风控模型开发验证报告

实验 ID: DEV-mw0b63sf-2afb · 生成时间: 2026-11-03 14:31 (GMT+8) · 报告结构:标准14节模型开发验证模板

00 模型概要信息

项目内容
模型名称exp_DEV-mw0b63sf-2afb
开发完成日期2026-11-03
模型开发人员(人工填写)
模型开发人员邮件地址(人工填写)
模型验证人员(人工填写)
模型验证人员邮件地址(人工填写)
模型应用环节collection_scorecard
模型应用频率(人工填写)
模型开发背景及业务用途描述Develop run with tuning
模型适用产品/客户描述(人工填写)
模型重开发原因(如有)—
模型拟监控频率(人工填写)
模型路线LR + WOE 评分卡
目标变量gbflag
实验 IDDEV-mw0b63sf-2afb
待补充:模型开发/验证人员、应用频率、产品描述等字段需人工补填。

01 模型设计

建模数据样本窗口

项目内容
切分方式外部 data_flag 划分(TRAIN 内 6/4 随机开发/测试,VALI 跨时间,PSI 时点)
开发样本(Train)窗口2021-10-17 ~ 2026-04-30
预留验证样本窗口2024-04-30 ~ 2026-04-30
跨时间验证样本窗口2021-12-31 ~ 2026-06-30
PSI时点样本窗口2023-11-18 ~ 2026-07-31

模型好坏定义

项目内容
目标变量gbflag
坏样本(BAD=1)占比8.03%
好样本(BAD=0)占比—
不确定样本本流程要求目标变量已预计算为 0/1 二值,不确定样本请在上游合并口径

样本排除规则

序号排除规则说明
1id_no主键/时间/口径字段,不入模
2create_time主键/时间/口径字段,不入模
3data_flag主键/时间/口径字段,不入模
4gbflag 为空排除 0 行

滚动率分析

待补充:滚动率分析(M0→M6+ 状态迁移矩阵)需要逾期状态口径数据,当前流程目标变量为上游已计算的二值标签,暂未生成。
待补充:表现窗口及坏账率分析、模型细分定义需结合业务口径人工补充。

02 数据流

注:中文字典已加载(dictionary.csv,匹配 1 个字段);以下字段未在字典中、报告显示英文原名,请补充字典:ps_perd_no、gbflag、data_flag、id_no、age、max_int_rat、min_int_rat、min_tnr、sum_apply_amt、min_apply_amt、min_apprv_amt、apprv_amt_day、end_loan_to_rep_loan_freq、end_loan_to_cur_freq、f_rep_loan_to_f_apply_dt_freq、ps_prcp_amt_mon、ps_int_amt_mon、ps_int_amt_sum、ps_prcp_amt_avg、ps_int_amt_avg、setl_prcp_mon_ratio、acc_ps_perd_no、max_od_days、m_s_ps_instm_amt、f_ps_od_days_max、s_ps_od_days_max、t_ps_od_days_max、ovd_days_d3_num_m1、orig_prcp_avg_m3、orig_prcp_avg_m12 等共 127 个。

开发时点 — 样本排除

类别样本数
全部样本115682
排除:目标变量为空0
排除后样本115682

注:按照01模型设计的排除规则统计;当前流程的排除项为目标变量缺失。

总体样本概况

数据集样本量好样本坏样本坏样本率占比
全部样本11568210638992938.03%100.00%
开发样本380883493731518.27%32.92%
测试样本253912329121008.27%21.95%
训练+测试样本634795822852518.27%54.87%
跨时间验证样本347303229824327.00%30.02%
PSI时点样本17473158631610—15.10%

开发/预留验证样本交叉统计

类别开发+预留样本数开发样本数开发占比预留验证样本数预留验证占比
好582283493760.00%2329140.00%
坏5251315160.01%210039.99%
合计634793808860.00%2539140.00%

细致变量分析(预分箱口径,开发样本)

变量IV缺失率取值集中度最大相关系数
last_od_to_cur_freq0.19280.02%70.20%0.4978
max_od_days0.18700.02%25.81%0.5557
last_loan_setl_tnr0.16690.02%18.96%0.3034
end_loan_to_cur_freq0.14950.02%6.32%0.6425
apprv_amt_day0.14300.02%79.67%0.6396
cre_l60_used_limit_rate_min0.12700.06%10.21%0.7274
ps_perd_no0.12440.00%13.10%0.6425
loan_os_prcp_credit_amount_m30.12220.07%9.66%0.5204
cre_l90_used_limit_amt_min0.11600.06%14.62%0.7390
odgt0_m12_subcurmonth0.10940.07%34.86%0.7415
shd_pre_repay_due_rate0.10881.11%16.57%0.5562
cre_now_l60_used_amt_m_min0.10490.06%74.02%0.6396
cre_now_l60_limit_left_d_max0.10130.48%67.92%0.7660
loan_os_prcp_credit_amount_m60.09440.07%2.98%0.5204
f_ps_od_days_max0.09320.02%53.63%0.5202
suc_apprv_num_m10.09090.02%79.23%0.6800
acc_ps_perd_no0.08750.02%25.81%0.4281
setl_prcp_ratio0.08660.02%21.83%0.6593
last_od_num_m60.08590.02%78.72%0.6476
od_amt_m6_loan_os_prcp_m6_pc0.07980.07%93.14%0.6644
loan_os_prcp_credit_amount_m120.07950.07%0.23%0.2371
ps_int_amt_mon0.07750.02%5.50%0.7503
available_limit_amt0.07710.02%32.48%0.8089
cre_l30_l90_used_amt_d_max0.07700.34%21.39%0.6593
orig_prcp_avg_m30.07640.02%60.04%0.5989
cre_l30_l90_used_amt_m_max0.07310.06%21.48%0.6810
cre_now_l30_used_amt_d_min0.07105.71%86.91%0.9733
shd_max_dpd0.06901.05%74.62%0.3801
cre_l60_used_limit_rate_max0.06850.06%6.42%0.7274
od_amt_m12_loan_os_prcp_m12_pc0.06490.07%92.39%0.6644
repay_amount_m3_loan_os_prcp_pc0.06340.07%38.42%0.6366
t_ps_od_days_max0.06070.02%54.18%0.5455
apprv_num_m10.05950.02%73.93%0.7533
s_ps_od_days_max0.05940.02%54.12%0.5455
shd_l30_max_dpd0.05762.96%71.65%0.7038
apprv_num_m30.05720.02%52.89%0.7533
useratemore50maxcontimonsm120.05300.07%10.71%0.6543
cre_l30_limit_left_amt_min0.05040.06%4.18%0.7875
cre_now_l60_limit_left_d_min0.04665.74%11.24%0.8178
ps_int_amt_sum0.04650.02%1.21%0.7396
avg_tnr
平均履约月数
0.04560.02%30.34%0.6841
odgt0_m12_mob_pc0.04470.07%33.87%0.6158
cre_now_l60_used_amt_d_min0.042910.25%81.68%0.7521
max_loan_os_prcp_m120.04110.07%4.04%0.7503
odgt0_m12_num0.04070.07%34.86%0.7141
useratemore70_m6_num0.04050.07%28.42%0.6941
cre_now_l60_used_amt_m_max0.04000.06%11.21%0.6810
suc_apprv_num_m60.03950.02%34.85%0.7214
cre_now_p1m_used_rate_d0.03865.64%4.54%0.9733
useratemore70monthsubcurmonth0.03840.07%35.85%0.6077
orig_prcp_avg_m120.03720.02%3.24%0.6695
cre_l30_l90_used_amt_m_min0.03690.06%78.40%0.5989
shd_pre_repay_amt_rate0.03390.08%1.29%0.6789
cre_usedrate_u30_daydelta_max0.033012.49%83.72%0.6366
repay_amount_m6_all_shd_amt_pc0.03210.07%3.66%0.7885
ps_int_amt_avg0.03060.02%1.21%0.6095
cre_now_l30_used_amt_d_max0.03040.52%14.27%0.6000
cre_now_l30_limit_left_d_min0.02904.23%14.95%0.3058
repay_amount_m6_loan_os_prcp_pc0.02820.07%54.31%0.5912
shd_l60_min_dpd0.02782.50%51.07%0.7466
min_apply_amt0.02730.02%10.57%0.4938
useratemore90maxcontimonsm120.02690.07%24.43%0.6895
cre_now_l90_limit_left_d_min0.02666.91%9.77%0.8178
repay_amount_all_shd_amt_pc0.02540.07%5.67%0.6536
shd_instm_amt_sum0.02480.08%0.03%0.7396
useratemore90_m3_num0.02400.07%61.13%0.6941
cre_l30_l90_usedrate_m_max0.02330.06%26.85%0.7107
odgt0_m3_num0.02270.07%53.85%0.7141
cre_now_l60_usedrate_m_max0.02260.06%13.38%0.7107
m_s_ps_instm_amt0.02160.02%70.25%0.4301
min_tnr0.02150.02%53.18%0.6841
setl_rem_prcp_paym_last_month_lo0.02030.07%87.85%0.6536
cre_now_p3m_used_rate_d0.020111.82%1.60%0.7521
shd_l180_instm_amt_sum0.02000.08%0.11%0.7458
last_clear_num_m30.02000.02%57.66%0.7214
loan_os_prcp_gt0_m12_num_pc0.01960.07%61.32%0.5912
ps_prcp_amt_mon0.01830.02%4.60%0.7458
cre_now_p1m_used_rate_m0.01710.93%4.84%0.7553
cre_now_p2m_used_rate_m0.01581.66%2.23%0.7553
shd_l30_pre_repay_due_rate0.01533.11%48.56%0.7855
end_loan_to_rep_loan_freq0.01330.02%6.97%0.4280
shd_pre_repay_cnt0.01330.08%1.22%0.5711
shd_remain_actual_cnt0.01250.08%7.33%0.6008
account_age0.01240.07%4.74%0.7206
shd_l90_ps_due_cnt0.01180.08%17.68%0.6782
cre_now_l30_limit_left_d_max0.01150.72%79.30%0.7660
min_apprv_amt0.01100.02%29.17%0.5699
last_clear_num_m10.01000.02%79.20%0.7141
shd_min_dpd0.00941.06%26.93%0.7113
sum_apply_amt0.00870.02%10.45%0.4938
account_age_loan_day0.00840.07%2.11%0.7206
last_apply_freq0.00800.02%87.14%0.6518
shd_remain_actual_rate0.00790.41%66.46%0.7855
max_od_due0.00690.07%74.02%0.7415
setl_prcp_mon_ratio0.00510.07%6.05%0.4346
grade_limit_amount0.00510.02%38.22%0.4065
shd_l180_max_dpd0.00431.69%57.75%0.7885
cre_usedrate_u30_daydelta_min0.004012.49%81.81%0.4730
age0.00400.02%4.89%0.1655
setl_rem_prcp_paym_m12_loan_os_p0.00360.07%36.61%0.3691

