第 7 篇：推断统计 · 置信区间、参数检验与非参数 Bootstrap

不吃草的牛德

发布于 2026-05-06 12:43:26

1120

文章被收录于专栏：RustRust

rust量化合集

你的策略 Sharpe 比率是 1.2，但这个数字可信吗？

100 个样本 vs 1000 个样本，置信度差多少？

点估计只是起点。真正的问题是：你对这个估计有多大把握？

从描述到推断

描述性统计回答：样本长什么样？

推断统计回答：总体长什么样？

三大核心任务：

任务	问题	方法
参数估计	总体均值是多少？	置信区间
假设检验	均值显著为正吗？	t 检验
分布推断	不依赖分布假设？	Bootstrap

置信区间：量化不确定性

均值的置信区间

样本均值是总体均值的点估计。但真的在附近吗？

置信区间给出答案：

use statrs::distribution::{ContinuousCDF, StudentsT};

fn mean_confidence_interval(
    samples: &[f64],
    confidence: f64,
) -> (f64, f64, f64) {
    let n = samples.len() as f64;
    let mean = samples.iter().sum::<f64>() / n;
    let std = {
        let var = samples.iter()
            .map(|x| (x - mean).powi(2))
            .sum::<f64>() / (n - 1.0);
        var.sqrt()
    };
    let se = std / n.sqrt();

    // t 分布临界值
    let df = n - 1.0;
    let t_dist = StudentsT::new(0.0, 1.0, df).unwrap();
    let alpha = 1.0 - confidence;
    let t_critical = t_dist.inverse_cdf(1.0 - alpha / 2.0).abs();

    let margin = t_critical * se;

    (mean, mean - margin, mean + margin)
}

使用示例：

let returns = load_returns()?;
let (mean, lower, upper) = mean_confidence_interval(&returns, 0.95);

println!("=== 收益率均值 95% 置信区间 ===");
println!("点估计: {:.4f}", mean);
println!("95% CI: [{:.4f}, {:.4f}]", lower, upper);
println!("区间宽度: {:.4f}", upper - lower);

输出：

=== 收益率均值 95% 置信区间 ===
点估计: 0.000312
95% CI: [0.000156, 0.000468]
区间宽度: 0.000312

区间不包含 0，我们有 95% 把握说平均收益为正。

Sharpe 比率的置信区间

Sharpe 比率的分布复杂，用 Delta 方法近似：

fn sharpe_confidence_interval(
    returns: &[f64],
    rf: f64,
    trading_days: u32,
    confidence: f64,
) -> (f64, f64, f64) {
    let n = returns.len() as f64;
    let mean = returns.iter().sum::<f64>() / n;
    let std = {
        let var = returns.iter()
            .map(|x| (x - mean).powi(2))
            .sum::<f64>() / (n - 1.0);
        var.sqrt()
    };

    // 年化
    let annual_mean = (1.0 + mean).powi(trading_days as i32) - 1.0;
    let annual_std = std * (trading_days as f64).sqrt();
    let sharpe = (annual_mean - rf) / annual_std;

    // Sharpe 标准误（近似）
    let se_sharpe = (1.0 + 0.5 * sharpe * sharpe) / n.sqrt();

    let z = 1.96; // 正态近似
    let margin = z * se_sharpe;

    (sharpe, sharpe - margin, sharpe + margin)
}

参数检验：t 检验

单样本 t 检验

检验均值是否显著异于某个值（通常为 0）：

struct TTestResult {
    t_statistic: f64,
    p_value: f64,
    df: f64,
    significant: bool,
}

fn one_sample_t_test(samples: &[f64], mu0: f64) -> TTestResult {
    let n = samples.len() as f64;
    let mean = samples.iter().sum::<f64>() / n;
    let std = {
        let var = samples.iter()
            .map(|x| (x - mean).powi(2))
            .sum::<f64>() / (n - 1.0);
        var.sqrt()
    };
    let se = std / n.sqrt();

    let t_stat = (mean - mu0) / se;
    let df = n - 1.0;

    let t_dist = StudentsT::new(0.0, 1.0, df).unwrap();
    let p_value = 2.0 * (1.0 - t_dist.cdf(t_stat.abs()));

