_enhanced.py

@@ -0,0 +1,195 @@
+"""
+增强型几何布朗运动 (Enhanced Geometric Brownian Motion)
+
+包含：
+  1. 多资产相关 GBM (Multi-asset GBM with Cholesky Decomposition)
+  2. 默顿跳跃扩散模型 (Merton Jump-Diffusion Model)
+  3. Heston 随机波动率模型 (Heston Stochastic Volatility)
+  4. 分数布朗运动 (Fractional Brownian Motion) - 选配
+"""
+
+import numpy as np
+import pandas as pd
+from typing import Optional, Tuple, List, Union
+
+
+class EnhancedGBM:
+    """
+    增强型 GBM 仿真器
+    """
+
+    @staticmethod
+    def simulate_multi_asset(
+        s0: np.ndarray,
+        mu: np.ndarray,
+        sigma: np.ndarray,
+        corr_matrix: np.ndarray,
+        T: float = 1.0,
+        n_steps: int = 252,
+        n_paths: int = 1000,
+        seed: Optional[int] = None
+    ) -> np.ndarray:
+        """
+        多资产相关 GBM 模拟
+
+        使用 Cholesky 分解生成相关随机增量：
+        dS_i = μ_i S_i dt + σ_i S_i dW_i
+        E[dW_i dW_j] = ρ_{ij} dt
+
+        Args:
+            s0          : 初始价格向量, shape (n_assets,)
+            mu          : 年化漂移率向量
+            sigma       : 年化波动率向量
+            corr_matrix : 相关系数矩阵, shape (n_assets, n_assets)
+            T           : 终止时间（年）
+            n_steps     : 时间步数
+            n_paths     : 路径数
+            seed        : 随机种子
+
+        Returns:
+            paths: shape (n_paths, n_assets, n_steps + 1)
+        """
+        if seed is not None:
+            np.random.seed(seed)
+
+        n_assets = len(s0)
+        dt = T / n_steps
+
+        # Cholesky 分解: L @ L.T = Correlation
+        L = np.linalg.cholesky(corr_matrix)
+
+        paths = np.zeros((n_paths, n_assets, n_steps + 1))
+        paths[:, :, 0] = s0
+
+        for i in range(n_steps):
+            # 生成独立标准正态分布 Z ~ N(0, 1)
+            Z = np.random.standard_normal((n_paths, n_assets))
+
+            # 转化为相关增量 ε = Z @ L.T
+            epsilon = Z @ L.T
+
+            # 伊藤漂移修正
+            drift = (mu - 0.5 * sigma**2) * dt
+            diffusion = sigma * np.sqrt(dt) * epsilon
+
+            # 更新价格: S_{t+dt} = S_t * exp(drift + diffusion)
+            paths[:, :, i + 1] = paths[:, :, i] * np.exp(drift + diffusion)
+
+        return paths
+
+    @staticmethod
+    def simulate_merton_jump_diffusion(
+        s0: float,
+        mu: float,
+        sigma: float,
+        lambda_j: float,
+        mu_j: float,
+        sigma_j: float,
+        T: float = 1.0,
+        n_steps: int = 252,
+        n_paths: int = 1000,
+        seed: Optional[int] = None
+    ) -> np.ndarray:
+        """
+        默顿跳跃扩散模型 (Merton Jump-Diffusion)
+        
+        dS/S = (μ - λk)dt + σdW + (J - 1)dN
+        其中 N 是强度为 λ 的泊松过程，ln(J) ~ N(μ_j, σ_j²)
+        k = E[J - 1] = exp(μ_j + 0.5σ_j²) - 1
+
+        Args:
+            lambda_j: 跳跃强度（每年平均跳跃次数）
+            mu_j    : 跳跃幅度的对数均值
+            sigma_j : 跳跃幅度的对数标准差
+        """
+        if seed is not None:
+            np.random.seed(seed)
+
+        dt = T / n_steps
+        k = np.exp(mu_j + 0.5 * sigma_j**2) - 1
+
+        paths = np.zeros((n_paths, n_steps + 1))
+        paths[:, 0] = s0
+
+        for i in range(n_steps):
+            # 扩散部分 (Diffusion)
+            Z = np.random.standard_normal(n_paths)
+            drift = (mu - 0.5 * sigma**2 - lambda_j * k) * dt
+            diffusion = sigma * np.sqrt(dt) * Z
+
+            # 跳跃部分 (Jump)
