（十一）隨機森林在催收評分卡(還款率模型)的應用

（十一）隨機森林在催收評分卡(還款率模型)的應用

催收評分卡和申請評分卡和行爲評分卡不太一樣，一般申請評分卡和行爲評分卡使用一個模型就可以了，但是催收評分卡由三個模型構成：（不同的模型功能目的不一樣，其中失聯預測模型是比較重要的）

• 還款率模型
• 賬齡滾動模型
• 失聯預測模型

這篇博客以還款率模型進行講解，要講解還款率模型，我們首先要了解一下隨機森林模型

基於迴歸樹的隨機森林（元分類器是由許多回歸樹構成，每一個元分類器模型並行運行得出一個預測值，取所有元分類器模型的平均值作爲最終的預測值）

 

隨機森林模型的訓練步驟

 

如何建立還款率模型呢？

 

（小象學院）代碼如下，數據可以在我的資源下載，當然了，還款率模型完之後還可以對其進行延伸，預測出來的催回還款率假設定一個閾值（80%，自己可以定），大於80%爲可摧回，小於爲不可催回，之後可以使用二分類的邏輯迴歸對客戶情況進行預測該客戶是可摧回還是不可催回：

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import GridSearchCV

def MakeupMissingCategorical(x):
    if str(x) == 'nan':
        return 'Unknown'
    else:
        return x

def MakeupMissingNumerical(x,replacement):
    if np.isnan(x):
        return replacement
    else:
        return x


'''
第一步：文件準備
'''
foldOfData = 'H:/'
mydata = pd.read_csv(foldOfData + "還款率模型.csv",header = 0,engine ='python')
#催收還款率等於催收金額/（所欠本息+催收費用）。其中催收費用以支出形式表示
mydata['rec_rate'] = mydata.apply(lambda x: x.LP_NonPrincipalRecoverypayments /(x.AmountDelinquent-x.LP_CollectionFees), axis=1)
#還款率假如大於1，按作1處理
mydata['rec_rate'] = mydata['rec_rate'].map(lambda x: min(x,1))
#整個開發數據分爲訓練集、測試集2個部分
trainData, testData = train_test_split(mydata,test_size=0.4)

'''
第二步：數據預處理
'''
#由於不存在數據字典，所以只分類了一些數據
categoricalFeatures = ['CreditGrade','Term','BorrowerState','Occupation','EmploymentStatus','IsBorrowerHomeowner','CurrentlyInGroup','IncomeVerifiable']

numFeatures = ['BorrowerAPR','BorrowerRate','LenderYield','ProsperRating (numeric)','ProsperScore','ListingCategory (numeric)','EmploymentStatusDuration','CurrentCreditLines',
                'OpenCreditLines','TotalCreditLinespast7years','CreditScoreRangeLower','OpenRevolvingAccounts','OpenRevolvingMonthlyPayment','InquiriesLast6Months','TotalInquiries',
               'CurrentDelinquencies','DelinquenciesLast7Years','PublicRecordsLast10Years','PublicRecordsLast12Months','BankcardUtilization','TradesNeverDelinquent (percentage)',
               'TradesOpenedLast6Months','DebtToIncomeRatio','LoanFirstDefaultedCycleNumber','LoanMonthsSinceOrigination','PercentFunded','Recommendations','InvestmentFromFriendsCount',
               'Investors']

'''
類別型變量需要用目標變量的均值進行編碼
'''
encodedFeatures = []
encodedDict = {}
for var in categoricalFeatures:
    trainData[var] = trainData[var].map(MakeupMissingCategorical)
    avgTarget = trainData.groupby([var])['rec_rate'].mean()
    avgTarget = avgTarget.to_dict()
    newVar = var + '_encoded'
    trainData[newVar] = trainData[var].map(avgTarget)
    encodedFeatures.append(newVar)
    encodedDict[var] = avgTarget

#對數值型數據的缺失進行補缺
trainData['ProsperRating (numeric)'] = trainData['ProsperRating (numeric)'].map(lambda x: MakeupMissingNumerical(x,0))
trainData['ProsperScore'] = trainData['ProsperScore'].map(lambda x: MakeupMissingNumerical(x,0))

avgDebtToIncomeRatio = np.mean(trainData['DebtToIncomeRatio'])
trainData['DebtToIncomeRatio'] = trainData['DebtToIncomeRatio'].map(lambda x: MakeupMissingNumerical(x,avgDebtToIncomeRatio))
numFeatures2 = numFeatures + encodedFeatures