注:共 124 个候选变量,按 IV 降序展示前 100 个;集中度 = 最大频次取值样本数 ÷ 非空样本数;相关性为皮尔逊(仅数值变量),完整数据见 variable_detail.json。

03 单变量分析(细分箱)

last_od_to_cur_freq — IV 0.1928,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1.5]4011321679519.82%-1.0084
2(1.5, inf]340693171423556.91%0.1943
f_ps_od_days_max — IV 0.1809,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]2042218144227811.15%-0.3309
2(0.5, 1.5]844180024395.20%0.4970
3(1.5, inf]921787844334.70%0.6040
last_loan_setl_tnr — IV 0.1487,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]723969942453.38%0.9456
2(0.5, 3.5]11453105429117.95%0.0426
3(3.5, inf]1938817394199410.28%-0.2400
apprv_amt_day — IV 0.1475,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 11]303442745228929.53%-0.1555
2(11, inf]773674782583.34%0.9608
suc_apprv_num_m1 — IV 0.1475,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]301692728728829.55%-0.1580
2(0.5, inf]791176432683.39%0.9446
ps_perd_no — IV 0.1423,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1.5]499148451462.93%1.0963
2(1.5, 3.5]639560033926.13%0.3229
3(3.5, inf]267022408926139.79%-0.1846
cre_now_l60_used_amt_m_min — IV 0.1300,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1.78]282012545827439.73%-0.1775
2(1.78, inf]986494574074.13%0.7402
end_loan_to_cur_freq — IV 0.1253,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, -68.5]2363521233240210.16%-0.2267
2(-68.5, -25.5]559252323606.44%0.2705
3(-25.5, inf]885384653884.38%0.6767
apprv_num_m1 — IV 0.1224,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]281522542527279.69%-0.1734
2(0.5, 1.5]535051072434.54%0.6394
3(1.5, inf]457843981803.93%0.7900
orig_prcp_avg_m3 — IV 0.1156,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 23.45]2286420513235110.28%-0.2397
2(23.45, inf]15216144177995.25%0.4869
s_ps_od_days_max — IV 0.1150,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]2061018430218010.58%-0.2713
2(0.5, inf]17470165009705.55%0.4279
max_od_days — IV 0.1123,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 7.5]337123127624367.23%0.1466
2(7.5, inf]4368365471416.35%-0.7732
last_od_num_m6 — IV 0.1092,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]299772791920586.87%0.2016
2(0.5, 3.5]4179362155813.35%-0.5358
3(3.5, inf]3924339053413.61%-0.5578
cre_l90_used_limit_amt_min — IV 0.1036,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 748.71]13694129347605.55%0.4288
2(748.71, 8516.23]201721835418189.01%-0.0934
3(8516.23, inf]4199362757213.62%-0.5585
cre_l60_used_limit_rate_min — IV 0.1034,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.28]14626138138135.56%0.4271
2(0.28, 0.92]189021715317499.25%-0.1224
3(0.92, inf]4537394958812.96%-0.5010
available_limit_amt — IV 0.1021,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.03]1238510918146711.85%-0.3987
2(0.03, 5847.69]182341699512396.79%0.2127
3(5847.69, inf]746170174445.95%0.3543
cre_now_l30_used_amt_d_min — IV 0.0995,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1]312112827329389.41%-0.0980
2(1, inf]470145471543.28%1.0231
apprv_num_m3 — IV 0.0951,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]2013918040209910.42%-0.2548
2(0.5, 2.5]1020095756256.13%0.3232
3(2.5, inf]774173154265.50%0.4373
setl_prcp_ratio — IV 0.0905,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.04]834679793674.40%0.6733
2(0.04, inf]297342695127839.36%-0.1355
loan_os_prcp_credit_amount_m6 — IV 0.0893,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 48.94]12564118756895.48%0.4412
2(48.94, 107.59]187351707416618.87%-0.0756
3(107.59, inf]6761596279911.82%-0.3959
acc_ps_perd_no — IV 0.0875,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]98278680114711.67%-0.3821
2(0.5, 7.5]132211211111108.40%-0.0162
3(7.5, inf]15032141398935.94%0.3562
t_ps_od_days_max — IV 0.0843,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]2063318517211610.26%-0.2368
2(0.5, inf]174471641310345.93%0.3587
cre_l30_l90_used_amt_m_max — IV 0.0833,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 35.01]863382194144.80%0.5828
2(35.01, 1951.38]216251973518908.74%-0.0597
3(1951.38, inf]7807696184610.84%-0.2980
od_amt_m6_loan_os_prcp_m6_pc — IV 0.0798,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0]355443284227027.60%0.0920
2(0, inf]2516206944717.77%-0.8735
cre_now_l60_used_amt_d_min — IV 0.0752,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1]279432520527389.80%-0.1106
2(1, inf]624259502924.68%0.6840
loan_os_prcp_credit_amount_m12 — IV 0.0725,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 44.21]919787294685.09%0.5202
2(44.21, 95.76]210881923118578.81%-0.0682
3(95.76, inf]7775695182410.60%-0.2732
ps_int_amt_mon — IV 0.0724,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 15.47]933887985405.78%0.3848
2(15.47, 132.38]217682000917598.08%0.0255
3(132.38, inf]6974612385112.20%-0.4325
cre_now_l30_limit_left_d_max — IV 0.0708,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1]721468873274.53%0.6418
2(1, inf]306002779828029.16%-0.1110
cre_l30_l90_used_amt_d_max — IV 0.0696,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.97]291212641127109.31%-0.1267
2(0.97, inf]883984024374.94%0.5528
shd_max_dpd — IV 0.0690,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]95658427113811.90%-0.4008
2(0.5, inf]281222613419887.07%0.1731
shd_pre_repay_due_rate — IV 0.0663,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.95]171041604510596.19%0.3158
2(0.95, inf]2056018493206710.05%-0.2110
loan_os_prcp_credit_amount_m3 — IV 0.0653,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 100.48]311102884322677.29%0.1377
2(100.48, inf]6950606888212.69%-0.4771
od_amt_m12_loan_os_prcp_m12_pc — IV 0.0649,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0]352073251726907.64%0.0865
2(0, inf]2853239445916.09%-0.7540
max_od_due — IV 0.0638,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]98078661114611.69%-0.3832
2(0.5, inf]282532625020037.09%0.1673
odgt0_m12_subcurmonth — IV 0.0635,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]98048659114511.68%-0.3825
2(0.5, inf]282562625220047.09%0.1669
cre_l60_used_limit_rate_max — IV 0.0595,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.22]417939572225.31%0.4750
2(0.22, 1.21]265812451820637.76%0.0697
3(1.21, inf]7305644086511.84%-0.3980
shd_l30_max_dpd — IV 0.0576,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]278582582320357.30%0.1535
2(0.5, inf]91018026107511.81%-0.3769
useratemore50maxcontimonsm12 — IV 0.0530,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 2.5]836579244415.27%0.4829
2(2.5, 4.5]589354014928.35%-0.0099
3(4.5, inf]238022158622169.31%-0.1294
cre_now_l60_used_amt_m_max — IV 0.0511,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 83.31]514849042444.74%0.5951
2(83.31, 2003.01]240882208320058.32%-0.0063
3(2003.01, inf]8829792890110.21%-0.2309
shd_l30_pre_repay_due_rate — IV 0.0495,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.88]6969614482511.84%-0.3781
2(0.88, 3.77]263242424620787.89%0.0709
3(3.77, inf]361034042065.71%0.4188
cre_l30_limit_left_amt_min — IV 0.0459,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, -132.06]7225636985611.85%-0.3986
2(-132.06, inf]308402854622947.44%0.1157
cre_usedrate_u30_daydelta_max — IV 0.0452,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 85.5]494146922495.04%0.5850
2(85.5, inf]283912574126509.33%-0.0776
setl_rem_prcp_paym_m3_num — IV 0.0433,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]320942924028548.89%-0.0789
2(0.5, inf]596656712954.94%0.5504
ps_int_amt_sum — IV 0.0424,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1082.7]169571578811696.89%0.1972
2(1082.7, 4579.91]145751334512308.44%-0.0218
3(4579.91, inf]6548579775111.47%-0.3623
useratemore70monthsubcurmonth — IV 0.0416,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 4.5]254292310223279.15%-0.1104
2(4.5, 9.5]603755644737.83%0.0593
3(9.5, inf]659462453495.29%0.4787
max_loan_os_prcp_m12 — IV 0.0415,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1747.28]410138602415.88%0.3679
2(1747.28, 20848.59]289052660622997.95%0.0429
3(20848.59, inf]5054444560912.05%-0.4180
avg_tnr 平均履约月数 — IV 0.0412,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 6.5]876882545145.86%0.3703
2(6.5, 10.5]126711164510268.10%0.0233
3(10.5, inf]166411503116109.67%-0.1721
odgt0_m12_num — IV 0.0409,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]1326811926134210.11%-0.2212
2(0.5, 3.5]170991575413457.87%0.0550
3(3.5, inf]769372314626.01%0.3449
useratemore70_m6_num — IV 0.0405,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1.5]15252142619916.50%0.2609
2(1.5, 5.5]178801623616449.19%-0.1156
3(5.5, inf]4928441451410.43%-0.2554
suc_apprv_num_m6 — IV 0.0404,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1.5]220111992520869.48%-0.1492
2(1.5, 3.5]781272405727.32%0.1323
3(3.5, inf]825777654925.96%0.3530
odgt0_m12_mob_pc — IV 0.0402,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 4.16]1327211930134210.11%-0.2208
2(4.16, 20.13]125321151210208.14%0.0179
3(20.13, inf]12256114697876.42%0.2735
orig_prcp_avg_m12 — IV 0.0357,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1031.97]11498107357636.64%0.2381