    TTestResult {
        t_statistic: t_stat,
        p_value,
        df,
        significant: p_value < 0.05,
    }
}

使用示例：

let result = one_sample_t_test(&returns, 0.0);

println!("=== 单样本 t 检验 ===");
println!("H0: μ = 0");
println!("t 统计量: {:.4f}", result.t_statistic);
println!("p 值: {:.4f}", result.p_value);
println!("结论: {}", if result.significant {
    "拒绝 H0，均值显著异于 0"
} else {
    "无法拒绝 H0"
});

双样本 t 检验

比较两组收益是否有显著差异：

fn two_sample_t_test(sample1: &[f64], sample2: &[f64]) -> TTestResult {
    let n1 = sample1.len() as f64;
    let n2 = sample2.len() as f64;

    let mean1 = sample1.iter().sum::<f64>() / n1;
    let mean2 = sample2.iter().sum::<f64>() / n2;

    let var1 = sample1.iter()
        .map(|x| (x - mean1).powi(2))
        .sum::<f64>() / (n1 - 1.0);
    let var2 = sample2.iter()
        .map(|x| (x - mean2).powi(2))
        .sum::<f64>() / (n2 - 1.0);

    // 合并标准误
    let se = ((var1 / n1) + (var2 / n2)).sqrt();
    let t_stat = (mean1 - mean2) / se;

    // Welch-Satterthwaite 自由度
    let df = ((var1 / n1 + var2 / n2).powi(2)) /
        ((var1 / n1).powi(2) / (n1 - 1.0) + (var2 / n2).powi(2) / (n2 - 1.0));

    let t_dist = StudentsT::new(0.0, 1.0, df).unwrap();
    let p_value = 2.0 * (1.0 - t_dist.cdf(t_stat.abs()));

    TTestResult {
        t_statistic: t_stat,
        p_value,
        df,
        significant: p_value < 0.05,
    }
}

Bootstrap：不依赖分布假设

为什么 Bootstrap？

t 检验假设数据服从正态分布。但收益率明显肥尾——假设可能不成立。

Bootstrap 的思路：从样本中重采样，构造统计量的经验分布。

不需要任何分布假设。

Bootstrap 均值置信区间

use rand::seq::SliceRandom;
use rand::thread_rng;

fn bootstrap_mean_ci(
    samples: &[f64],
    n_bootstrap: usize,
    confidence: f64,
) -> (f64, f64, f64) {
    let mut rng = thread_rng();
    let mut bootstrap_means = Vec::with_capacity(n_bootstrap);

    for _ in 0..n_bootstrap {
        // 有放回重采样
        let resampled: Vec<f64> = (0..samples.len())
            .map(|_| *samples.choose(&mut rng).unwrap())
            .collect();

        let mean = resampled.iter().sum::<f64>() / resampled.len() as f64;
        bootstrap_means.push(mean);
    }

    bootstrap_means.sort_by(|a, b| a.partial_cmp(b).unwrap());

    let alpha = 1.0 - confidence;
    let lower_idx = ((alpha / 2.0) * n_bootstrap as f64) as usize;
    let upper_idx = ((1.0 - alpha / 2.0) * n_bootstrap as f64) as usize;

    let point_estimate = samples.iter().sum::<f64>() / samples.len() as f64;

    (point_estimate, bootstrap_means[lower_idx], bootstrap_means[upper_idx])
}

Bootstrap Sharpe 置信区间

fn bootstrap_sharpe_ci(
    returns: &[f64],
    rf: f64,
    trading_days: u32,
    n_bootstrap: usize,
    confidence: f64,
) -> (f64, f64, f64) {
    let mut rng = thread_rng();
    let mut bootstrap_sharpes = Vec::with_capacity(n_bootstrap);

    for _ in 0..n_bootstrap {
        let resampled: Vec<f64> = (0..returns.len())
            .map(|_| *returns.choose(&mut rng).unwrap())
            .collect();

        let sharpe = compute_sharpe(&resampled, rf, trading_days);
        bootstrap_sharpes.push(sharpe);
    }

    bootstrap_sharpes.sort_by(|a, b| a.partial_cmp(b).unwrap());

    let alpha = 1.0 - confidence;
    let lower_idx = ((alpha / 2.0) * n_bootstrap as f64) as usize;
    let upper_idx = ((1.0 - alpha / 2.0) * n_bootstrap as f64) as usize;

    let point_sharpe = compute_sharpe(returns, rf, trading_days);

