信用評分卡模型在Python中實踐（下）
python金融風(fēng)控評分卡模型和數(shù)據(jù)分析微專業(yè)課：
http://dwz.date/b9vv
1項目背景介紹
? ? 信用評分卡的模型一般采用邏輯回歸模型，屬于二分類模型，Python 中的sklearn.linear_model導(dǎo)入LogisticRegression即可。
model_data?=?data[np.append(quant_model_vars,qual_model_vars)]#model_data_WOE?=?pd.DataFrame()model_data_WOE['duration']=duration_WoEmodel_data_WOE['amount']=amount_WoEmodel_data_WOE['age']=age_WoEmodel_data_WOE['installment_rate']=installment_rate_WoEmodel_data_WOE['status']=status_WoEmodel_data_WOE['credit_history']=credit_history_WoEmodel_data_WOE['savings']=savings_WoEmodel_data_WOE['property']=property_WoEmodel_data_WOE['employment_duration']=employment_duration_WoEmodel_data_WOE['purpose']=purpose_WoE#model_data_WOE['credit_risk']=credit_risk#邏輯回歸model?=?LogisticRegression()model.fit(model_data_WOE,credit_risk)coefficients?=?model.coef_.ravel()intercept?=?model.intercept_[0]
注：Python中的模型不夠R中模型友好，想看模型的變量、系數(shù)、檢驗之類的都比較麻煩，要一個變量一個變量去找，然后輸出打印，反之R的模型結(jié)果就友好很多了，一個summary函數(shù)就把全部概況顯示出來了。
###########自定義ks函數(shù)#############def?predict_df(model,data,label,feature=None):????if?feature:????????df_feature=data.loc[:,feature]????else:????????all_feature?=?list(data.columns.values)????????all_feature.remove(label)????????df_feature=data.loc[:,all_feature]????df_prob=model.predict(df_feature)????df_pred=pd.Series(df_prob).map(lambda?x:1?if?x>0.5?else?0)????df=pd.DataFrame()????df['predict']=df_pred????df['label']=data.loc[:,label].values????df['score']=df_prob????return?df??def?ks(data,model,label):????data_df?=?predict_df(model,data,label)????KS_data?=?data_df.sort_values(by='score',ascending=True)????KS_data['Bad']?=?KS_data['label'].cumsum()?/?KS_data['label'].sum()????KS_data['Count']?=?np.arange(1?,?len(KS_data['label'])?+?1)????KS_data['Good']?=?(KS_data['Count']?-?KS_data['label'].cumsum() )?/?(len(KS_data['label'])?-?KS_data['label'].sum())????KS_data.index=KS_data['Count']?????????ks?=?KS_data.iloc[::int(len(KS_data)/100),:]????ks.index?=?np.arange(len(ks))????return?ks??def?ks_plot(ks_df):????plt.figure(figsize=(6,?5))????plt.subplot(111)????plt.plot(ks_df['Bad'], lw=3.5, color='r', label='Bad')??# train_ks['Bad']????plt.plot(ks_df['Good'], lw=3.5, color='g',?????????????label='Good')??# train_ks['Good']????plt.legend(loc=4)????plt.grid(True)????plt.axis('tight')????plt.title('The KS Curve of data')????plt.show()
KS(Kolmogorov-Smirnov)：KS用于模型風(fēng)險區(qū)分能力進(jìn)行評估，?
指標(biāo)衡量的是好壞樣本累計分部之間的差值。?好壞樣本累計差異越大，KS指標(biāo)越大，那么模型的風(fēng)險區(qū)分能力越強(qiáng)，通常來講，KS>0.2即表示模型有較好的預(yù)測準(zhǔn)確性。經(jīng)過計算，模型的KS值為0.35，模型效果較好，如下：
?
?
?
六、評分卡
引用文獻(xiàn)的評分卡計算方法：
一般評分卡公式：Score=A - B * log(Odds)
通常情況下，需要設(shè)定兩個假設(shè)：?