2(1031.97, 3819.86]156521439812548.01%0.0348
3(3819.86, inf]109309797113310.37%-0.2487
shd_pre_repay_amt_rate — IV 0.0350,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.62]12414116038116.53%0.2551
2(0.62, 0.75]881481166987.92%0.0477
3(0.75, inf]168311519116409.74%-0.1797
m_s_ps_instm_amt — IV 0.0347,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 33.56]268902491619747.34%0.1295
2(33.56, 374.29]5391484754410.09%-0.2188
3(374.29, inf]5799516763210.90%-0.3048
repay_amount_m6_all_shd_amt_pc — IV 0.0340,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 93.99]132681195713119.88%-0.1952
2(93.99, 99.49]197411819115507.85%0.0570
3(99.49, inf]505147632885.70%0.4000
useratemore90maxcontimonsm12 — IV 0.0335,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]929887235756.18%0.3136
2(0.5, 4.5]216531981018438.51%-0.0309
3(4.5, inf]7109637873110.28%-0.2395
cre_l30_l90_used_amt_m_min — IV 0.0320,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.63]298562718526718.95%-0.0853
2(0.63, inf]820977304795.84%0.3756
last_clear_num_m3 — IV 0.0302,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]219561990820489.33%-0.1317
2(0.5, 2.5]1046097027587.25%0.1435
3(2.5, inf]566453203446.07%0.3326
useratemore90_m3_num — IV 0.0302,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1.5]293832714722367.61%0.0909
2(1.5, 2.5]397936233568.95%-0.0856
3(2.5, inf]4698414155711.86%-0.3996
repay_amount_m3_loan_os_prcp_pc — IV 0.0283,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, -4999]480445542505.20%0.4966
2(-4999, inf]332563035728998.72%-0.0571
ps_int_amt_avg — IV 0.0271,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 13.08]163091517011396.98%0.1832
2(13.08, 38.25]143551310912468.68%-0.0526
3(38.25, inf]7416665176510.32%-0.2433
min_apply_amt — IV 0.0259,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1960]3831341541610.86%-0.3007
2(1960, 7300]267612449822638.46%-0.0240
3(7300, inf]748870174716.29%0.2953
shd_instm_amt_sum — IV 0.0258,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 17727.41]154541433311217.25%0.1427
2(17727.41, 134568.02]178261634114858.33%-0.0074
3(134568.02, inf]4779423654311.36%-0.3514
shd_l180_instm_amt_sum — IV 0.0217,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 5366.87]147451367710687.24%0.1442
2(5366.87, 34552.49]187101713715738.41%-0.0174
3(34552.49, inf]4604409650811.03%-0.3184
repay_amount_all_shd_amt_pc — IV 0.0216,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 97.22]274692502624438.89%-0.0790
2(97.22, 98.42]643159674647.22%0.1484
3(98.42, inf]416039182425.82%0.3787
min_tnr — IV 0.0215,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 5]229492126016897.36%0.1267
2(5, inf]151311367014619.66%-0.1699
cre_now_l60_limit_left_d_min — IV 0.0213,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.97]8036718585110.59%-0.2536
2(0.97, inf]278652569421717.79%0.0842
cre_now_l60_usedrate_m_max — IV 0.0210,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.08]10747100407076.58%0.2478
2(0.08, inf]273182487524438.94%-0.0849
tothismonthodcontireducenumm12 — IV 0.0209,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]328993004428558.68%-0.0521
2(0.5, inf]516148672945.70%0.4009
cre_l30_l90_usedrate_m_max — IV 0.0207,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.08]151961413910576.96%0.1880
2(0.08, inf]228692077620939.15%-0.1103
setl_rem_prcp_paym_last_month_lo — IV 0.0203,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0]363363325630808.48%-0.0264
2(0, inf]17241655694.00%0.7717
cre_now_l90_limit_left_d_min — IV 0.0203,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.99]9682869199110.24%-0.2180
2(0.99, inf]257742378719877.71%0.0932
cre_now_l30_limit_left_d_min — IV 0.0202,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1]115961046211349.78%-0.1697
2(1, 1.73]199591835516048.04%0.0457
3(1.73, inf]492046013196.48%0.2772
loan_os_prcp_gt0_m12_num_pc — IV 0.0200,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 60.39]444241952475.56%0.4265
2(60.39, 92.32]1008892388508.43%-0.0199
3(92.32, inf]235302147820528.72%-0.0575
last_clear_num_m1 — IV 0.0197,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]301612750926528.79%-0.0667
2(0.5, inf]791974214986.29%0.2955
ps_prcp_amt_mon — IV 0.0189,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 89.38]401237632496.21%0.3096
2(89.38, 955.47]206301900416267.88%0.0526
3(955.47, inf]134381216312759.49%-0.1505
tothismonthodcontiaddnumm12 — IV 0.0187,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]330573046325947.85%0.0576
2(0.5, inf]5003444855511.09%-0.3245
cre_now_p2m_used_rate_m — IV 0.0177,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0]11203104277766.93%0.1961
2(0, 0.08]447241253477.76%0.0736
3(0.08, inf]217821979519879.12%-0.1031
shd_pre_repay_cnt — IV 0.0175,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 25.5]956088766847.15%0.1575
2(25.5, 362.5]215941983917558.13%0.0195
3(362.5, inf]6905619571010.28%-0.2395
last_apply_freq — IV 0.0169,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 6.5]333023067926237.88%0.0533
2(6.5, inf]4778425152711.03%-0.3182
cre_now_p1m_used_rate_d — IV 0.0166,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.68]659361624316.54%0.2927
2(0.68, inf]293472669826499.03%-0.0569
odgt0_m3_num — IV 0.0155,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1.5]300542771923357.77%0.0684
2(1.5, inf]8006719281410.17%-0.2270
account_age — IV 0.0148,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 21.5]178221650113217.41%0.1193
2(21.5, 29.5]751468976178.21%0.0083
3(29.5, inf]127241151312119.52%-0.1537
cre_now_p3m_used_rate_d — IV 0.0145,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.41]454642532936.45%0.3289
2(0.41, inf]290392639826419.09%-0.0442
cre_now_l30_used_amt_d_max — IV 0.0145,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.68]743369484856.53%0.2605
2(0.68, inf]304582779626628.74%-0.0557
shd_remain_actual_cnt — IV 0.0129,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 3.5]712966624676.55%0.2522
2(3.5, inf]309302824826828.67%-0.0512
shd_pre_repay_cnt_rate — IV 0.0124,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]523249043286.27%0.2991
2(0.5, inf]328273000628218.59%-0.0414
cre_now_p1m_used_rate_m — IV 0.0122,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.02]11149103667837.02%0.1799
2(0.02, inf]265852423923468.82%-0.0680
cre_usedrate_u30_daydelta_min — IV 0.0116,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 2.5]273532486924849.08%-0.0474
2(2.5, inf]597955644156.94%0.2446
shd_remain_actual_rate — IV 0.0084,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.42]384936032466.39%0.2814
2(0.42, 0.74]486544743918.04%0.0346
3(0.74, inf]292192670925108.59%-0.0380
max_int_rat — IV 0.0079,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.21]372634902366.33%0.2879
2(0.21, inf]343543144029148.48%-0.0274
cre_now_l60_limit_left_d_max — IV 0.0073,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1]372193418430358.15%0.0130
2(1, inf]6855929313.58%-0.5576
grade_limit_amount — IV 0.0072,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 180500]197591800217578.89%-0.0791
2(180500, inf]183211692813937.60%0.0916
repay_amount_m6_loan_os_prcp_pc — IV 0.0070,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, -4999]682863534756.96%0.1877
2(-4999, inf]312322855826748.56%-0.0374
bh_cur_od_num — IV 0.0064,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, -49999.5]150401368213589.03%-0.0959
2(-49999.5, inf]230402124817927.78%0.0670
shd_l90_ps_due_cnt — IV 0.0053,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 5.5]11841109578847.47%0.1116
2(5.5, inf]262182395322658.64%-0.0472
shd_min_dpd — IV 0.0050,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, -371.5]534648385089.50%-0.1492
2(-371.5, -142.5]172641582414408.34%-0.0061
3(-142.5, inf]150761389811787.81%0.0650
sum_apply_amt — IV 0.0049,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 4180]181321653116018.83%-0.0713
2(4180, inf]199481839915497.77%0.0687
account_age_loan_day — IV 0.0048,2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 10.56]273282515821707.94%0.0447
2(10.56, inf]1073297539799.12%-0.1069
setl_rem_prcp_paym_m12_loan_os_p — IV 0.0047,3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, -9998.5]678962725177.62%0.0901
2(-9998.5, 54.41]240622209919638.16%0.0153
3(54.41, inf]720965406699.28%-0.1258

注:统计全部候选变量的自动分箱明细(初分组),按 IV 降序展示前 100 个变量,共 124 个;完整数据见 feature_report.json。

04 单变量分析(粗分箱)

f_ps_od_days_max — 3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]2042218144227811.15%-0.3309
2(0.5, 1.5]844180024395.20%0.4970
3(1.5, inf]921787844334.70%0.6040
2026-11-03T14:31:22.734530 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(11.15% → 4.70%),WOE 单调上升(-0.3309 → 0.6040),符合风控分箱预期。IV=0.1809,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.5])为坏率最高区间:坏率高达 11.15%,可作为拒绝或重点风控策略的阈值参考

apprv_amt_day — 细分箱 3 箱 / 粗分箱 2 箱

细分箱(初始分箱)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 11]303442745228929.53%-0.1555
2(11, 2385.5]398838651233.08%1.0416
3(2385.5, inf]374836131353.60%0.8811

