I am a data scientist with around one year of professional experience and an MSc in Statistics with Data Science, with Merit, from the University of Edinburgh. My projects use statistics, numerical modelling and machine learning to investigate forecasting, financial risk and trading questions.
Previously, I worked on delivery staffing, customer-order forecasting and rider incentives at Swiggy, a food-delivery business, and promotion-planning models at PUMA. My earlier degree is an Integrated MSc in Chemistry from IIT Kharagpur. I am based in Bengaluru, India.
Email · All projects · How I evaluate research
| Project | What I investigated | Main finding |
|---|---|---|
| BTC/ETH spot–futures carry | When does a futures premium pay for trading costs and holding capital? | 19.35% simulated net return versus 16.81% for entering every eligible contract, including cash interest. |
| GridPulse: electricity demand | Can past demand predict tomorrow's hourly electricity needs? | Average test error fell 43.3% versus copying the same hour from the previous week. |
| ClearCredit: repayment-history risk | Do ordered repayment histories improve default-risk predictions? | GRU ranked 50.6% of defaults in the highest-risk 20% of accounts; its advantage over boosted trees was small. |
| Used-tractor prices — MSc dissertation | How much can cleaned auction records tell us about sale prices? | Median percentage error fell from 40.2% to 32.0% against quadratic regression. |
| River pollution — MSc dissertation | How do direct-discharge and runoff contributions change over time? | Hierarchical time-varying models helped zinc and orthophosphate predictions, but not every pollutant. |
| Option-price calibration | Does an adaptive penalty recover volatility more reliably from sparse quotes? | Some pricing gains, but unstable known-volatility recovery and weaker performance on some unseen expiries. |
Each repository begins with the problem and result, then links to model details, evaluation and code. Results are labelled as simulations, historical report findings or executed prediction studies as appropriate. The full index also covers market-state volatility forecasting, ride-hailing simulation, statistical coursework and earlier machine-learning notebooks.
I start with a useful baseline and make the prediction target explicit. Then I check whether the extra model complexity helps, examine where it fails and keep the result traceable to evidence. In the credit-risk study, for example, adding deep learning led to a useful failure analysis rather than a claim that neural networks clearly beat trees.
Programming: Python · R · SQL · C++
Methods: time series · Bayesian inference · machine learning · numerical optimisation · stochastic simulation
Tools: NumPy · pandas · SciPy · scikit-learn · PyTorch · LightGBM · CatBoost · brms/Stan · DuckDB · Git · Databricks · Snowflake
I am interested in quantitative research and data science roles where careful analysis and collaboration turn models into useful decisions.