（1）給某個特定的比率設(shè)定特定的預(yù)期分值；?
（2）確定比率翻番的分?jǐn)?shù)（PDO）?
根據(jù)以上的分析，我們首先假設(shè)比率為x的特定點的分值為P。則比率為2x的點的分值應(yīng)該為P+PDO。代入式中，可以得到如下兩個等式：?
P = A - B * log(x)?
P - PDO = A - B * log(2x)
本文中通過指定特定比率（好壞比）（1/20）的特定分值（50）和比率翻番的分?jǐn)?shù)（10），來計算評分卡的系數(shù)alpha和beta
def?alpha_beta(basepoints,baseodds,pdo):????beta?=?pdo/math.log(2)????alpha?=?basepoints?+?beta?*?math.log(baseodds)????return?alpha,beta
評分卡公式：Score=6.78?- 14.43 * log(Odds)
而 log(Odds) = \beta _{0} + \beta _{1} x _{1}+ \beta _{2} x _{2} + ... +\beta _{n } x _{n}
，代入WOE轉(zhuǎn)換后的變量并進(jìn)行變化，可得到最終的評分卡公式：
式中ωijωij?為第i行第j個變量的WOE，為已知變量；βiβi為邏輯回歸方程中的系數(shù)，為已知變量；δijδij為二元變量，表示變量i是否取第j個值。
根據(jù)以上表格可計算出指標(biāo)各分段的分值
#計算基礎(chǔ)分值basepoint?=?round(alpha?-?beta?*?intercept)#變量_scoreduration_score?=?np.round(model_data_WOE['duration']*coefficients[0]*beta)amount_score?=?np.round(model_data_WOE['amount']*coefficients[1]*beta)age_score?=?np.round(model_data_WOE['age']*coefficients[2]*beta)installment_rate_score?=?np.round(model_data_WOE['installment_rate']*coefficients[2]*beta)status_score?=?np.round(model_data_WOE['status']*coefficients[4]*beta)credit_history_score?=?np.round(model_data_WOE['credit_history']*coefficients[5]*beta)savings_score?=?np.round(model_data_WOE['savings']*coefficients[6]*beta)property_score?=?np.round(model_data_WOE['property']*coefficients[7]*beta)employment_duration_score?=?np.round(model_data_WOE['employment_duration']*coefficients[8]*beta)purpose_score?=?np.round(model_data_WOE['purpose']*coefficients[9]*beta)#變量的分值duration_scoreCard?=?pd.DataFrame(duration_Cutpoint,duration_score).drop_duplicates()amount_scoreCard?=?pd.DataFrame(amount_Cutpoint,amount_score).drop_duplicates()age_scoreCard?=?pd.DataFrame(age_Cutpoint,age_score).drop_duplicates()installment_rate_scoreCard?=?pd.DataFrame(installment_rate_Cutpoint,installment_rate_score).drop_duplicates()status_scoreCard?=?pd.DataFrame(np.array(discrete_data['status']),status_score).drop_duplicates()credit_history_scoreCard?=?pd.DataFrame(np.array(discrete_data['credit_history']),credit_history_score).drop_duplicates()savings_scoreCard?=?pd.DataFrame(np.array(discrete_data['savings']),savings_score).drop_duplicates()property_scoreCard?=?pd.DataFrame(np.array(discrete_data['property']),property_score).drop_duplicates()employment_duration_scoreCard?=?pd.DataFrame(np.array(discrete_data['employment_duration']),employment_duration_score).drop_duplicates()purpose_scoreCard?=?pd.DataFrame(np.array(discrete_data['purpose']),purpose_score).drop_duplicates()
至此，信用評分卡的
轉(zhuǎn)載https://blog.csdn.net/kxiaozhuk/article/details/84613794
python信用評分卡（附代碼，博主錄制）
http://dwz.date/b62p