粗分箱(WOE 单调性合并后,入模口径)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 11]303442745228929.53%-0.1555
2(11, inf]773674782583.34%0.9608
2026-11-03T14:31:22.842621 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.53% → 3.34%),WOE 单调上升(-0.1555 → 0.9608),符合风控分箱预期。IV=0.1475,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 11])为坏率最高区间:坏率高达 9.53%,可作为拒绝或重点风控策略的阈值参考

suc_apprv_num_m1 — 2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]301692728728829.55%-0.1580
2(0.5, inf]791176432683.39%0.9446
2026-11-03T14:31:22.946107 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.55% → 3.39%),WOE 单调上升(-0.1580 → 0.9446),符合风控分箱预期。IV=0.1475,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.5])为坏率最高区间:坏率高达 9.55%,可作为拒绝或重点风控策略的阈值参考

cre_now_l60_used_amt_m_min — 细分箱 3 箱 / 粗分箱 2 箱

细分箱(初始分箱)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1.78]282012545827439.73%-0.1775
2(1.78, 3000.02]587256602123.61%0.8791
3(3000.02, inf]399237971954.88%0.5635

粗分箱(WOE 单调性合并后,入模口径)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1.78]282012545827439.73%-0.1775
2(1.78, inf]986494574074.13%0.7402
2026-11-03T14:31:23.049364 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.73% → 4.13%),WOE 单调上升(-0.1775 → 0.7402),符合风控分箱预期。IV=0.1300,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 1.78])为坏率最高区间:坏率高达 9.73%,可作为拒绝或重点风控策略的阈值参考

end_loan_to_cur_freq — 3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, -68.5]2363521233240210.16%-0.2267
2(-68.5, -25.5]559252323606.44%0.2705
3(-25.5, inf]885384653884.38%0.6767
2026-11-03T14:31:23.153554 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(10.16% → 4.38%),WOE 单调上升(-0.2267 → 0.6767),符合风控分箱预期。IV=0.1253,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 1 箱((-inf, -68.5])为坏率最高区间:坏率高达 10.16%,可作为拒绝或重点风控策略的阈值参考

apprv_num_m1 — 3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]281522542527279.69%-0.1734
2(0.5, 1.5]535051072434.54%0.6394
3(1.5, inf]457843981803.93%0.7900
2026-11-03T14:31:23.261200 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.69% → 3.93%),WOE 单调上升(-0.1734 → 0.7900),符合风控分箱预期。IV=0.1224,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.5])为坏率最高区间:坏率高达 9.69%,可作为拒绝或重点风控策略的阈值参考

s_ps_od_days_max — 细分箱 3 箱 / 粗分箱 2 箱

细分箱(初始分箱)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]2061018430218010.58%-0.2713
2(0.5, 3.5]13474127916835.07%0.5241
3(3.5, inf]399637092877.18%0.1531

粗分箱(WOE 单调性合并后,入模口径)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]2061018430218010.58%-0.2713
2(0.5, inf]17470165009705.55%0.4279
2026-11-03T14:31:23.410312 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(10.58% → 5.55%),WOE 单调上升(-0.2713 → 0.4279),符合风控分箱预期。IV=0.1150,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.5])为坏率最高区间:坏率高达 10.58%,可作为拒绝或重点风控策略的阈值参考

max_od_days — 细分箱 3 箱 / 粗分箱 2 箱

细分箱(初始分箱)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]98278680114711.67%-0.3821
2(0.5, 7.5]238852259612895.40%0.4580
3(7.5, inf]4368365471416.35%-0.7732

粗分箱(WOE 单调性合并后,入模口径)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 7.5]337123127624367.23%0.1466
2(7.5, inf]4368365471416.35%-0.7732
2026-11-03T14:31:23.509686 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调上升(7.23% → 16.35%),WOE 单调下降(0.1466 → -0.7732),符合风控分箱预期。IV=0.1123,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 2 箱((7.5, inf])为坏率最高区间:坏率高达 16.35%,可作为拒绝或重点风控策略的阈值参考

cre_l90_used_limit_amt_min — 3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 748.71]13694129347605.55%0.4288
2(748.71, 8516.23]201721835418189.01%-0.0934
3(8516.23, inf]4199362757213.62%-0.5585
2026-11-03T14:31:23.617917 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调上升(5.55% → 13.62%),WOE 单调下降(0.4288 → -0.5585),符合风控分箱预期。IV=0.1036,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 3 箱((8516.23, inf])为坏率最高区间:坏率高达 13.62%,可作为拒绝或重点风控策略的阈值参考

cre_l60_used_limit_rate_min — 3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.28]14626138138135.56%0.4271
2(0.28, 0.92]189021715317499.25%-0.1224
3(0.92, inf]4537394958812.96%-0.5010
2026-11-03T14:31:23.732037 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调上升(5.56% → 12.96%),WOE 单调下降(0.4271 → -0.5010),符合风控分箱预期。IV=0.1034,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 3 箱((0.92, inf])为坏率最高区间:坏率高达 12.96%,可作为拒绝或重点风控策略的阈值参考

available_limit_amt — 3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.03]1238510918146711.85%-0.3987
2(0.03, 5847.69]182341699512396.79%0.2127
3(5847.69, inf]746170174445.95%0.3543
2026-11-03T14:31:23.846981 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(11.85% → 5.95%),WOE 单调上升(-0.3987 → 0.3543),符合风控分箱预期。IV=0.1021,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.03])为坏率最高区间:坏率高达 11.85%,可作为拒绝或重点风控策略的阈值参考

cre_now_l30_used_amt_d_min — 2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1]312112827329389.41%-0.0980
2(1, inf]470145471543.28%1.0231
2026-11-03T14:31:23.957351 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.41% → 3.28%),WOE 单调上升(-0.0980 → 1.0231),符合风控分箱预期。IV=0.0995,该变量区分度偏弱但可用。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 1])为坏率最高区间:坏率高达 9.41%,可作为拒绝或重点风控策略的阈值参考

setl_prcp_ratio — 细分箱 3 箱 / 粗分箱 2 箱

细分箱(初始分箱)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.04]834679793674.40%0.6733
2(0.04, 0.66]2108618922216410.26%-0.2376
3(0.66, inf]864880296197.16%0.1568

粗分箱(WOE 单调性合并后,入模口径)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.04]834679793674.40%0.6733
2(0.04, inf]297342695127839.36%-0.1355
2026-11-03T14:31:24.064008 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调上升(4.40% → 9.36%),WOE 单调下降(0.6733 → -0.1355),符合风控分箱预期。IV=0.0905,该变量区分度偏弱但可用。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 2 箱((0.04, inf])为坏率最高区间:坏率高达 9.36%,可作为拒绝或重点风控策略的阈值参考

loan_os_prcp_credit_amount_m6 — 3 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 48.94]12564118756895.48%0.4412
2(48.94, 107.59]187351707416618.87%-0.0756
3(107.59, inf]6761596279911.82%-0.3959
2026-11-03T14:31:24.171988 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调上升(5.48% → 11.82%),WOE 单调下降(0.4412 → -0.3959),符合风控分箱预期。IV=0.0893,该变量区分度偏弱但可用。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 3 箱((107.59, inf])为坏率最高区间:坏率高达 11.82%,可作为拒绝或重点风控策略的阈值参考

t_ps_od_days_max — 细分箱 3 箱 / 粗分箱 2 箱

细分箱(初始分箱)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]2063318517211610.26%-0.2368
2(0.5, 3.5]13145125016444.90%0.5599
3(3.5, inf]430239123909.07%-0.1003

粗分箱(WOE 单调性合并后,入模口径)

分组号分栏TotalGoodBadBad RateWOE
1(-inf, 0.5]2063318517211610.26%-0.2368
2(0.5, inf]174471641310345.93%0.3587
2026-11-03T14:31:24.277626 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(10.26% → 5.93%),WOE 单调上升(-0.2368 → 0.3587),符合风控分箱预期。IV=0.0843,该变量区分度偏弱但可用。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.5])为坏率最高区间:坏率高达 10.26%,可作为拒绝或重点风控策略的阈值参考

cre_now_l60_used_amt_d_min — 2 箱
分组号分栏TotalGoodBadBad RateWOE
1(-inf, 1]279432520527389.80%-0.1106
2(1, inf]624259502924.68%0.6840
2026-11-03T14:31:24.385924 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.80% → 4.68%),WOE 单调上升(-0.1106 → 0.6840),符合风控分箱预期。IV=0.0752,该变量区分度偏弱但可用。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 1])为坏率最高区间:坏率高达 9.80%,可作为拒绝或重点风控策略的阈值参考

注:统计入模变量经合并调整后的最终分箱明细(开发样本口径)。20261022 口径:发生 WOE 单调性合并的变量双列展示——细分箱为初始(决策树)分箱,粗分箱为单调合并后的入模口径(本次共 6 个变量发生合并);未合并变量仅一套分箱。

05 变量筛选

项目内容
筛选优先级黑名单 > 白名单 > 自动规则
IV 下限0.022219602542548865
缺失率上限0.8
集中度上限0.9
相关性上限(皮尔逊)0.7
PSI 上限0.25
重要性 Top-K16
黑名单无
白名单无

