"""
==============================================================
基于随机森林的农户农地经营权抵押贷款信用风险识别
Random Forest Model for Farmland Mortgage Loan Credit Risk
==============================================================
作者：董轩羽 (27123125)
课程：智能风控
参考文献：
彭艳玲等（2025）系统工程理论与实践
汪寿阳等（2025）农业经济问题
张珩等（2018）中国农村经济
==============================================================
"""
import os, logging
import numpy as np
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split, cross_val_score, StratifiedKFold
from sklearn.metrics import (roc_auc_score, accuracy_score, recall_score,
precision_score, f1_score, confusion_matrix, roc_curve)
from sklearn.preprocessing import StandardScaler
import warnings
warnings.filterwarnings('ignore')
# ─── 日志配置 ───────────────────────────────
LOG_FILE = '/kaggle/working/dxy_rf_log_20260619.txt'
os.makedirs(os.path.dirname(LOG_FILE), exist_ok=True)
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s %(levelname)s %(message)s',
datefmt='%Y-%m-%d %H:%M:%S',
handlers=[logging.FileHandler(LOG_FILE, 'w', 'utf-8'), logging.StreamHandler()]
)
log = logging.getLogger()
log.info("=" * 68)
log.info(" 智能风控课程作业 | 随机森林农户信用风险识别模型")
log.info(" 作者：董轩羽 学号：27123125")
log.info("=" * 68)
# ═══════════════════════════════════════════════════════════════
# 1. 生成模拟调查数据（宁夏/重庆/四川 三省区农户）
# ═══════════════════════════════════════════════════════════════
log.info("─── 步骤 1：生成模拟调查数据 ──────────────────────────────────")
DATA_FILE = '/kaggle/working/dxy_farmland_credit_data.csv'
SEED = 42
rng = np.random.default_rng(SEED)
N = 840
region = rng.choice([1, 2, 3], N, p=[0.38, 0.32, 0.30]) # 1=宁夏 2=重庆 3=四川
age = rng.normal(47, 9, N).clip(24, 74).astype(int)
edu = rng.choice([6,9,12,15], N, p=[0.18,0.52,0.22,0.08]) + rng.integers(-1,2,N)
edu = edu.clip(4, 16)
family_size = rng.integers(2, 8, N)
land_base = np.where(region==1, 9.5, np.where(region==2, 4.8, 5.9))
land_area = (rng.exponential(land_base) + 1.0).clip(0.5, 55.0)
agri_income = (rng.exponential(3.0, N) * (0.4 + 0.6*land_area/land_area.mean())).clip(0.2, 28)
loan_amount = rng.choice([1,2,3,5,8,10,15,20],N,p=[0.08,0.15,0.22,0.25,0.15,0.09,0.04,0.02]) \
+ rng.uniform(-0.4, 0.4, N)
loan_amount = loan_amount.clip(0.5, 20)
insured = rng.binomial(1, 0.47, N)
dist_bank = rng.exponential(11.5, N).clip(1, 75)
loan_term = rng.choice([6,12,24,36], N, p=[0.12,0.68,0.15,0.05])
other_debt = rng.binomial(1, 0.27, N)
non_agri = rng.exponential(2.8, N).clip(0, 18)
# 违约概率（含非线性项：阈值效应 + 交互项）
def sigmoid(x): return 1 / (1 + np.exp(-x))
Z = (- 0.60*(agri_income/agri_income.std())
+ 0.45*(loan_amount/loan_amount.std())
- 0.38*insured
- 0.32*(land_area/land_area.std())
+ 0.20*(family_size/family_size.std())
- 0.18*(edu/edu.std())
+ 0.15*(dist_bank/dist_bank.std())
+ 0.12*other_debt
- 0.10*(non_agri/non_agri.std())
# 非线性：低收入×高贷款 的交互项
+ 0.25*((agri_income < agri_income.mean()*0.6).astype(float)
* (loan_amount > loan_amount.mean()*1.4).astype(float))
+ rng.normal(0, 0.6, N))
prob = sigmoid(Z)
# 以 top-10% 阈值强制控制违约率为 10%
threshold = np.percentile(prob, 90)
default = (prob >= threshold).astype(int)
df = pd.DataFrame({
'地区': region,
'户主年龄': age,
'受教育程度': edu,
'家庭人口': family_size,
'耕地面积': land_area.round(2),