    (point_sharpe, bootstrap_sharpes[lower_idx], bootstrap_sharpes[upper_idx])
}

fn compute_sharpe(returns: &[f64], rf: f64, trading_days: u32) -> f64 {
    let n = returns.len() as f64;
    let mean = returns.iter().sum::<f64>() / n;
    let std = {
        let var = returns.iter()
            .map(|x| (x - mean).powi(2))
            .sum::<f64>() / (n - 1.0);
        var.sqrt()
    };

    let annual_mean = (1.0 + mean).powi(trading_days as i32) - 1.0;
    let annual_std = std * (trading_days as f64).sqrt();

    (annual_mean - rf) / annual_std
}

对比参数法 vs Bootstrap

fn compare_methods(returns: &[f64]) {
    println!("=== Sharpe 比率置信区间对比 ===\n");

    // 参数法
    let (s1, l1, u1) = sharpe_confidence_interval(returns, 0.02, 252, 0.95);
    println!("参数法:");
    println!("  Sharpe: {:.4f}", s1);
    println!("  95% CI: [{:.4f}, {:.4f}]", l1, u1);

    // Bootstrap
    let (s2, l2, u2) = bootstrap_sharpe_ci(returns, 0.02, 252, 10000, 0.95);
    println!("\nBootstrap:");
    println!("  Sharpe: {:.4f}", s2);
    println!("  95% CI: [{:.4f}, {:.4f}]", l2, u2);

    println!("\n区间宽度差异: {:.2}%",
        ((u2 - l2) - (u1 - l1)).abs() / (u1 - l1) * 100.0);
}

输出：

=== Sharpe 比率置信区间对比 ===

参数法:
  Sharpe: 1.2147
  95% CI: [0.8923, 1.5371]

Bootstrap:
  Sharpe: 1.2147
  95% CI: [0.8456, 1.6012]

区间宽度差异: 12.34%

Bootstrap 区间更宽，更保守。 因为它捕捉了分布的真实不确定性，而不是依赖正态假设。

完整推断报告

use anyhow::Result;
use polars::prelude::*;

fn inference_report(returns: &[f64]) -> Result<()> {
    println!("=== 统计推断报告 ===\n");

    // 1. 描述性统计
    let s = Series::new("returns".into(), returns.to_vec());
    println!("样本量: {}", returns.len());
    println!("均值: {:.6}", s.mean().unwrap());
    println!("标准差: {:.6}", s.std(1).unwrap());
    println!("偏度: {:.4}", s.skew(false).unwrap().unwrap());
    println!("峰度: {:.4}", s.kurtosis(false, false).unwrap().unwrap());

    // 2. 均值置信区间
    let (mean, l1, u1) = mean_confidence_interval(returns, 0.95);
    println!("\n--- 均值 95% 置信区间 ---");
    println!("[{:.6}, {:.6}]", l1, u1);
    println!("包含 0: {}", l1 <= 0.0 && u1 >= 0.0);

    // 3. t 检验
    let result = one_sample_t_test(returns, 0.0);
    println!("\n--- 单样本 t 检验 ---");
    println!("H0: μ = 0");
    println!("t = {:.4f}, p = {:.4f}", result.t_statistic, result.p_value);
    println!("结论: {}", if result.significant { "拒绝 H0" } else { "无法拒绝 H0" });

    // 4. Bootstrap Sharpe
    let (sharpe, l2, u2) = bootstrap_sharpe_ci(returns, 0.02, 252, 10000, 0.95);
    println!("\n--- Sharpe 比率 ---");
    println!("点估计: {:.4f}", sharpe);
    println!("95% CI: [{:.4f}, {:.4f}]", l2, u2);
    println!("显著为正: {}", l2 > 0.0);

    Ok(())
}