候选变量 124,成功分箱 124,入模 16

变量白名单入模IV缺失率集中度IV门槛相关性PSI剔除原因
ps_perd_no—N0.1423PASSPASSPASS—0.3194PSI>0.25
age—N0.0044PASSPASSFAIL——IV<0.022219602542548865
max_int_rat—N0.0079PASSPASSFAIL——IV<0.022219602542548865
min_int_rat—N0.0022PASSPASSFAIL——IV<0.022219602542548865
avg_tnr
平均履约月数
—N0.0412PASSPASSPASSPASS0.0223重要性排名超出Top-16
min_tnr—N0.0215PASSPASSFAIL——IV<0.022219602542548865
sum_apply_amt—N0.0049PASSPASSFAIL——IV<0.022219602542548865
min_apply_amt—N0.0259PASSPASSPASSPASS0.0360重要性排名超出Top-16
min_apprv_amt—N0.0041PASSPASSFAIL——IV<0.022219602542548865
apprv_amt_day—Y0.1475PASSPASSPASSPASS0.1320
end_loan_to_rep_loan_freq—N0.0004PASSPASSFAIL——IV<0.022219602542548865
end_loan_to_cur_freq—Y0.1253PASSPASSPASSPASS0.2135
f_rep_loan_to_f_apply_dt_freq—N0.0010PASSPASSFAIL——IV<0.022219602542548865
ps_prcp_amt_mon—N0.0189PASSPASSFAIL——IV<0.022219602542548865
ps_int_amt_mon—N0.0724PASSPASSPASSFAIL0.0962与 cre_l90_used_limit_amt_min 相关>0.7
ps_int_amt_sum—N0.0424PASSPASSPASSPASS0.0254重要性排名超出Top-16
ps_prcp_amt_avg—N0.0043PASSPASSFAIL——IV<0.022219602542548865
ps_int_amt_avg—N0.0271PASSPASSPASSPASS0.0171重要性排名超出Top-16
setl_prcp_mon_ratio—N0.0042PASSPASSFAIL——IV<0.022219602542548865
acc_ps_perd_no—N0.0875PASSPASSPASS—0.4557PSI>0.25
max_od_days—Y0.1123PASSPASSPASSPASS0.1885
m_s_ps_instm_amt—N0.0347PASSPASSPASS—0.2758PSI>0.25
f_ps_od_days_max—Y0.1809PASSPASSPASSPASS0.1497
s_ps_od_days_max—Y0.1150PASSPASSPASSPASS0.1764
t_ps_od_days_max—Y0.0843PASSPASSPASSPASS0.2128
ovd_days_d3_num_m1—N0.0000PASSFAIL———集中度>0.9
orig_prcp_avg_m3—N0.1156PASSPASSPASS—0.2544PSI>0.25
orig_prcp_avg_m12—N0.0357PASSPASSPASSPASS0.0231重要性排名超出Top-16
suc_apply_trench_num—N0.0000PASSFAIL———集中度>0.9
suc_apply_top_pc_code_num—N0.0022PASSPASSFAIL——IV<0.022219602542548865
suc_apprv_num_m1—Y0.1475PASSPASSPASSPASS0.1385
apprv_num_m1—Y0.1224PASSPASSPASSPASS0.1073
suc_apply_num_m3—N0.0000PASSFAIL———集中度>0.9
apprv_num_m3—N0.0951PASSPASSPASSFAIL0.2387与 apprv_num_m1 相关>0.7
suc_apprv_num_m6—N0.0404PASSPASSPASSPASS0.0130重要性排名超出Top-16
apply_num_m12—N0.0007PASSPASSFAIL——IV<0.022219602542548865
last_apply_od_num_m12—N0.0000PASSFAIL———集中度>0.9
last_apply_freq—N0.0169PASSPASSFAIL——IV<0.022219602542548865
dec_loan_num—N0.0000PASSFAIL———集中度>0.9
bh_cur_od_num—N0.0064PASSPASSFAIL——IV<0.022219602542548865
hj_cur_od_num—N0.0005PASSFAIL———集中度>0.9
setl_prcp_ratio—Y0.0905PASSPASSPASSPASS0.2370
last_loan_setl_tnr—N0.1487PASSPASSPASS—0.2771PSI>0.25
last_od_num_m6—N0.1092PASSPASSPASS—0.5997PSI>0.25
last_od_to_cur_freq—N0.1928PASSPASSPASS—0.2883PSI>0.25
last_clear_num_m1—N0.0197PASSPASSFAIL——IV<0.022219602542548865
last_clear_num_m3—N0.0302PASSPASSPASSFAIL0.0129与 suc_apprv_num_m6 相关>0.7
grade_limit_amount—N0.0072PASSPASSFAIL——IV<0.022219602542548865
exposure—N0.0002PASSPASSFAIL——IV<0.022219602542548865
available_limit_amt—Y0.1021PASSPASSPASSPASS0.0077
account_age—N0.0148PASSPASSFAIL——IV<0.022219602542548865
account_age_loan_day—N0.0048PASSPASSFAIL——IV<0.022219602542548865
max_od_due—N0.0638PASSPASSPASS—0.4345PSI>0.25
odgt0_m12_mob_pc—N0.0402PASSPASSPASS—0.8830PSI>0.25
odgt0_m3_num—N0.0155PASSPASSFAIL——IV<0.022219602542548865
odgt0_m12_num—N0.0409PASSPASSPASS—0.8575PSI>0.25
odgt0_m12_subcurmonth—N0.0635PASSPASSPASS—1.2145PSI>0.25
tothismonthodcontiaddnumm12—N0.0187PASSPASSFAIL——IV<0.022219602542548865
tothismonthodcontireducenumm12—N0.0209PASSPASSFAIL——IV<0.022219602542548865
od_amt_m6_loan_os_prcp_m6_pc—N0.0798PASSFAIL———集中度>0.9
od_amt_m12_loan_os_prcp_m12_pc—N0.0649PASSFAIL———集中度>0.9
avg_od_amt_m6_orig_prcp_pc—N0.0000PASSFAIL———集中度>0.9
max_loan_os_prcp_m12—N0.0415PASSPASSPASSPASS0.0539重要性排名超出Top-16
loan_os_prcp_gt0_m3_num_pc—N0.0004PASSPASSFAIL——IV<0.022219602542548865
loan_os_prcp_gt0_m12_num_pc—N0.0200PASSPASSFAIL——IV<0.022219602542548865
repay_amount_m3_loan_os_prcp_pc—N0.0283PASSPASSPASS—0.3012PSI>0.25
repay_amount_m6_loan_os_prcp_pc—N0.0070PASSPASSFAIL——IV<0.022219602542548865
repay_amount_all_shd_amt_pc—N0.0216PASSPASSFAIL——IV<0.022219602542548865
repay_amount_m6_all_shd_amt_pc—N0.0340PASSPASSPASSPASS0.0382重要性排名超出Top-16
setl_rem_prcp_paym_last_month_lo—N0.0203PASSPASSFAIL——IV<0.022219602542548865
setl_rem_prcp_paym_m12_loan_os_p—N0.0047PASSPASSFAIL——IV<0.022219602542548865
setl_rem_prcp_paym_m3_num—N0.0433PASSPASSPASSPASS0.0163重要性排名超出Top-16
setl_rem_prcp_paym_m12_num—N0.0042PASSPASSFAIL——IV<0.022219602542548865
loan_os_prcp_credit_amount_m3—N0.0653PASSPASSPASSPASS0.0000重要性排名超出Top-16
loan_os_prcp_credit_amount_m6—Y0.0893PASSPASSPASSPASS0.0360
loan_os_prcp_credit_amount_m12—N0.0725PASSPASSPASSPASS0.0684重要性排名超出Top-16
useratemore50maxcontimonsm12—N0.0530PASSPASSPASSPASS0.2260重要性排名超出Top-16
useratemore70_m6_num—N0.0405PASSPASSPASSPASS0.0705重要性排名超出Top-16
useratemore70monthsubcurmonth—N0.0416PASSPASSPASSPASS0.0847重要性排名超出Top-16
useratemore90_m3_num—N0.0302PASSPASSPASSPASS0.0012重要性排名超出Top-16
useratemore90maxcontimonsm12—N0.0335PASSPASSPASSPASS0.0908重要性排名超出Top-16
shd_max_dpd—N0.0690PASSPASSPASS—0.3133PSI>0.25
shd_min_dpd—N0.0050PASSPASSFAIL——IV<0.022219602542548865
shd_pre_repay_cnt—N0.0175PASSPASSFAIL——IV<0.022219602542548865
shd_perd_max—N0.0003PASSFAIL———集中度>0.9
shd_instm_amt_sum—N0.0258PASSPASS———唯一率过高(ID列)
shd_l30_max_dpd—N0.0576PASSPASSPASS—1.1956PSI>0.25
shd_l60_min_dpd—N0.0022PASSPASSFAIL——IV<0.022219602542548865
shd_l90_ps_due_cnt—N0.0053PASSPASSFAIL——IV<0.022219602542548865
shd_l180_max_dpd—N0.0043PASSPASSFAIL——IV<0.022219602542548865
shd_l180_instm_amt_sum—N0.0217PASSPASS———唯一率过高(ID列)
shd_remain_actual_cnt—N0.0129PASSPASSFAIL——IV<0.022219602542548865
shd_pre_repay_cnt_rate—N0.0124PASSPASSFAIL——IV<0.022219602542548865
shd_pre_repay_amt_rate—N0.0350PASSPASSPASSPASS0.0381重要性排名超出Top-16
shd_pre_repay_30d_amt_rate—N0.0006PASSPASSFAIL——IV<0.022219602542548865
shd_pre_repay_due_rate—N0.0663PASSPASSPASS—0.2564PSI>0.25
shd_l30_pre_repay_due_rate—N0.0495PASSPASSPASS—0.8874PSI>0.25
shd_remain_actual_rate—N0.0084PASSPASSFAIL——IV<0.022219602542548865
cre_usedrate_u30_daydelta_min—N0.0116PASSPASSFAIL——IV<0.022219602542548865
cre_usedrate_u30_daydelta_max—N0.0452PASSPASSPASSPASS0.2464重要性排名超出Top-16
cre_l30_limit_left_amt_min—N0.0459PASSPASSPASSFAIL0.0010与 available_limit_amt 相关>0.7
cre_l60_used_limit_rate_max—N0.0595PASSPASSPASSFAIL0.0126与 cre_l60_used_limit_rate_min 相关>0.7
cre_l60_used_limit_rate_min—Y0.1034PASSPASSPASSPASS0.0284
cre_l90_used_limit_amt_min—Y0.1036PASSPASSPASSPASS0.0288
cre_now_p1m_used_rate_d—N0.0166PASSPASSFAIL——IV<0.022219602542548865
cre_now_p3m_used_rate_d—N0.0145PASSPASSFAIL——IV<0.022219602542548865
cre_now_p1m_used_rate_m—N0.0122PASSPASSFAIL——IV<0.022219602542548865
cre_now_p2m_used_rate_m—N0.0177PASSPASSFAIL——IV<0.022219602542548865
cre_now_l30_used_amt_d_max—N0.0145PASSPASSFAIL——IV<0.022219602542548865
cre_now_l30_used_amt_d_min—Y0.0995PASSPASSPASSPASS0.0333
cre_now_l30_limit_left_d_max—N0.0708PASSPASSPASSPASS0.0105重要性排名超出Top-16
cre_now_l30_limit_left_d_min—N0.0202PASSPASSFAIL——IV<0.022219602542548865
cre_now_l60_used_amt_d_min—Y0.0752PASSPASSPASSPASS0.0572
cre_now_l60_limit_left_d_max—N0.0073PASSPASSFAIL——IV<0.022219602542548865
cre_now_l60_limit_left_d_min—N0.0213PASSPASSFAIL——IV<0.022219602542548865
cre_now_l60_used_amt_m_max—N0.0511PASSPASSPASSPASS0.2386重要性排名超出Top-16
cre_now_l60_used_amt_m_min—Y0.1300PASSPASSPASSPASS0.0710
cre_now_l60_usedrate_m_max—N0.0210PASSPASSFAIL——IV<0.022219602542548865
cre_now_l90_limit_left_d_min—N0.0203PASSPASSFAIL——IV<0.022219602542548865
cre_l30_l90_used_amt_d_max—N0.0696PASSPASSPASS—0.3032PSI>0.25
cre_l30_l90_used_amt_m_max—N0.0833PASSPASSPASS—0.3006PSI>0.25
cre_l30_l90_used_amt_m_min—N0.0320PASSPASSPASSPASS0.0504重要性排名超出Top-16
cre_l30_l90_usedrate_m_max—N0.0207PASSPASSFAIL——IV<0.022219602542548865
cre_l30_l90_usedrate_m_min—N0.0034PASSPASSFAIL——IV<0.022219602542548865

