Loan-Delequency-Prediction

Dataset Description:

It is a India ML hiring hackathon dataset hosted by analytics vidya which can be seen in the below given link

https://datahack.analyticsvidhya.com/contest/india-ml-hiring-hackathon-2019/

Problem statement:

1. To predict the delequency in the 13th month of each loan which is missing in the test data
2. Predict as delequent or non-delequent (1, 0)

Preprocessing and Feature Engineering:

Step:1

1. Converting all categorical columns to numericals :(source, financial_institution,loan_purpose)
2. Converting dates to Date_Time format and extracting
   • Month
   • Week
   • Day of th month
   • Day of the week

Step-2:

1. Finding the feature importance in normal Machine learning algorithm (Tree Classifier)
2. Finding the correlated features

• Feature importance - Gives what are the features in the data influence the label (here, 1- delequent, 0- non-delequent)
• Correlated features - The features that are correlated to the label are important in classification algorithm (but not in regression algorithm)

As we are in need of finding the delequency of a particual loan , it is a classification problem

Correlation plot of preprocessed data

Correlation plot of only selected features

Step:3

1. Extract the following features from the selected fatures:[m0 to m12]

   • Max delequent of previous months

   • Min delequent of previous months

   • Max min diffetence of previous months - gives how the delequency fluus months

   • Sum of delequent of previous monthsctuates

   • Mean delequent of previous months

   • Variance delequent of previous months

   • Number of delequent months delequent of previo

   • Median of delequent of previous months

   • Skewness of previous months - gives whether the delequency is more in the later or earlier months

   • Kurtosis of previous months - tailedness - i.e, how the delequency is in recent months

   • Trend of the previous month delequency

Corelation of final train data

Train data - 69634 data with 23 features each Validation data - 46424 data with 23 features each Test data - 35866 data with 23 features each

Building the Model- GRU

GRU - Gated Recurrent Unit

• Why GRU?
  • The GRU has an advantage of storing the previous sequnce
  • Overcomes vanishing gradient problem - i.e., The backpropagation will give importance to initial data in sequence as well
  • Works as efficient as LSTM also with less parameters

Model:

• The model has a GRU with 32 cells, Dense of 10, and then final Dense for 2 classes
• The model is trained with class weights of ratio 1:7 as the data is imbalanced towards the class 1

Performance analysis of valid data

Confusion_matrix:

 [[46020   150]
 [  109   145]]

Accuracy: 0.9944209891435465

F1-score: 0.528233151183971

Classification Report:
               precision    recall  f1-score   support

           0       1.00      1.00      1.00     46170
           1       0.49      0.57      0.53       254

   micro avg       0.99      0.99      0.99     46424
   macro avg       0.74      0.78      0.76     46424
weighted avg       0.99      0.99      0.99     46424

The result was published in the private leader board of the hackathon competition

Public board -244/1232 Private board - 149/712

The above ranking is among 3740 participants