4. PSI分箱策略：等距分箱、等频分箱、最优分箱在PSI计算中的选择

PSI（群体稳定性指标）的计算高度依赖分箱策略。不同的分箱方式会直接影响PSI数值的敏感度与稳定性，错误的分箱可能导致“假告警”或“漏告警”。本节将系统对比三种主流分箱策略的数学原理、适用场景及在贷后预警中的选择原则。

4.1 分箱策略的核心矛盾

PSI的本质是衡量两个分布（预期分布 vs 实际分布）在各分箱上的偏差。分箱策略需要平衡以下矛盾：

• 粒度 vs 稳定性：箱数越多，对局部偏移越敏感，但噪声也越大；箱数越少，稳定性越好，但可能掩盖关键偏移。
• 边界 vs 业务含义：等距分箱边界固定但可能破坏数据自然分布；等频分箱保证箱内样本量均衡但边界随分布变化；最优分箱追求统计最优但可能牺牲可解释性。

4.2 三种分箱策略详解

4.3 分箱策略选择决策流程

在贷后预警系统中，分箱策略的选择应遵循以下决策流程：

1. 变量类型判断：连续变量（如额度使用率）vs 离散变量（如信用等级）。
2. 分布形态评估：使用KS检验或直方图判断变量分布是否均匀、是否偏态。
3. 业务可解释性要求：是否需要向业务人员解释分箱边界（如“额度使用率>80%”比“分箱3”更直观）。
4. 跨期稳定性需求：如果需要在不同时间点保持箱边界一致，优先选择等距分箱。
5. 回测验证：使用历史数据模拟PSI告警，选择误报率最低的策略。

4.4 分箱策略对PSI数值的影响对比

以下表格展示了同一变量在不同分箱策略下的PSI计算结果差异（基于某贷后数据集）：

4.5 分箱策略选择流程图

4.6 代码实现：三种分箱策略的PSI计算

以下Python代码展示了如何使用三种分箱策略计算PSI，并对比结果：

import numpy as np
import pandas as pd
from scipy.stats import chi2_contingency

def psi_equidistant(actual, expected, bins=10):
    """等距分箱PSI计算"""
    min_val = min(min(actual), min(expected))
    max_val = max(max(actual), max(expected))
    bin_edges = np.linspace(min_val, max_val, bins + 1)
    
    actual_counts, _ = np.histogram(actual, bins=bin_edges)
    expected_counts, _ = np.histogram(expected, bins=bin_edges)
    
    actual_pct = actual_counts / len(actual)
    expected_pct = expected_counts / len(expected)
    
    # 处理零值
    actual_pct = np.where(actual_pct == 0, 0.0001, actual_pct)
    expected_pct = np.where(expected_pct == 0, 0.0001, expected_pct)
    
    psi = np.sum((actual_pct - expected_pct) * np.log(actual_pct / expected_pct))
    return psi

def psi_equifrequent(actual, expected, bins=10):
    """等频分箱PSI计算（基于实际分布分箱）"""
    # 使用实际数据的分位数作为分箱边界
    percentiles = np.linspace(0, 100, bins + 1)[1:-1]
    bin_edges = np.percentile(actual, percentiles)
    bin_edges = np.concatenate([[min(actual)], bin_edges, [max(actual)]])
    
    actual_counts, _ = np.histogram(actual, bins=bin_edges)
    expected_counts, _ = np.histogram(expected, bins=bin_edges)
    
    actual_pct = actual_counts / len(actual)
    expected_pct = expected_counts / len(expected)
    
    actual_pct = np.where(actual_pct == 0, 0.0001, actual_pct)
    expected_pct = np.where(expected_pct == 0, 0.0001, expected_pct)
    
    psi = np.sum((actual_pct - expected_pct) * np.log(actual_pct / expected_pct))
    return psi

def psi_optimal(actual, expected, max_bins=10, p_threshold=0.05):
    """最优分箱PSI计算（基于卡方合并）"""
    # 初始分箱：使用等频分箱作为起点
    percentiles = np.linspace(0, 100, max_bins + 1)[1:-1]
    bin_edges = np.percentile(actual, percentiles)
    bin_edges = np.concatenate([[min(actual)], bin_edges, [max(actual)]])
    
    # 卡方合并：合并p值大于阈值的相邻箱
    merged = True
    while merged and len(bin_edges) > 3:
        merged = False
        min_p = float('inf')
        min_idx = -1
        
        for i in range(len(bin_edges) - 2):
            # 构建2x2列联表
            table = np.array([
                [np.sum((actual >= bin_edges[i]) & (actual < bin_edges[i+1])),
                 np.sum((actual >= bin_edges[i+1]) & (actual < bin_edges[i+2]))],
                [np.sum((expected >= bin_edges[i]) & (expected < bin_edges[i+1])),
                 np.sum((expected >= bin_edges[i+1]) & (expected < bin_edges[i+2]))]
            ])
            
            if np.any(table == 0):
                continue
                
            _, p, _, _ = chi2_contingency(table)
            if p < min_p:
                min_p = p
                min_idx = i
        
        if min_p > p_threshold and min_idx >= 0:
            # 合并相邻箱
            bin_edges = np.delete(bin_edges, min_idx + 1)
            merged = True
    
    # 使用合并后的边界计算PSI
    actual_counts, _ = np.histogram(actual, bins=bin_edges)
    expected_counts, _ = np.histogram(expected, bins=bin_edges)
    
    actual_pct = actual_counts / len(actual)
    expected_pct = expected_counts / len(expected)
    
    actual_pct = np.where(actual_pct == 0, 0.0001, actual_pct)
    expected_pct = np.where(expected_pct == 0, 0.0001, expected_pct)
    
    psi = np.sum((actual_pct - expected_pct) * np.log(actual_pct / expected_pct))
    return psi

# 示例数据
np.random.seed(42)
actual = np.random.normal(50, 15, 10000)
expected = np.random.normal(55, 18, 10000)

# 计算三种PSI
psi_eqd = psi_equidistant(actual, expected, bins=10)
psi_eqf = psi_equifrequent(actual, expected, bins=10)
psi_opt = psi_optimal(actual, expected, max_bins=10)

print(f"等距分箱 PSI: {psi_eqd:.4f}")
print(f"等频分箱 PSI: {psi_eqf:.4f}")
print(f"最优分箱 PSI: {psi_opt:.4f}")

4.7 分箱策略选择总结