Sayantan Chakraborty

Call Volume Forecasting for Retail Bank, Branded Cards & Costco

Objective- Replaced a naive judgemental based approach with a robust multivariate segment level modeling framework to predict inbound call volume month over month, optimizing budget allocation for workforce management (WFM)

Roles & Responsibilities

1. Segmented call volume data based on customer attributes like month on book (MOB), Products, Interactive voice response(IVR)
2. Developed a time series predictive model SARIMAX, Ensemble methods( Random Forest, Boosting techniques) for multiple answer groups.
3. Socialized with workforce management(WFM) team and helped reduce overstaffing costs by bringing down error rate from 20-25% to 8-12%
4. Month over month Variance walk to quantify increase/ decrease in call volume based on exogenous drivers used for staffing allocation.

Benefits- Improvement in workforce optimization yielded savings of $10-$12 M annually.

Predicting customers entering into early delinquency bucket in retail bank collections portfolio

Objective- To predict whether customer enters into 1st delinquency bucket in next month to efficiently strategize outbound calls and reduce operational cost.

Role & Responsibilities

1. Analyze data from multiple tables to derive synthetic predictor variables PTP_last_3M, DPD_3M, no of outbound calls, payment indicator, revolver, product type at customer id level.
2. Predictive classification model using a combination of logistic regression, Bagging, Boosting technique.
3. Evaluation of metrics like Precision, F1 score to test efficacy of model performance.

Benefits- The model correctly predicted 75% of the accounts entering into target delinquency bucket which in turn reduced outbound calls by 22% quarterly.

Automation of headcount reconciliation of Capacity Planning

Objective- To automate cap plan x region wise headcount KPI's for all cap plans within a line of Business (LOB) from a monthly generated raw excel file.

Roles & Responsibilities- Prepared an automated python script to transform unstructured data in an excel and convert it to a summary level view used for monthly decision making of the entire capacity planning process.

Benefits- Reduced 48-50 hours manual process to a 2 hour job worth 0.5 FTE's.

Customer attrition in retail business

Objective- Analyze customer's behavior in B2B to understand the propensity of churn/attrition of individual distributors in order to derive retention strategies by means of bulk discount, promotional campaigns on different product categories as well as other granular levels of distributor attributes.

Roles & Responsibilities- Rank group of distributors based on their historical attributes of engagement with business. Segmentation using K-Means clustering to identify distinct groups followed by employing classification methodologies like Logistic Regression/Tree based algorithms to predict
probability of churn. The entire activity has been conducted for each distributors with varied metric(Recall,Precision,F1 Score) importance for each cluster having different distinguishable characteristics.

Benefits- 45% of the customers whose probability of churn was greater than 72% were successfully retained with incorporation of several strategies.

Supply chain & Sales Analytics

Objective-Vehicle Routing to minimize overall cost of delivery to various customers by employing optimum no of vehicles with fixed capacity under specific time window, traffic halt constraints as well as targeting a specific set of
priority customers in each routes imposed as a constraint to the problem. The model is deployed using Flask with user specific input web page followed by an
equivalent summary page to depict the output.

Roles & Responsibilities-Mathematical problem solved in Python using
pulP(Open source solver for Linear optimization).

Benefits-Overall reduction in service delivery cost by 12-14% in B2B.

Predictor model for silicon control in Blast Furnace

Objective-Prediction of silicon content in hot metal 2hours prior to casting through data driven modeling technique as well as recommendation of operating variable ranges to minimize silicon content in hot metal during
tapping of Pig iron.

Roles & Responsibilities- Exploratory Data Analysis to identify regimes of silicon content and provide operating range of controllable variables, predictive modeling employed with optimum no of features using the best algorithm pertaining to the data set in context which explains the variability of silicon content in hot metal reasonably well.

Benefits- Trials resulted in reduction of average silicon content from 0.1% to 0.64% which in turn impacted in 7kg/tons of reduction in lime consumption as well as 2.75kg/tons of hot metal reduction in fuel rate.

Tuyere Leak Detection System

Objective-To minimize daily no of false alarms (KPI< 5) and prescribe a methodology to correctly identify true leakages in tuyere and provide an early signal in order to take appropriate actions.

Roles & Responsibilities- Simulating no of alarms daily using moving average technique in time series data, calculating a synthetic variable to measure deviation in differential
flow, trial and errors with several first principle based algorithms, identifying set points to trigger alarms using statistical techniques, implementing logic in source code.

Benefits-Quarterly decrease of false alarms by 70% and 55% true cases captured by the model.