+            # 1. 产生泊松分布的跳跃次数 (通常 dt 很小，N 主要是 0 或 1)
+            N = np.random.poisson(lambda_j * dt, n_paths)
+
+            # 2. 对每个跳跃计算幅度
+            jump_factor = np.ones(n_paths)
+            for path_idx in range(n_paths):
+                if N[path_idx] > 0:
+                    # 总跳跃幅度 = Σ ln(J_i)
+                    total_jump_log = np.random.normal(mu_j, sigma_j, N[path_idx]).sum()
+                    jump_factor[path_idx] = np.exp(total_jump_log)
+
+            # 更新价格
+            paths[:, i + 1] = paths[:, i] * np.exp(drift + diffusion) * jump_factor
+
+        return paths
+
+    @staticmethod
+    def simulate_heston(
+        s0: float,
+        v0: float,
+        mu: float,
+        kappa: float,
+        theta: float,
+        sigma_v: float,
+        rho: float,
+        T: float = 1.0,
+        n_steps: int = 252,
+        n_paths: int = 1000,
+        seed: Optional[int] = None
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Heston 随机波动率模型
+        
+        dS = μS dt + √v S dW1
+        dv = κ(θ - v)dt + σ_v √v dW2
+        E[dW1 dW2] = ρ dt
+
+        Args:
+            v0     : 初始方差 (Initial Variance)
+            kappa  : 均值回归速度 (Mean Reversion Speed)
+            theta  : 长期平均方差 (Long-term Variance)
+            sigma_v: 波动率的波动率 (Vol of Vol)
+            rho    : 价格与波动率的相关性
+        """
+        if seed is not None:
+            np.random.seed(seed)
+
+        dt = T / n_steps
+
+        s_paths = np.zeros((n_paths, n_steps + 1))
+        v_paths = np.zeros((n_paths, n_steps + 1))
+        s_paths[:, 0] = s0
+        v_paths[:, 0] = v0
+
+        for i in range(n_steps):
+            # 生成相关随机变量
+            Z1 = np.random.standard_normal(n_paths)
+            Z2 = np.random.standard_normal(n_paths)
+            W1 = Z1
+            W2 = rho * Z1 + np.sqrt(1 - rho**2) * Z2
+
+            v_curr = v_paths[:, i]
+            # 保证方差非负 (使用 Full Truncation 或 Reflection)
+            v_plus = np.maximum(v_curr, 0)
+
+            # 更新方差 (Euler-Maruyama)
+            v_paths[:, i + 1] = v_curr + kappa * (theta - v_plus) * dt + \
+                               sigma_v * np.sqrt(v_plus * dt) * W2
+
+            # 更新价格
+            s_paths[:, i + 1] = s_paths[:, i] * np.exp(
+                (mu - 0.5 * v_plus) * dt + np.sqrt(v_plus * dt) * W1
+            )
+
+        return s_paths, v_paths

aria_code/packages/quant_engine/stochastic/ito_calculus.py

@@ -0,0 +1,477 @@
+"""
+伊藤积分与伊藤引理实现
+Itô's Lemma and Itô Stochastic Integral
+
+理论基础：
+  设 X(t) 为伊藤过程：dX = μ(X,t)dt + σ(X,t)dW
+  对光滑函数 f(X,t)，伊藤引理给出：
+  df = (∂f/∂t + μ·∂f/∂X + ½σ²·∂²f/∂X²)dt + σ·∂f/∂X·dW
+
+关键应用：
+  1. GBM → S(t) = S₀·exp((μ - σ²/2)t + σW(t))
+  2. Black-Scholes PDE推导
+  3. Feynman-Kac定理（PDE ↔ 期望）
+  4. 测度变换（Girsanov定理）
+"""
+
+from __future__ import annotations
+
+import numpy as np
+import pandas as pd
+from dataclasses import dataclass, field
+from typing import Callable, Optional, Tuple, List
+import warnings
+
+
+@dataclass
+class ItoProcess:
+    """
+    伊藤过程：dX = μ(X,t)dt + σ(X,t)dW
+
+    参数：
+        mu_func  : 漂移系数函数 μ(x, t) -> float
+        sigma_func: 扩散系数函数 σ(x, t) -> float
+        x0       : 初始值
+        name     : 过程名称
+    """
+    mu_func: Callable[[float, float], float]