06 回归结果

硬门槛:P≥0.1 / VIF>4.0 迭代剔除,共 3 个变量被剔除(4 轮):suc_apprv_num_m1、apprv_num_m1、cre_now_l60_used_amt_m_min。

ParameterLabelEstimate(无惩罚)Standard ErrorWald Chi-SquarePr > ChiSqStandardized Estimate
InterceptIntercept-2.40810.020014567.02<0.0001—
WOE_f_ps_od_days_maxf_ps_od_days_max-0.75460.0567176.81<0.0001-0.3337
WOE_apprv_amt_dayapprv_amt_day-0.42820.077230.77<0.0001-0.1923
WOE_end_loan_to_cur_freqend_loan_to_cur_freq-0.39690.075527.61<0.0001-0.1516
WOE_s_ps_od_days_maxs_ps_od_days_max-0.62170.070777.36<0.0001-0.2166
WOE_max_od_daysmax_od_days-1.14140.0535455.58<0.0001-0.3346
WOE_cre_l90_used_limit_amt_mincre_l90_used_limit_amt_min-0.72850.0689111.77<0.0001-0.2341
WOE_cre_l60_used_limit_rate_mincre_l60_used_limit_rate_min-0.40650.068135.68<0.0001-0.1320

注:Estimate 列为无惩罚(精确推断)拟合;生产系数保持 L2 正则(model_results.json / score.py 口径),两套估计并列展示。标准误基于无惩罚似然观测信息矩阵。

07 VIF

VIF

ParameterVIF
f_ps_od_days_max1.55719
apprv_amt_day2.28322
end_loan_to_cur_freq2.29083
s_ps_od_days_max1.56570
max_od_days1.05575
cre_l90_used_limit_amt_min1.31882
cre_l60_used_limit_rate_min1.33564

相关系数(WOE 转换后,开发样本)

f_ps_od_days_maxapprv_amt_dayend_loan_to_cur_freqs_ps_od_days_maxmax_od_dayscre_l90_used_limit_amt_mincre_l60_used_limit_rate_min
f_ps_od_days_max1.0000.0690.0100.594-0.0870.0550.071
apprv_amt_day0.0691.0000.7440.0200.1460.1410.202
end_loan_to_cur_freq0.0100.7441.000-0.0530.1640.1540.191
s_ps_od_days_max0.5940.020-0.0531.000-0.1260.0320.034
max_od_days-0.0870.1460.164-0.1261.000-0.0670.019
cre_l90_used_limit_amt_min0.0550.1410.1540.032-0.0671.0000.479
cre_l60_used_limit_rate_min0.0710.2020.1910.0340.0190.4791.000

08 最终评分卡

标准分600.0
标准ODD20.0
PDO50.0
变量分栏WOECoefficientScorecard Points
f_ps_od_days_max(-inf, 0.5]-0.3309-0.714262
(0.5, 1.5]0.4970-0.7142105
(1.5, inf]0.6040-0.7142110
apprv_amt_day(-inf, 11]-0.1555-0.410875
(11, inf]0.9608-0.4108108
end_loan_to_cur_freq(-inf, -68.5]-0.2267-0.391573
(-68.5, -25.5]0.2705-0.391587
(-25.5, inf]0.6767-0.391598
s_ps_od_days_max(-inf, 0.5]-0.2713-0.570168
(0.5, inf]0.4279-0.570197
max_od_days(-inf, 7.5]0.1466-1.034290
(7.5, inf]-0.7732-1.034222
cre_l90_used_limit_amt_min(-inf, 748.71]0.4288-0.6586100
(748.71, 8516.23]-0.0934-0.658675
(8516.23, inf]-0.5585-0.658653
cre_l60_used_limit_rate_min(-inf, 0.28]0.4271-0.391591
(0.28, 0.92]-0.1224-0.391576
(0.92, inf]-0.5010-0.391565

09 模型表现

模型区分度评估(按样本)

模型样本KSAUCGINI
LR + WOE 评分卡开发样本0.30980.70500.4100
LR + WOE 评分卡测试样本0.31030.70420.4084
LR + WOE 评分卡训练+测试样本0.30940.70470.4094
LR + WOE 评分卡跨时间验证样本0.26860.67950.3590

LR + WOE 评分卡 — 4 样本评估表(score_cut 粒度)

开发样本(KS 0.3098,AUC 0.7050)

cutscntbadgoodbad_ratepopulation_pctcum_bad_ratecum_good_rate
[-inf, 508)4047802324519.82%10.63%25.45%9.29%
[508, 519)4196613358314.61%11.02%44.91%19.54%
[519, 537)3223387283612.01%8.46%57.19%27.66%
[537, 556)377835134279.29%9.92%68.33%37.47%
[556, 567)381426735477.00%10.01%76.80%47.62%
[567, 585)382624335836.35%10.05%84.51%57.88%
[585, 602)390617837284.56%10.26%90.16%68.55%
[602, 618)431514941663.45%11.33%94.89%80.47%
[618, 645)335811132473.31%8.82%98.41%89.77%
[645, inf)36255035751.38%9.52%100.00%100.00%
2026-11-03T14:31:24.511854 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-11-03T14:31:24.692982 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

测试样本(KS 0.3103,AUC 0.7042)

cutscntbadgoodbad_ratepopulation_pctcum_bad_ratecum_good_rate
[-inf, 508)2680515216519.22%10.55%24.52%9.30%
[508, 519)2736403233314.73%10.78%43.71%19.31%
[519, 537)2169274189512.63%8.54%56.76%27.45%
[537, 556)250923322769.29%9.88%67.86%37.22%
[556, 567)258418424007.12%10.18%76.62%47.52%
[567, 585)249916223376.48%9.84%84.33%57.56%
[585, 602)260412524794.80%10.26%90.29%68.20%
[602, 618)285010127493.54%11.22%95.10%80.01%
[618, 645)23466922772.94%9.24%98.38%89.78%
[645, inf)24143423801.41%9.51%100.00%100.00%
2026-11-03T14:31:24.819209 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-11-03T14:31:24.942087 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

训练+测试样本(KS 0.3094,AUC 0.7047)

cutscntbadgoodbad_ratepopulation_pctcum_bad_ratecum_good_rate
[-inf, 508)67271317541019.58%10.60%25.08%9.29%
[508, 519)69321016591614.66%10.92%44.43%19.45%
[519, 537)5392661473112.26%8.49%57.02%27.58%
[537, 556)628758457039.29%9.90%68.14%37.37%
[556, 567)639845159477.05%10.08%76.73%47.58%
[567, 585)632540559206.40%9.96%84.44%57.75%
[585, 602)651030362074.65%10.26%90.21%68.41%
[602, 618)716525069153.49%11.29%94.97%80.29%
[618, 645)570418055243.16%8.99%98.40%89.77%
[645, inf)60398459551.39%9.51%100.00%100.00%
2026-11-03T14:31:25.068234 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-11-03T14:31:25.191175 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

跨时间验证样本(KS 0.2686,AUC 0.6795)

cutscntbadgoodbad_ratepopulation_pctcum_bad_ratecum_good_rate
[-inf, 508)2526418210816.55%7.27%17.19%6.53%
[508, 519)3495437305812.50%10.06%35.16%15.99%
[519, 537)3031333269810.99%8.73%48.85%24.35%
[537, 556)352029232288.30%10.14%60.86%34.34%
[556, 567)362323633876.51%10.43%70.56%44.83%
[567, 585)418325139326.00%12.04%80.88%57.00%
[585, 602)304512529204.11%8.77%86.02%66.04%
[602, 618)462119144304.13%13.31%93.87%79.76%
[618, 645)28768327932.89%8.28%97.29%88.41%
[645, inf)38106637441.73%10.97%100.00%100.00%
2026-11-03T14:31:25.319281 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-11-03T14:31:25.446687 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