'''
第三步：調參
對基於CART的隨機森林的調參，主要有：
1，樹的個數
2，樹的最大深度
3，內部節點最少樣本數與葉節點最少樣本數
4，特徵個數
此外，調參過程中選擇的誤差函數是均值誤差，5倍摺疊
'''
X, y= trainData[numFeatures2],trainData['rec_rate']

param_test1 = {'n_estimators':range(60,91,5)}
gsearch1 = GridSearchCV(estimator = RandomForestRegressor(min_samples_split=50,min_samples_leaf=10,max_depth=8,max_features='sqrt' ,random_state=10),param_grid = param_test1, scoring='neg_mean_squared_error',cv=5)
gsearch1.fit(X,y)
gsearch1.best_params_, gsearch1.best_score_
best_n_estimators = gsearch1.best_params_['n_estimators']

param_test2 = {'max_depth':range(3,15), 'min_samples_split':range(10,101,10)}
gsearch2 = GridSearchCV(estimator = RandomForestRegressor(n_estimators=best_n_estimators, min_samples_leaf=10,max_features='sqrt' ,random_state=10,oob_score=True),param_grid = param_test2, scoring='neg_mean_squared_error',cv=5)
gsearch2.fit(X,y)
gsearch2.best_params_, gsearch2.best_score_
best_max_depth = gsearch2.best_params_['max_depth']
best_min_samples_split = gsearch2.best_params_['min_samples_split']

param_test3 = {'min_samples_leaf':range(1,20,2)}
gsearch3 = GridSearchCV(estimator = RandomForestRegressor(n_estimators=best_n_estimators, max_depth = best_max_depth,max_features='sqrt',min_samples_split=best_min_samples_split,random_state=10,oob_score=True),param_grid = param_test3, scoring='neg_mean_squared_error',cv=5)
gsearch3.fit(X,y)
gsearch3.best_params_, gsearch3.best_score_
best_min_samples_leaf = gsearch3.best_params_['min_samples_leaf']

numOfFeatures = len(numFeatures2)
mostSelectedFeatures = numOfFeatures/2
param_test4 = {'max_features':range(3,numOfFeatures+1)}
gsearch4 = GridSearchCV(estimator = RandomForestRegressor(n_estimators=best_n_estimators, max_depth=best_max_depth,min_samples_leaf=best_min_samples_leaf,min_samples_split=best_min_samples_split,random_state=10,oob_score=True),param_grid = param_test4, scoring='neg_mean_squared_error',cv=5)
gsearch4.fit(X,y)
gsearch4.best_params_, gsearch4.best_score_
best_max_features = gsearch4.best_params_['max_features']

#把最優參數全部獲取去做隨機森林擬合
cls = RandomForestRegressor(n_estimators=best_n_estimators,max_depth=best_max_depth,min_samples_leaf=best_min_samples_leaf,min_samples_split=best_min_samples_split,max_features=best_max_features,random_state=10,oob_score=True)
cls.fit(X,y)
trainData['pred'] = cls.predict(trainData[numFeatures2])
trainData['less_rr'] = trainData.apply(lambda x: int(x.pred > x.rec_rate), axis=1)
np.mean(trainData['less_rr'])
err = trainData.apply(lambda x: np.abs(x.pred - x.rec_rate), axis=1)
np.mean(err)

#隨機森林評估變量重要性
importance=cls.feature_importances_
featureImportance=dict(zip(numFeatures2,importance))
featureImportance=sorted(featureImportance.items(),key=lambda x:x[1],reverse=True)

'''
第四步：在測試集上測試效果
'''
#類別型數據處理
for var in categoricalFeatures:
    testData[var] = testData[var].map(MakeupMissingCategorical)
    newVar = var + '_encoded'
    testData[newVar] = testData[var].map(encodedDict[var])
    avgnewVar = np.mean(trainData[newVar])
    testData[newVar] = testData[newVar].map(lambda x: MakeupMissingNumerical(x, avgnewVar))

#連續性數據處理
testData['ProsperRating (numeric)'] = testData['ProsperRating (numeric)'].map(lambda x: MakeupMissingNumerical(x,0))
testData['ProsperScore'] = testData['ProsperScore'].map(lambda x: MakeupMissingNumerical(x,0))
testData['DebtToIncomeRatio'] = testData['DebtToIncomeRatio'].map(lambda x: MakeupMissingNumerical(x,avgDebtToIncomeRatio))

testData['pred'] = cls.predict(testData[numFeatures2])
testData['less_rr'] = testData.apply(lambda x: int(x.pred > x.rec_rate), axis=1)
np.mean(testData['less_rr'])
err = testData.apply(lambda x: np.abs(x.pred - x.rec_rate), axis=1)
np.mean(err)