'农业收入': agri_income.round(2),
'贷款金额': loan_amount.round(2),
'是否参农险': insured,
'到银行距离': dist_bank.round(1),
'贷款期限': loan_term,
'是否有其他负债': other_debt,
'非农收入': non_agri.round(2),
'违约': default
})
df.to_csv(DATA_FILE, index=False, encoding='utf-8-sig')
log.info(f" 样本数：{len(df)} 条 | 特征数：{df.shape[1]-1} 个")
log.info(f" 样本违约率：{df['违约'].mean()*100:.2f}%（违约 {df['违约'].sum()} 条）")
log.info(f" 地区分布：宁夏={( df['地区']==1).sum()} 重庆={(df['地区']==2).sum()} 四川={(df['地区']==3).sum()}")
log.info(f" 数据文件：./数据/dxy_farmland_credit_data.csv")
# ═══════════════════════════════════════════════════════════════
# 2. 描述性统计
# ═══════════════════════════════════════════════════════════════
log.info("─── 步骤 2：描述性统计 ────────────────────────────────────────")
FEAT = ['农业收入','贷款金额','是否参农险','耕地面积','家庭人口',
'受教育程度','到银行距离','贷款期限','是否有其他负债','非农收入','户主年龄']
log.info(f" {'变量':<12} {'均值':>7} {'标准差':>7} {'违约组均值':>10} {'正常组均值':>10}")
log.info(f" {'-'*60}")
for c in FEAT:
mu = df[c].mean(); sd = df[c].std()
md = df[df['违约']==1][c].mean()
mn = df[df['违约']==0][c].mean()
log.info(f" {c:<12} {mu:>7.2f} {sd:>7.2f} {md:>10.2f} {mn:>10.2f}")
# ═══════════════════════════════════════════════════════════════
# 3. 数据划分
# ═══════════════════════════════════════════════════════════════
log.info("─── 步骤 3：训练集 / 测试集划分（75% / 25%，分层抽样）──────────")
X, y = df[FEAT], df['违约']
X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.25, random_state=SEED, stratify=y)
log.info(f" 训练集：{len(X_tr)} 条（违约 {y_tr.sum()}，{y_tr.mean()*100:.1f}%）")
log.info(f" 测试集：{len(X_te)} 条（违约 {y_te.sum()}，{y_te.mean()*100:.1f}%）")
# ═══════════════════════════════════════════════════════════════
# 4. 模型训练
# ═══════════════════════════════════════════════════════════════
log.info("─── 步骤 4：模型训练 ──────────────────────────────────────────")
# ── 随机森林 ──
log.info(" [随机森林] 开始训练 (n_estimators=200, max_depth=10, class_weight='balanced')")
rf = RandomForestClassifier(
n_estimators=200, max_depth=10, min_samples_leaf=4,
max_features='sqrt', class_weight='balanced', random_state=SEED, n_jobs=-1
)
rf.fit(X_tr, y_tr)
cv_rf = cross_val_score(rf, X_tr, y_tr, cv=StratifiedKFold(5, shuffle=True, random_state=SEED),
scoring='roc_auc')
log.info(f" [随机森林] 训练完成 | 5折CV AUC：{cv_rf.mean():.4f} ± {cv_rf.std():.4f}")
# ── Logistic 回归 ──
log.info(" [Logistic回归] 开始训练 (C=1.0, class_weight='balanced')")
scaler = StandardScaler()
X_tr_s = scaler.fit_transform(X_tr); X_te_s = scaler.transform(X_te)
lr = LogisticRegression(C=1.0, class_weight='balanced', max_iter=500, random_state=SEED)
lr.fit(X_tr_s, y_tr)
cv_lr = cross_val_score(lr, X_tr_s, y_tr, cv=StratifiedKFold(5, shuffle=True, random_state=SEED),
scoring='roc_auc')
log.info(f" [Logistic回归] 训练完成 | 5折CV AUC：{cv_lr.mean():.4f} ± {cv_lr.std():.4f}")
# ═══════════════════════════════════════════════════════════════
# 5. 模型评估
# ═══════════════════════════════════════════════════════════════
log.info("─── 步骤 5：测试集模型性能评估 ────────────────────────────────")
def eval_model(model, Xtest, ytest, name, scaler=None):
Xe = scaler.transform(Xtest) if scaler else Xtest
yp = model.predict_proba(Xe)[:,1]
yc = model.predict(Xe)
auc = roc_auc_score(ytest, yp)
acc = accuracy_score(ytest, yc)
rec = recall_score(ytest, yc)