+    sigma_func: Callable[[float, float], float]
+    x0: float = 1.0
+    name: str = "ItoProcess"
+
+    def simulate_path(
+        self,
+        T: float = 1.0,
+        n_steps: int = 252,
+        seed: Optional[int] = None,
+        method: str = "euler_maruyama"
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        模拟伊藤过程路径
+
+        Args:
+            T      : 终止时间（年）
+            n_steps: 时间步数
+            seed   : 随机种子
+            method : "euler_maruyama"（一阶）或 "milstein"（带二阶修正）
+
+        Returns:
+            (t_grid, X_path): 时间网格和过程路径
+        """
+        if seed is not None:
+            np.random.seed(seed)
+
+        dt = T / n_steps
+        sqrt_dt = np.sqrt(dt)
+        t_grid = np.linspace(0, T, n_steps + 1)
+        X = np.zeros(n_steps + 1)
+        X[0] = self.x0
+
+        dW = np.random.normal(0, sqrt_dt, n_steps)
+
+        for i in range(n_steps):
+            x_i = X[i]
+            t_i = t_grid[i]
+            mu_i = self.mu_func(x_i, t_i)
+            sigma_i = self.sigma_func(x_i, t_i)
+
+            if method == "euler_maruyama":
+                # Euler-Maruyama：X_{n+1} = X_n + μ·Δt + σ·ΔW
+                X[i + 1] = x_i + mu_i * dt + sigma_i * dW[i]
+
+            elif method == "milstein":
+                # Milstein 方法：加入二阶修正项 ½σ·σ'·((ΔW)² - Δt)
+                # σ' ≈ (σ(x+δ,t) - σ(x-δ,t)) / 2δ（数值微分）
+                delta = x_i * 1e-5 if x_i != 0 else 1e-5
+                sigma_prime = (self.sigma_func(x_i + delta, t_i) -
+                               self.sigma_func(x_i - delta, t_i)) / (2 * delta)
+                milstein_correction = 0.5 * sigma_i * sigma_prime * (dW[i] ** 2 - dt)
+                X[i + 1] = x_i + mu_i * dt + sigma_i * dW[i] + milstein_correction
+
+            else:
+                raise ValueError(f"Unknown method: {method}")
+
+        return t_grid, X
+
+    def simulate_ensemble(
+        self,
+        T: float = 1.0,
+        n_steps: int = 252,
+        n_paths: int = 1000,
+        method: str = "euler_maruyama",
+        antithetic: bool = False,
+    ) -> np.ndarray:
+        """
+        模拟多条路径（支持对偶变量方差缩减）
+
+        Returns:
+            paths: shape (n_paths, n_steps+1)
+        """
+        dt = T / n_steps
+        sqrt_dt = np.sqrt(dt)
+        paths = np.zeros((n_paths, n_steps + 1))
+        paths[:, 0] = self.x0
+
+        if antithetic:
+            half = n_paths // 2
+            dW_base = np.random.normal(0, sqrt_dt, (half, n_steps))
+            dW_all = np.vstack([dW_base, -dW_base])  # 对偶变量
+        else:
+            dW_all = np.random.normal(0, sqrt_dt, (n_paths, n_steps))
+
+        t_grid = np.linspace(0, T, n_steps + 1)
+
+        for i in range(n_steps):
+            x_i = paths[:, i]
+            t_i = t_grid[i]
+            mu_i = np.vectorize(self.mu_func)(x_i, t_i)
+            sigma_i = np.vectorize(self.sigma_func)(x_i, t_i)
+
+            if method == "euler_maruyama":
+                paths[:, i + 1] = x_i + mu_i * dt + sigma_i * dW_all[:, i]
+            elif method == "milstein":
+                delta = np.where(x_i != 0, np.abs(x_i) * 1e-5, 1e-5)
+                sig_p = np.vectorize(self.sigma_func)(x_i + delta, t_i)
+                sig_m = np.vectorize(self.sigma_func)(x_i - delta, t_i)
+                sigma_prime = (sig_p - sig_m) / (2 * delta)
