├── Financial fraud prediction.ipynb
├── README.md
└── result1.PNG
/README.md:
--------------------------------------------------------------------------------
1 | # Financial-fraud-prediction
2 | 1、导入数据，并查看数据信息
3 | ----------
4 | ```
5 | # 导入数据文件
6 | df = pd.read_csv('F:\Python\data\LoanStats_2016Q3.csv', skiprows=1, low_memory=False)
7 | df.info() # 查看文件属性
8 | ```
9 | 2、对数据进行预处理，将相关度不大的属性列删除
10 | ----------
11 | ```
12 | # drop id and member_id
13 | df.drop('id',1,inplace=True)
14 | df.drop('member_id',1,inplace=True)
15 |
16 | # drop nan rows
17 | df.dropna(axis=0,how='all',inplace=True)
18 | df.info()
19 |
20 | # emp_title is too many to ignore
21 | df.drop('emp_title',1,inplace=True)
22 | ```
23 |
24 | 3、查看数据类别，将缺失值较大的属性列删除
25 | ----------
26 | ```
27 | # drop hige missing_pct
28 | df.drop('desc',1,inplace=True)
29 | df.drop('verification_status_joint',1,inplace=True)
30 |
31 | # drop unnecessary object
32 | df.drop('term',1,inplace=True)
33 | df.drop('issue_d',1,inplace=True)
34 | df.drop('purpose',1,inplace=True)
35 | df.drop('title',1,inplace=True)
36 | df.drop('zip_code',1,inplace=True)
37 | df.drop('addr_state',1,inplace=True)
38 | df.drop('earliest_cr_line',1,inplace=True)
39 | df.drop('revol_util',1,inplace=True)
40 |
41 | # drop after loan
42 | df.drop(['out_prncp','out_prncp_inv','total_pymnt','total_pymnt_inv','total_rec_prncp', 'grade', 'sub_grade'] ,1, inplace=True)
43 | df.drop(['total_rec_int','total_rec_late_fee','recoveries','collection_recovery_fee','collection_recovery_fee'],1, inplace=True)
44 | df.drop(['last_pymnt_d','last_pymnt_amnt','next_pymnt_d','last_credit_pull_d'],1, inplace=True)
45 | df.drop(['policy_code'],1, inplace=True)
46 | ```
47 |
48 | 4.整理数据列，关注loan_status属性列，并计算与该属性列相关系数，相关系数高的可以删除
49 | ------------
50 | ```
51 | # focus on loan_status
52 | df.loan_status.value_counts()
53 |
54 | # Current and Fully Paid regard as excellent 1, late regard as bad loan 0, others regard ass nan
55 | df.loan_status.replace('Current',int(1),inplace=True)
56 | df.loan_status.replace('Fully Paid',int(1),inplace=True)
57 | df.loan_status.replace('Charged Off',np.nan,inplace=True)
58 | df.loan_status.replace('In Grace Period',np.nan,inplace=True)
59 | df.loan_status.replace('Default',np.nan,inplace=True)
60 | df.loan_status.replace('Late (31-120 days)',int(0),inplace=True)
61 | df.loan_status.replace('Late (16-30 days)',int(0),inplace=True)
62 | ```
63 | ```
64 | # find highly corr Data
65 | cor = df.corr()
66 | cor.loc[:,:] = np.tril(cor,k=-1)
67 | cor = cor.stack()
68 | cor[abs(cor) > 0.5]
69 |
70 | # drop the columns highly correlated with loan_status
71 | df.drop(['funded_amnt','funded_amnt_inv','installment'],1,inplace=True)
72 | ```
73 | 5.对已处理好的数据进行建模训练，测试集为0.2
74 | ----------
75 | ```
76 | Y = df.loan_status
77 | df.drop('loan_status',1,inplace=True)
78 | X = df
79 |
80 | # train_data and test_data
81 | x_train, x_test, y_train, y_test = train_test_split(X,Y,test_size=0.2,random_state=1)
82 | ```
83 | 6.使用GBRT对数据进行建模训练
84 | ---------
85 | ```
86 | from sklearn.model_selection import train_test_split
87 | from sklearn.model_selection import GridSearchCV
88 | from sklearn import ensemble
89 | from sklearn.preprocessing import OneHotEncoder
90 | ```
91 | ```
92 | param_grid = {'learning_rate': [0.1],
93 | 'max_depth': [2],
94 | 'min_samples_split': [50,100],
95 | 'n_estimators': [100,200]
96 | }
97 |
98 |
99 | est = GridSearchCV(ensemble.GradientBoostingRegressor(),
100 | param_grid, n_jobs=4, refit=True)
101 |
102 | est.fit(x_train, y_train)
103 |
104 | best_params = est.best_params_
105 | print(best_params)
106 | ```
107 | ```
108 | %%time
109 | est = ensemble.GradientBoostingRegressor(min_samples_split=50,n_estimators=300,
110 | learning_rate=0.1,max_depth=1,random_state=0,
111 | loss='ls').fit(x_train,y_train)
112 | est.score(x_test,y_test)
113 | ```
114 | 7、测试模型，找出10个最重要属性
115 | -------
116 | ```
117 | def computer_ks(data):
118 | sorted_list = data.sort_values(['predict'],ascending=True)
119 | total_bad = sorted_list['label'].sum(axis=None,skipna=None,level=None,numeric_only=None)/3
120 | total_good = sorted_list.shape[0] - total_bad
121 |
122 | max_ks = 0.0
123 | good_count = 0.0
124 | bad_count = 0.0
125 | for index, row in sorted_list.iterrows():
126 | if row['label'] == 3:
127 | bad_count += 1.0
128 | else:
129 | good_count += 1.0
130 |
131 | val = bad_count/total_bad - good_count/total_good
132 | max_ks = max(max_ks,val)
133 |
134 | return max_ks
135 | ```
136 |
137 | ```
138 | feature_importance = est.feature_importances_
139 | feature_importance = 100.0 * (feature_importance / feature_importance.max())
140 |
141 | indices = np.argsort(feature_importance)[-10:]
142 | plt.barh(np.arange(10),feature_importance[indices],color='green',alpha=0.4)
143 | plt.yticks(np.arange(10+0.25),np.array(X.columns)[indices])
144 | _ = plt.xlabel('Relative importance'), plt.title('Top Ten Important Variables')
145 | ```
146 | 8、最终结果
147 | -------
148 | 
149 |
150 |
--------------------------------------------------------------------------------
/result1.PNG:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/fanzhihai/Financial-fraud-prediction/2a281db6059ec598711936490afcdf705e994ebf/result1.PNG
--------------------------------------------------------------------------------