Customer Potential Value for US Consumer (Card, Retail and Multiple Relationship Customer) Objective: Build Customer centric approach by predicting Potential Value for efficient marketing strategy Identified different segments based on bureau behaviour to predict segment based potential value Identified 8 key segments for marketing strategy based on Citi value and potential value for each customer Overlay the Potential value with customer attrition model to devise more… Show more Customer Potential Value for US Consumer (Card, Retail and Multiple Relationship Customer) Objective: Build Customer centric approach by predicting Potential Value for efficient marketing strategy Identified different segments based on bureau behaviour to predict segment based potential value Identified 8 key segments for marketing strategy based on Citi value and potential value for each customer Overlay the Potential value with customer attrition model to devise more efficient Targeting strategy. Show less

Trade Finance Need Prediction for Institution client Group (ICG) of Citi Objective: Worked on Trade Finance need prediction for Institution client Group (ICG), so that we can optimally allocate the funds. Built Rule based Engine (an expert system which incorporate Business knowledge) Bulit XGBoost model Built the following Bayesian model: 1) Bayesian assuming variables are independent (assuming continuous distribution) 2) Bayesian Considering… Show more Trade Finance Need Prediction for Institution client Group (ICG) of Citi Objective: Worked on Trade Finance need prediction for Institution client Group (ICG), so that we can optimally allocate the funds. Built Rule based Engine (an expert system which incorporate Business knowledge) Bulit XGBoost model Built the following Bayesian model: 1) Bayesian assuming variables are independent (assuming continuous distribution) 2) Bayesian Considering dependencies between the variables (assuming continuous distribution) 3) Bayesian assuming variables are independent (assuming discrete distribution) 4) Bayesian Considering dependencies between the variables (assuming discrete distribution)Customer Attrition Model for US Consumer bank (Data Science Hackathon) Objective: Predict the likelihood of a Citi Gold customer to voluntarily attrite in next 3 months Built the benchmark XGBoost Model with 4.8 lift in response rate in top decile Experimented Anomaly detection Techniques (Autoencoder, Isolation Forest), Deep Neural Network, Oversampling (Random, Smote) and Undersampling Technique (Tomek Link, Edited nearest Neighbour) Finally, selected oversampling followed by Undersampling (SMOTE-ENN) to handle the class imbalance  Built XGBoost on Data modified using SMOTE-ENN and achieved lift of 5.8 in top decile Commended by Senior leadership for ease of implementation while also following Citi modelling guidelines. Show less

 Worked on a Contact Strategy Optimization for Credit card customers in North America.  Developed following Reinforcement learning algorithms in python for maximizing long term rewards by making more meaningful contacts. 1) Q Learning 2) Fitted Q Learning 3) Deep Q Learning Used a Keras library in python for building an neural network and Bellman Equation (Markov decision Process ) for calculating long term reward.

 Developed a segmented (for different delinquency segment) model to identify the Customer who are more likely to default in next month for Home Loans to enable collection team to prioritize efforts  Worked on Collection Scorecard model for prioritizing collection effort for Home Loans Statistical Techniques used – XGBoost, Decision Tree and Logistic Regression

Remote Internship in Data Science and Analytics