跨时间验证样本结论

风险高度集中在低分尾端——最低 20% 样本集中了全量 48.9% 的坏客户,策略切分空间很大。

① 样本与分箱结构:共 34,730 样本 / 2,432 坏样本,整体坏率 7.00%,好坏比约 13.3:1;10 档等频分数分栏,每箱约 10%,分箱均匀,无空箱无过厚箱。

② 排序性与单调性:逐箱 Bad Rate 16.55% → … → 1.73%(低分 → 高分),存在局部倒挂,建议复核分箱边界,最低分段坏率是最高分段的 9.6 倍。

③ 区分度与坏客户捕获:最低 20% 样本捕获全量 48.9% 的坏客户,最低 50% 捕获 80.9%;最高 20% 样本仅含 6.1% 的坏客户,其中风险最高一箱单独占 18.0%。

(口径:分数 = 按基准分/优比/PDO 刻度由模型概率转换,分数越高风险越低;结论由程序按分箱数据自动生成。)

注:score_cut 表按开发样本(train)分数等频 10 档取边界(低分 → 高分),各样本共用同一套边界、同栏可比(20261022 口径);分数由模型概率按基准分/优比/PDO 刻度转换,分数越高风险越低;每表附 KS/AUC(按概率计算,与分栏口径无关);训练+测试 = TRAIN 全体。旧实验无分数转换时为 PD 等频口径(按本样本分位切分)。

10 KS和GINI

LR + WOE 评分卡

样本KSGINIAUC
开发样本0.30980.41000.7050
测试样本0.31030.40840.7042
训练+测试样本0.30940.40940.7047
跨时间验证样本0.26860.35900.6795

开发样本

2026-11-03T14:31:25.535249 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-11-03T14:31:25.604111 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

测试样本

2026-11-03T14:31:25.672032 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-11-03T14:31:25.740226 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

训练+测试样本

2026-11-03T14:31:25.808389 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-11-03T14:31:25.877826 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

跨时间验证样本

2026-11-03T14:31:26.005670 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-11-03T14:31:26.075655 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

注:GINI = 2×AUC−1。KS 图 X 轴为分数(升序),红虚线为最大间隔位置;ROC 图例标注 AUC。

11 PSI

模型稳定性评估

模型训练集与测试集训练集与跨时间验证训练集与psi验证集
LR + WOE 评分卡0.00030.02540.0534

分数段 PSI 明细(开发样本 vs 跨时间验证样本)

分数段开发样本占比跨时间验证样本占比PSI 贡献累计 PSI
[-inf, 508)8.56%5.75%0.01120.0112
[508, 519)3.78%3.21%0.00090.0121
[519, 537)17.60%17.08%0.00020.0123
[537, 556)9.93%10.06%0.00000.0123
[556, 567)9.34%10.23%0.00080.0131
[567, 585)10.13%12.10%0.00350.0166
[585, 602)10.50%8.51%0.00420.0208
[602, 618)8.56%7.88%0.00060.0214
[618, 645)11.53%13.42%0.00290.0243
[645, inf)10.06%11.77%0.00270.0269

分数段 PSI 明细(开发样本 vs PSI时点样本)

分数段开发样本占比PSI时点占比PSI 贡献累计 PSI
[-inf, 508)8.56%6.16%0.00790.0079
[508, 519)3.78%3.85%0.00000.0079
[519, 537)17.60%16.05%0.00140.0093
[537, 556)9.93%10.27%0.00010.0094
[556, 567)9.34%10.02%0.00050.0099
[567, 585)10.13%14.34%0.01460.0246
[585, 602)10.50%8.12%0.00610.0307
[602, 618)8.56%8.01%0.00040.0311
[618, 645)11.53%14.12%0.00530.0363
[645, inf)10.06%9.08%0.00100.0373
2026-11-03T14:31:26.158552 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

注:分数段 PSI 明细按分数刻度计算(概率→分数转换,20261022 口径),分栏边界取自开发样本(train)分数等频分位,各对比样本共用同一套边界。

模型样本量分布评估

评分段开发样本样本量开发样本占比测试样本样本量测试样本占比训练+测试样本样本量训练+测试样本占比跨时间验证样本样本量跨时间验证样本占比PSI时点样本样本量PSI时点样本占比
(-inf, 508]32598.56%21298.38%53888.49%19965.75%10766.16%
(508, 519]14413.78%9673.81%24083.79%11153.21%6723.85%
(519, 537]670517.60%445217.53%1115717.58%593217.08%280516.05%
(537, 556]37849.93%25159.91%62999.92%349510.06%179410.27%
(556, 567]35569.34%23779.36%59339.35%355310.23%175010.02%
(567, 585]385710.13%255810.07%641510.11%420312.10%250514.34%
(585, 602]399910.50%267210.52%667110.51%29548.51%14188.12%
(602, 618]32628.56%21498.46%54118.52%27367.88%13998.01%
(618, 645]439211.53%300411.83%739611.65%466013.42%246814.12%
(645, inf]383310.06%256810.11%640110.08%408611.77%15869.08%
sum38088100.00%25391100.00%63479100.00%34730100.00%17473100.00%

变量 PSI(开发样本 vs PSI时点样本,按入模变量分箱)

特征项PSI状态
setl_prcp_ratio0.2370关注
end_loan_to_cur_freq0.2135关注
t_ps_od_days_max0.2128关注
max_od_days0.1885关注
s_ps_od_days_max0.1764关注
f_ps_od_days_max0.1497关注
suc_apprv_num_m10.1385关注
apprv_amt_day0.1320关注
apprv_num_m10.1073关注
cre_now_l60_used_amt_m_min0.0710稳定
cre_now_l60_used_amt_d_min0.0572稳定
loan_os_prcp_credit_amount_m60.0360稳定
cre_now_l30_used_amt_d_min0.0333稳定
cre_l90_used_limit_amt_min0.0288稳定
cre_l60_used_limit_rate_min0.0284稳定
available_limit_amt0.0077稳定
f_ps_od_days_max — PSI 0.1497
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 0.5]53.63%71.67%18.04%0.0523
(0.5, 1.5]22.17%15.68%-6.49%0.0225
(1.5, inf]24.20%12.65%-11.55%0.0749
apprv_amt_day — PSI 0.1320
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 11]79.68%63.48%-16.21%0.0369
(11, inf]20.32%36.52%16.21%0.0951
suc_apprv_num_m1 — PSI 0.1385
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 0.5]79.23%62.50%-16.73%0.0397
(0.5, inf]20.77%37.50%16.73%0.0988
cre_now_l60_used_amt_m_min — PSI 0.0710
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 1.78]74.09%61.70%-12.38%0.0227
(1.78, inf]25.91%38.30%12.38%0.0484
end_loan_to_cur_freq — PSI 0.2135
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, -68.5]62.07%39.34%-22.73%0.1036
(-68.5, -25.5]14.68%21.11%6.43%0.0233
(-25.5, inf]23.25%39.55%16.30%0.0866
apprv_num_m1 — PSI 0.1073
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 0.5]73.93%58.74%-15.19%0.0349
(0.5, 1.5]14.05%24.06%10.01%0.0539
(1.5, inf]12.02%17.20%5.18%0.0185
s_ps_od_days_max — PSI 0.1764
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 0.5]54.12%74.07%19.95%0.0626
(0.5, inf]45.88%25.93%-19.95%0.1138
max_od_days — PSI 0.1885
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 7.5]88.53%98.19%9.66%0.0100
(7.5, inf]11.47%1.81%-9.66%0.1785
cre_l90_used_limit_amt_min — PSI 0.0288
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 748.71]35.98%38.67%2.70%0.0019
(748.71, 8516.23]52.99%45.70%-7.30%0.0108
(8516.23, inf]11.03%15.63%4.60%0.0160
cre_l60_used_limit_rate_min — PSI 0.0284
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 0.28]38.42%43.84%5.42%0.0071
(0.28, 0.92]49.66%41.38%-8.28%0.0151
(0.92, inf]11.92%14.78%2.86%0.0062
available_limit_amt — PSI 0.0077
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 0.03]32.52%32.39%-0.14%0.0000
(0.03, 5847.69]47.88%44.64%-3.24%0.0023
(5847.69, inf]19.59%22.97%3.38%0.0054
cre_now_l30_used_amt_d_min — PSI 0.0333
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 1]86.91%80.18%-6.73%0.0054
(1, inf]13.09%19.82%6.73%0.0279
setl_prcp_ratio — PSI 0.2370
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 0.04]21.92%44.50%22.58%0.1599
(0.04, inf]78.08%55.50%-22.58%0.0771
loan_os_prcp_credit_amount_m6 — PSI 0.0360
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 48.94]33.01%41.65%8.64%0.0201
(48.94, 107.59]49.22%40.80%-8.43%0.0158
(107.59, inf]17.76%17.55%-0.21%0.0000
t_ps_od_days_max — PSI 0.2128
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 0.5]54.18%75.91%21.72%0.0732
(0.5, inf]45.82%24.09%-21.72%0.1396
cre_now_l60_used_amt_d_min — PSI 0.0572
分栏开发基准比例PSI时点样本 比例变更贡献
(-inf, 1]81.74%71.70%-10.04%0.0132
(1, inf]18.26%28.30%10.04%0.0440