+                corr = 0.5 * sigma_i * sigma_prime * (dW_all[:, i] ** 2 - dt)
+                paths[:, i + 1] = x_i + mu_i * dt + sigma_i * dW_all[:, i] + corr
+
+        return paths
+
+
+class ItoCalculus:
+    """
+    伊藤微积分工具集
+
+    提供：
+      1. 伊藤积分（离散近似）
+      2. 伊藤引理（数值验证）
+      3. 测度变换（Girsanov）
+      4. Feynman-Kac 期望定理
+    """
+
+    @staticmethod
+    def ito_integral(
+        integrand: np.ndarray,
+        brownian_increments: np.ndarray
+    ) -> float:
+        """
+        伊藤积分：∫₀ᵀ f(t)dW(t) 的离散近似
+
+        使用左端点 Riemann-Stieltjes 求和（伊藤型）：
+          I = Σ f(t_i) · ΔW_i
+
+        注意：与 Stratonovich 积分的区别在于使用左端点（非中点）
+
+        Args:
+            integrand          : f(t_i) 在各时间节点的值，shape (n,)
+            brownian_increments: ΔW_i = W(t_{i+1}) - W(t_i)，shape (n,)
+
+        Returns:
+            伊藤积分近似值
+        """
+        if len(integrand) != len(brownian_increments):
+            raise ValueError("integrand and brownian_increments must have same length")
+        return float(np.sum(integrand * brownian_increments))
+
+    @staticmethod
+    def ito_isometry_check(
+        integrand: np.ndarray,
+        dt: float,
+        n_trials: int = 10000
+    ) -> dict:
+        """
+        验证伊藤等距性：E[（∫f dW）²] = E[∫f² dt] = ∫E[f²]dt
+
+        Args:
+            integrand: 确定性被积函数（测试用）
+            dt       : 时间步长
+            n_trials : 蒙特卡罗验证次数
+
+        Returns:
+            {'theoretical': float, 'empirical': float, 'error_pct': float}
+        """
+        n = len(integrand)
+        theoretical = float(np.sum(integrand ** 2) * dt)
+
+        squared_integrals = []
+        for _ in range(n_trials):
+            dW = np.random.normal(0, np.sqrt(dt), n)
+            I = np.sum(integrand * dW)
+            squared_integrals.append(I ** 2)
+
+        empirical = float(np.mean(squared_integrals))
+        error_pct = abs(empirical - theoretical) / max(abs(theoretical), 1e-10) * 100
+
+        return {
+            "theoretical": theoretical,
+            "empirical": empirical,
+            "error_pct": error_pct,
+            "isometry_verified": error_pct < 5.0
+        }
+
+    @staticmethod
+    def stratonovich_to_ito(
+        integrand: np.ndarray,
+        sigma_deriv: np.ndarray,
+        dt: float
+    ) -> np.ndarray:
+        """
+        Stratonovich 积分转化为伊藤积分
+
+        关系：∫f ∘ dW = ∫f dW + ½∫f'·f dt
+        修正项 = ½·σ·(∂σ/∂X)·dt（二次变差修正）
+
+        Args:
+            integrand  : Stratonovich 被积函数 f(t_i)
+            sigma_deriv: ∂σ/∂X 在各时间节点的值
+            dt         : 时间步长
+
+        Returns:
+            等价的伊藤积分被积函数
+        """
+        ito_correction = 0.5 * integrand * sigma_deriv * dt
+        return integrand - ito_correction
+
+
+def apply_ito_lemma(
+    f: Callable[[float, float], float],
+    df_dt: Callable[[float, float], float],
+    df_dx: Callable[[float, float], float],
+    d2f_dx2: Callable[[float, float], float],
+    mu: Callable[[float, float], float],
+    sigma: Callable[[float, float], float],
+    x_path: np.ndarray,
+    t_grid: np.ndarray,
+) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """
+    伊藤引理数值实现
+
+    对伊藤过程 dX = μ(X,t)dt + σ(X,t)dW，
+    函数 f(X,t) 满足：
+      df = [∂f/∂t + μ·∂f/∂X + ½σ²·∂²f/∂X²]dt + σ·∂f/∂X·dW