pre = precision_score(ytest, yc)
f1 = f1_score(ytest, yc)
fpr,tpr,_ = roc_curve(ytest, yp)
ks = (tpr - fpr).max()
cm = confusion_matrix(ytest, yc)
log.info(f"\n ── {name} ──────────────────────────")
log.info(f" AUC = {auc:.4f}")
log.info(f" 准确率 = {acc:.4f} ({acc*100:.1f}%)")
log.info(f" 违约召回率 = {rec:.4f} ({rec*100:.1f}%)")
log.info(f" 精确率 = {pre:.4f}")
log.info(f" F1 分数 = {f1:.4f}")
log.info(f" KS 统计量 = {ks:.4f}")
log.info(f" 混淆矩阵：TN={cm[0,0]} FP={cm[0,1]} FN={cm[1,0]} TP={cm[1,1]}")
return dict(name=name, AUC=auc, 准确率=acc, 违约召回率=rec, 精确率=pre, F1=f1, KS=ks, prob=yp)
rf_r = eval_model(rf, X_te, y_te, "随机森林 (Random Forest)")
lr_r = eval_model(lr, X_te, y_te, "Logistic 回归", scaler=scaler)
# ═══════════════════════════════════════════════════════════════
# 6. 特征重要性
# ═══════════════════════════════════════════════════════════════
log.info("\n─── 步骤 6：特征重要性分析 (Gini 不纯度降低量) ───────────────")
imp = pd.Series(rf.feature_importances_, index=FEAT).sort_values(ascending=False)
log.info(f" {'排名':>4} {'特征':<14} {'重要性':>8}")
log.info(f" {'-'*34}")
for r,(f,v) in enumerate(imp.items(),1):
log.info(f" {r:>4} {f:<14} {v:>8.4f}")
# ═══════════════════════════════════════════════════════════════
# 7. Credit Risk+ 期望损失测量
# ═══════════════════════════════════════════════════════════════
log.info("\n─── 步骤 7：Credit Risk+ 期望损失测量 ──────────────────────────")
pd_arr = rf.predict_proba(X_te)[:,1]
LGD = 0.65
ead_arr = X_te['贷款金额'].values
el_arr = pd_arr * LGD * ead_arr
total_ead = ead_arr.sum()
total_el = el_arr.sum()
el_rate = total_el / total_ead * 100
shock_el = pd_arr * (LGD * 1.35) * ead_arr
shock_inc = (shock_el.sum() - total_el) / total_el * 100
log.info(f" 总风险敞口（EAD） : {total_ead:.2f} 万元")
log.info(f" 违约损失率（LGD）假设 : {LGD:.0%}（农地流转变现难，参照行业均值）")
log.info(f" 组合期望损失（EL） : {total_el:.2f} 万元")
log.info(f" 期望损失率（EL/EAD） : {el_rate:.2f}%")
log.info(f" [压力测试] 极端气候冲击（LGD×1.35={LGD*1.35:.0%}）：期望损失增幅 +{shock_inc:.1f}%")
# ═══════════════════════════════════════════════════════════════
# 8. 汇总
# ═══════════════════════════════════════════════════════════════
log.info("\n─── 步骤 8：结果汇总 ──────────────────────────────────────────")
log.info("\n 表2 模型性能对比（测试集）")
log.info(f" {'指标':<12} {'随机森林':>10} {'Logistic':>10} {'差值':>8} {'提升幅度':>9}")
log.info(f" {'-'*56}")
for k in ['AUC','准确率','违约召回率','F1','KS']:
rv, lv = rf_r[k], lr_r[k]
log.info(f" {k:<12} {rv:>10.4f} {lv:>10.4f} {rv-lv:>+8.4f} {(rv-lv)/lv*100:>+8.1f}%")
log.info("\n 表3 特征重要性前7位")
log.info(f" {'排名':>4} {'特征变量':<14} {'重要性得分':>10} {'经济含义'}")
log.info(f" {'-'*60}")
meanings = {'农业收入':'还款能力核心指标','贷款金额':'资金需求与偿债压力',
'耕地面积':'抵押物价值基础','是否参农险':'风险保障程度',
'家庭人口':'家庭负担水平','受教育程度':'风险认知与管理能力',
'到银行距离':'金融服务可及性','非农收入':'收入多元化程度',
'户主年龄':'经验与风险偏好','贷款期限':'还款期限结构','是否有其他负债':'债务负担'}
for r,(f,v) in enumerate(imp.head(7).items(),1):
log.info(f" {r:>4} {f:<14} {v:>10.4f} {meanings.get(f,'—')}")
log.info("\n Credit Risk+ 损失测量汇总")
log.info(f" 基准情景 EL = {total_el:.2f} 万元 (EL率 {el_rate:.2f}%)")
log.info(f" 极端情景 EL = {shock_el.sum():.2f} 万元 (增幅 +{shock_inc:.1f}%)")
log.info("\n" + "=" * 68)
log.info(" 模型运行完成")
log.info("=" * 68)
print(f"\n✓ 全部步骤运行完成")
print(f" 运行日志：./运行日志/dxy_rf_log_20260619.txt")
print(f" 模拟数据：./数据/dxy_farmland_credit_data.csv")