Suman Ghosh

IRRBB: Hedging Strategy Optimization

• This was to help the Treasury team optimize hedging to protect the bank from interest rate risk.
• Build a mortgage prepayment model to predict prepayment behavior to help hedge rebalancing, net interest income, using multinomial logistic regression.
• Used a single-layer neural network to optimize parameters across prepayment types
• Written modular production code in Python and followed test-driven development principles.
• Prepayment curve assumptions helped the business optimize the hedge by $3.2 billion.
• Build an internal transfer price engine to close the gap between balance amortization and the ITP profile, attributing P&L contribution by products.

Anomaly Detection in Assurance

• The primary responsibility was to help financial controllers and the Assurance team in identifying unusual patterns in the general ledger for analytical review.
• Build rules-based weighted scoring mechanisms to identify and prioritize riskier observations.
• Build unsupervised machine learning solutions like 'Isolation Forest,' 'Auto-Encoder,' and 'Variational Auto Encoder' to identify unknown patterns given.
• Deployed a dashboard using 'Looker' for users, with a feedback loop.
• Users are leveraging the solution for the monthly analytic review process and identified more than 10 unusual cases on average.

NxtGen Toolkit

• Design and build an automated model-building platform for finance analytics, similar to AutoML, but in-house.
• Platform for users (developers) with a no-code solution and a UI to choose candidate models, making the decision-making process faster.

RAG - Finance FIM - Prototype

• The primary responsibility was to build Retrieval Augmented Generation (RAG) to help businesses with process improvement and resource optimization by providing draft answers on the questions from the first line of defense.
• Leveraged Langchain with the text-embedding model of OpenAI and GPT-4.
• Both FAISS and the Azure Vector database were used to store the embedding.
• Used multi-query, reranking, Azure Cognitive Search. UI has been developed using Gradio.

Group Forecast Modeling (Stress Testing Scenarios)

• The primary responsibility was to help businesses in building driver-based balance and P&L projection models.
• Created the model for forecasting deposit balances (inflow, outflow, new origination) across the UK portfolio using macroeconomic drivers and traditional approaches, followed by Net Interest Income calculation using balance, value of funds, and liquidity premium.

Campaign Optimization (Response model for big ticket purchase)

• The primary responsibility was to identify customers who can make purchases of over $799 in the next three months for a retail partner.
• Built the framework for account-level targeting with the use of on-us, contact history, and off-us bureau behaviors.
• Worked closely with the Model Risk Management, Fair Lending, and Model Implementation teams for required approvals.
• Impact: Each campaign is expected to provide $13.2 million of additional sales, with an annual impact of $78 million.

Covid-19 Analysis - Sales Change Driver Analysis

• Identify drivers for sales change in the post/pre-COVID situation, and create triggers to arrest the sales drop in a campaign.
• Created a multi-class solution for getting the drivers across segments of sales change (Decrease, No Change, Increase).
• Developed rules based on important drivers for sales, arrest triggers, and implemented them in the campaign.

Account Targeting (Propensity Model for new lead generation)

• The primary responsibility was to identify customers whom GE can target and when to target.
• Built the framework for account-level targeting in B2B using hospital demographics and purchase history for multiple modalities (CT, MR, etc.). across the US, Canada, Japan, and Africa.
• Helped the commercial team enhance the future funnel by prioritizing existing customers and adding new customers to the funnel.
• Impact: Generated new orders worth $600K in Africa for the value CT machine and $2M in the funnel.

Forecast Market Revenue

• The primary responsibility was to generate future market revenue (quarterly) for strategic decisions.
• Build and deploy a forecast engine on the cloud, and product managers leveraged the same for market numbers, which drive discussions for quarterly op-mech.
• Leveraged GDELT for qualitative justification.

• The primary responsibility was to identify the buzz about the client's product on social media, and how to increase engagement levels across all platforms.
• Developed tools for collecting data from Twitter, Facebook, and YouTube.
• Extract the trending topics on Twitter about their products using topic modeling.
• Leveraged association analysis and sentiment analysis on the post to see the trends in social media.