+           |-----------漂移项（ITO漂移）------------|  |扩散项|
+
+    Args:
+        f       : 目标函数 f(x, t)
+        df_dt   : ∂f/∂t
+        df_dx   : ∂f/∂X（一阶偏导）
+        d2f_dx2 : ∂²f/∂X²（二阶偏导）
+        mu      : 漂移系数 μ(x, t)
+        sigma   : 扩散系数 σ(x, t)
+        x_path  : X 的路径，shape (n_steps+1,)
+        t_grid  : 时间网格，shape (n_steps+1,)
+
+    Returns:
+        (f_path, drift_components, diffusion_components)
+        f_path: f(X(t)) 的路径
+        drift : 漂移项 dt 系数在各时间点的值
+        diffusion: 扩散项 dW 系数在各时间点的值
+
+    示例（GBM 验证）：
+        X = S（股价），f = ln(S)
+        μ_lnS = μ - σ²/2    ← 伊藤修正项 -σ²/2
+        σ_lnS = σ
+        → ln S(T) = ln S(0) + (μ - σ²/2)T + σW(T)  ✓
+    """
+    n = len(x_path)
+    f_path = np.zeros(n)
+    drift_components = np.zeros(n)
+    diffusion_components = np.zeros(n)
+
+    for i in range(n):
+        x_i = x_path[i]
+        t_i = t_grid[i]
+        f_path[i] = f(x_i, t_i)
+
+        mu_i = mu(x_i, t_i)
+        sigma_i = sigma(x_i, t_i)
+
+        # 伊藤漂移：∂f/∂t + μ·∂f/∂X + ½σ²·∂²f/∂X²
+        drift_components[i] = (
+            df_dt(x_i, t_i)
+            + mu_i * df_dx(x_i, t_i)
+            + 0.5 * sigma_i ** 2 * d2f_dx2(x_i, t_i)
+        )
+        # 扩散项：σ·∂f/∂X
+        diffusion_components[i] = sigma_i * df_dx(x_i, t_i)
+
+    return f_path, drift_components, diffusion_components
+
+
+class GirsanovTransform:
+    """
+    Girsanov 测度变换（风险中性测度）
+
+    在风险中性世界中，股价过程从：
+      dS = μS dt + σS dW^P  （真实世界）
+    变换为：
+      dS = rS dt + σS dW^Q  （风险中性世界）
+
+    Girsanov 密度（Radon-Nikodym 导数）：
+      dQ/dP = exp(-θ·W(T) - ½θ²·T)
+      其中 θ = (μ - r) / σ  （市场价格风险）
+    """
+
+    def __init__(self, mu: float, r: float, sigma: float):
+        """
+        Args:
+            mu   : 真实世界漂移率（年化）
+            r    : 无风险利率（年化）
+            sigma: 波动率（年化）
+        """
+        self.mu = mu
+        self.r = r
+        self.sigma = sigma
+        self.theta = (mu - r) / sigma  # 市场价格风险
+
+    def radon_nikodym(self, W_T: float, T: float) -> float:
+        """
+        计算 Radon-Nikodym 导数（测度变换权重）
+
+        dQ/dP|_T = exp(-θ·W(T) - ½θ²·T)
+        """
+        return np.exp(-self.theta * W_T - 0.5 * self.theta ** 2 * T)
+
+    def real_to_risk_neutral_brownian(
+        self,
+        W_path: np.ndarray,
+        t_grid: np.ndarray
+    ) -> np.ndarray:
+        """
+        将真实世界布朗运动 W^P 转化为风险中性布朗运动 W^Q
+
+        Girsanov 定理：W^Q(t) = W^P(t) + θ·t
+        """
+        return W_path + self.theta * t_grid
+
+    def price_option_by_measure_change(
+        self,
+        S0: float,
+        K: float,
+        T: float,
+        n_paths: int = 50000,
+        n_steps: int = 252
+    ) -> dict:
+        """
+        通过 Girsanov 测度变换定价欧式看涨期权
+
+        在 Q 测度下：S(T) = S0·exp((r - σ²/2)T + σW^Q(T))
+        C = e^{-rT}·E^Q[max(S(T)-K, 0)]
+
+        Returns:
+            {'price': float, 'delta': float, 'se': float}
+        """
+        dt = T / n_steps
+        sqrt_dt = np.sqrt(dt)
+
+        # 在 Q 测度下模拟（使用风险中性漂移 r - σ²/2）
+        dW = np.random.normal(0, sqrt_dt, (n_paths, n_steps))
+        log_returns = (self.r - 0.5 * self.sigma ** 2) * dt + self.sigma * dW
+        S_T = S0 * np.exp(np.sum(log_returns, axis=1))
+
+        payoffs = np.maximum(S_T - K, 0)
+        discounted = np.exp(-self.r * T) * payoffs