注:PSI ≤ 0.1 视为稳定,0.1~0.25 需关注,> 0.25 不稳定。

12 单变量分析

注:统计所有入模变量的单变量分析(开发样本口径)。预留验证/总体/OOT 口径的分组明细待扩展。

f_ps_od_days_max
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 0.5]2042253.63%18144227811.15%7.96
2(0.5, 1.5]844175.80%80024395.20%18.23
3(1.5, inf]9217100.00%87844334.70%20.29
合计380803493031508.27%11.09
2026-11-03T14:31:26.263212 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(11.15% → 4.70%),WOE 单调上升(-0.3309 → 0.6040),符合风控分箱预期。IV=0.1809,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.5])为坏率最高区间:坏率高达 11.15%,可作为拒绝或重点风控策略的阈值参考

apprv_amt_day
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 11]3034479.68%2745228929.53%9.49
2(11, inf]7736100.00%74782583.34%28.98
合计380803493031508.27%11.09
2026-11-03T14:31:26.368130 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.53% → 3.34%),WOE 单调上升(-0.1555 → 0.9608),符合风控分箱预期。IV=0.1475,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 11])为坏率最高区间:坏率高达 9.53%,可作为拒绝或重点风控策略的阈值参考

suc_apprv_num_m1
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 0.5]3016979.23%2728728829.55%9.47
2(0.5, inf]7911100.00%76432683.39%28.52
合计380803493031508.27%11.09
2026-11-03T14:31:26.470225 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.55% → 3.39%),WOE 单调上升(-0.1580 → 0.9446),符合风控分箱预期。IV=0.1475,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.5])为坏率最高区间:坏率高达 9.55%,可作为拒绝或重点风控策略的阈值参考

cre_now_l60_used_amt_m_min
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 1.78]2820174.09%2545827439.73%9.28
2(1.78, inf]9864100.00%94574074.13%23.24
合计380653491531508.28%11.08
2026-11-03T14:31:26.572536 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.73% → 4.13%),WOE 单调上升(-0.1775 → 0.7402),符合风控分箱预期。IV=0.1300,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 1.78])为坏率最高区间:坏率高达 9.73%,可作为拒绝或重点风控策略的阈值参考

end_loan_to_cur_freq
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, -68.5]2363562.07%21233240210.16%8.84
2(-68.5, -25.5]559276.75%52323606.44%14.53
3(-25.5, inf]8853100.00%84653884.38%21.82
合计380803493031508.27%11.09
2026-11-03T14:31:26.676457 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(10.16% → 4.38%),WOE 单调上升(-0.2267 → 0.6767),符合风控分箱预期。IV=0.1253,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 1 箱((-inf, -68.5])为坏率最高区间:坏率高达 10.16%,可作为拒绝或重点风控策略的阈值参考

apprv_num_m1
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 0.5]2815273.93%2542527279.69%9.32
2(0.5, 1.5]535087.98%51072434.54%21.02
3(1.5, inf]4578100.00%43981803.93%24.43
合计380803493031508.27%11.09
2026-11-03T14:31:26.783354 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.69% → 3.93%),WOE 单调上升(-0.1734 → 0.7900),符合风控分箱预期。IV=0.1224,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.5])为坏率最高区间:坏率高达 9.69%,可作为拒绝或重点风控策略的阈值参考

s_ps_od_days_max
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 0.5]2061054.12%18430218010.58%8.45
2(0.5, inf]17470100.00%165009705.55%17.01
合计380803493031508.27%11.09
2026-11-03T14:31:26.891592 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(10.58% → 5.55%),WOE 单调上升(-0.2713 → 0.4279),符合风控分箱预期。IV=0.1150,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.5])为坏率最高区间:坏率高达 10.58%,可作为拒绝或重点风控策略的阈值参考

max_od_days
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 7.5]3371288.53%3127624367.23%12.84
2(7.5, inf]4368100.00%365471416.35%5.12
合计380803493031508.27%11.09
2026-11-03T14:31:26.990485 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调上升(7.23% → 16.35%),WOE 单调下降(0.1466 → -0.7732),符合风控分箱预期。IV=0.1123,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 2 箱((7.5, inf])为坏率最高区间:坏率高达 16.35%,可作为拒绝或重点风控策略的阈值参考

cre_l90_used_limit_amt_min
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 748.71]1369435.98%129347605.55%17.02
2(748.71, 8516.23]2017288.97%1835418189.01%10.10
3(8516.23, inf]4199100.00%362757213.62%6.34
合计380653491531508.28%11.08
2026-11-03T14:31:27.098960 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调上升(5.55% → 13.62%),WOE 单调下降(0.4288 → -0.5585),符合风控分箱预期。IV=0.1036,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 3 箱((8516.23, inf])为坏率最高区间:坏率高达 13.62%,可作为拒绝或重点风控策略的阈值参考

cre_l60_used_limit_rate_min
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 0.28]1462638.42%138138135.56%16.99
2(0.28, 0.92]1890288.08%1715317499.25%9.81
3(0.92, inf]4537100.00%394958812.96%6.72
合计380653491531508.28%11.08
2026-11-03T14:31:27.279018 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调上升(5.56% → 12.96%),WOE 单调下降(0.4271 → -0.5010),符合风控分箱预期。IV=0.1034,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 3 箱((0.92, inf])为坏率最高区间:坏率高达 12.96%,可作为拒绝或重点风控策略的阈值参考

available_limit_amt
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 0.03]1238532.52%10918146711.85%7.44
2(0.03, 5847.69]1823480.41%1699512396.79%13.72
3(5847.69, inf]7461100.00%70174445.95%15.80
合计380803493031508.27%11.09
2026-11-03T14:31:27.400935 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(11.85% → 5.95%),WOE 单调上升(-0.3987 → 0.3543),符合风控分箱预期。IV=0.1021,该变量区分度尚可。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.03])为坏率最高区间:坏率高达 11.85%,可作为拒绝或重点风控策略的阈值参考

cre_now_l30_used_amt_d_min
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 1]3121186.91%2827329389.41%9.62
2(1, inf]4701100.00%45471543.28%29.53
合计359123282030928.61%10.61
2026-11-03T14:31:27.510328 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.41% → 3.28%),WOE 单调上升(-0.0980 → 1.0231),符合风控分箱预期。IV=0.0995,该变量区分度偏弱但可用。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 1])为坏率最高区间:坏率高达 9.41%,可作为拒绝或重点风控策略的阈值参考

setl_prcp_ratio
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 0.04]834621.92%79793674.40%21.74
2(0.04, inf]29734100.00%2695127839.36%9.68
合计380803493031508.27%11.09
2026-11-03T14:31:27.616899 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调上升(4.40% → 9.36%),WOE 单调下降(0.6733 → -0.1355),符合风控分箱预期。IV=0.0905,该变量区分度偏弱但可用。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 2 箱((0.04, inf])为坏率最高区间:坏率高达 9.36%,可作为拒绝或重点风控策略的阈值参考

loan_os_prcp_credit_amount_m6
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 48.94]1256433.01%118756895.48%17.24
2(48.94, 107.59]1873582.24%1707416618.87%10.28
3(107.59, inf]6761100.00%596279911.82%7.46
合计380603491131498.27%11.09
2026-11-03T14:31:27.734553 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调上升(5.48% → 11.82%),WOE 单调下降(0.4412 → -0.3959),符合风控分箱预期。IV=0.0893,该变量区分度偏弱但可用。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 2 箱 WOE<0(风险高于平均)

第 3 箱((107.59, inf])为坏率最高区间:坏率高达 11.82%,可作为拒绝或重点风控策略的阈值参考

t_ps_od_days_max
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 0.5]2063354.18%18517211610.26%8.75
2(0.5, inf]17447100.00%1641310345.93%15.87
合计380803493031508.27%11.09
2026-11-03T14:31:27.839973 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(10.26% → 5.93%),WOE 单调上升(-0.2368 → 0.3587),符合风控分箱预期。IV=0.0843,该变量区分度偏弱但可用。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 0.5])为坏率最高区间:坏率高达 10.26%,可作为拒绝或重点风控策略的阈值参考

cre_now_l60_used_amt_d_min
组号ValueTot. Records累计占比#Goods#BadsBad RateG/B Odds
1(-inf, 1]2794381.74%2520527389.80%9.21
2(1, inf]6242100.00%59502924.68%20.38
合计341853115530308.86%10.28
2026-11-03T14:31:27.945307 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

关键结论

单调性良好:随分栏增大,Bad Rate 单调下降(9.80% → 4.68%),WOE 单调上升(-0.1106 → 0.6840),符合风控分箱预期。IV=0.0752,该变量区分度偏弱但可用。

WOE 在第 1、2 箱之间穿越 0 线:前 1 箱 WOE>0(客户质量优于整体),后 1 箱 WOE<0(风险高于平均)

第 1 箱((-inf, 1])为坏率最高区间:坏率高达 9.80%,可作为拒绝或重点风控策略的阈值参考

13 压力测试

注:验证压力情境下模型的有效性和区分能力稳定性:a.坏浓度更高样本 b.时间外样本 c.特定特征客群 d.极端时间样本。当前流程自动生成情景1(贡献度最高的变量调用失败,20261026 批注2)与情景2(客群资质变差——跨时间样本剔除一半好样本,批注16)。

测试情景1:贡献度最高的变量调用失败

变量贡献度

变量贡献度
max_od_days0.3346
f_ps_od_days_max0.3337
cre_l90_used_limit_amt_min0.2341
s_ps_od_days_max0.2166
apprv_amt_day0.1923
end_loan_to_cur_freq0.1516
cre_l60_used_limit_rate_min0.1320
失效变量贡献度口径值KS(前 → 后)AUC(前 → 后)KS 衰减AUC 衰减
max_od_days0.33460.2686 → 0.27000.6795 → 0.68230.00140.0028

注:失效处理口径:LR 将该变量 WOE 置 0(中性箱)后按系数重算概率;树模型将该特征置缺失重打分;评估样本 = 跨时间验证样本。衰减幅度反映模型对头部变量的依赖程度。

区分度验证 — LR + WOE 评分卡 @ 跨时间样本

KS = 0.2686,GINI = 0.3590

2026-11-03T14:31:28.071486 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-11-03T14:31:28.198489 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

测试情景2:客群资质变差

指标数值
样本量34730 → 18581(剔除好样本 16149)
KS(前 → 后)0.2686 → 0.2706(Δ 0.0020)
AUC(前 → 后)0.6795 → 0.6800(Δ 0.0005)
GINI(后)0.3600
2026-11-03T14:31:28.324241 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-11-03T14:31:28.450804 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

注:情景2 模拟客群资质变差的压力情形(好客户变少、坏客户占比升高):随机剔除一半好样本(固定 seed=42,可复现)后复算 KS/AUC。