+
+        price = float(np.mean(discounted))
+        se = float(np.std(discounted) / np.sqrt(n_paths))
+
+        # 数值 Delta：ΔC/ΔS ≈ (C(S+ε) - C(S-ε)) / 2ε
+        eps = S0 * 0.01
+        payoffs_up = np.maximum(S_T * (1 + eps / S0) - K, 0)
+        payoffs_dn = np.maximum(S_T * (1 - eps / S0) - K, 0)
+        delta = float(np.exp(-self.r * T) * np.mean(payoffs_up - payoffs_dn) / (2 * eps))
+
+        return {"price": price, "delta": delta, "standard_error": se}
+
+
+class FeynmanKac:
+    """
+    Feynman-Kac 定理：随机过程期望 ↔ PDE
+
+    对于 SDE: dX = μ(X,t)dt + σ(X,t)dW，
+    边值问题 u_t + μ·u_x + ½σ²·u_xx - r·u = -g(x,t)
+    的解为：u(x,t) = E[∫_t^T e^{-r(s-t)} g(X_s,s)ds + e^{-r(T-t)} Φ(X_T) | X_t=x]
+
+    应用：期权定价（Black-Scholes PDE ↔ 风险中性期望）
+    """
+
+    def __init__(
+        self,
+        mu: Callable,
+        sigma: Callable,
+        r: float = 0.0
+    ):
+        self.mu = mu
+        self.sigma = sigma
+        self.r = r
+
+    def solve_by_mc(
+        self,
+        x0: float,
+        t0: float,
+        T: float,
+        terminal_condition: Callable[[float], float],
+        running_cost: Optional[Callable[[float, float], float]] = None,
+        n_paths: int = 50000,
+        n_steps: int = 252,
+        antithetic: bool = True,
+    ) -> dict:
+        """
+        用蒙特卡罗方法求解 Feynman-Kac 公式
+
+        u(x,t) ≈ (1/N) Σ [e^{-r(T-t)}·Φ(X_T^i) + ∫running_cost dt]
+
+        Args:
+            x0                : 初始状态
+            t0                : 初始时间
+            T                 : 终止时间
+            terminal_condition: Φ(X_T) 终端条件（如期权 payoff）
+            running_cost      : g(X_t, t) 持续成本（默认为0）
+            n_paths           : 路径数
+            n_steps           : 时间步数
+            antithetic        : 是否使用对偶变量
+
+        Returns:
+            {'value': float, 'se': float, 'ci_95': tuple}
+        """
+        process = ItoProcess(
+            mu_func=self.mu,
+            sigma_func=self.sigma,
+            x0=x0
+        )
+        tau = T - t0
+        paths = process.simulate_ensemble(
+            T=tau, n_steps=n_steps, n_paths=n_paths,
+            method="euler_maruyama", antithetic=antithetic
+        )
+
+        # 终端 payoff
+        X_T = paths[:, -1]
+        terminal_values = np.array([terminal_condition(x) for x in X_T])
+        discounted_terminal = np.exp(-self.r * tau) * terminal_values
+
+        # 运行成本（积分）
+        if running_cost is not None:
+            t_grid = np.linspace(t0, T, n_steps + 1)
+            dt = tau / n_steps
+            running_costs = np.zeros(n_paths)
+            for j in range(n_steps):
+                g_vals = np.array([running_cost(paths[k, j], t_grid[j]) for k in range(n_paths)])
+                running_costs += np.exp(-self.r * (t_grid[j] - t0)) * g_vals * dt
+            total_values = discounted_terminal + running_costs
+        else:
+            total_values = discounted_terminal
+
+        value = float(np.mean(total_values))
+        se = float(np.std(total_values) / np.sqrt(n_paths))
+        ci_lower = value - 1.96 * se
+        ci_upper = value + 1.96 * se
+
+        return {
+            "value": value,
+            "standard_error": se,
+            "ci_95": (ci_lower, ci_upper),
+            "n_paths": n_paths,
+        }