Parth
Rangarajan
I spent three years at Chubb modelling what goes wrong in insurance: claim severity, litigation duration, and the cost of waiting. Now I am building my financial mathematics intuition to price that risk in markets as well.
Looking for roles in insurance risk, actuarial analytics and quantitative finance.
| Result | Outcome | Method |
|---|---|---|
| 3%*** | Total operational cost saved by forecasting claim-resolution timelines and flagging early settlement | KOPPA survival-weighted duration model (§2) |
| 1st | Place at a company-wide hackathon; the proof of concept was adopted by senior leadership | Claims severity model, built with an Actuary in 3 days |
| 120 | Insurance products onboarded in one month by reusing eligibility and compensability rules | LangChain retrieval framework |
| NAICS | Industry codes predicted from scraped company data as a Data Science Trainee | Custom named-entity recognition |
| 3+ | Years building and deploying statistical models in production, with CI/CD | Python · R · SQL · GitHub Actions |
Dribbling through my career
Each cone is a checkpoint in my career. Pick a cone, or a role below, and watch me dribble to it. The goal at the far end is the next step.
KOPPA: survival analysis for claim flags
At Chubb I built KOPPA (Key-based Optimum Penalized Proportional-hazard Aggregations) to predict how long a claim stays open. A Cox survival model turns yes-or-no claim flags into weights, then a random forest learns from the weighted data. Build a claim below and watch each step.
- Step 1Cox proportional hazards modelFits one hazard ratio per binary flag
- Step 2Hazard ratios × flag matrixEach 1 becomes its ratio; each 0 stays 0
- Step 3Random forest regressionLearns from the weighted matrix
- Step 4Predicted claim durationFlags early settlement and long litigation
- h(t | x)the hazard: the rate at which a claim closes at day t, given it is still open
- h₀(t)the baseline hazard, for a claim with every flag set to 0
- xka binary flag on the claim file: 1 if present, 0 if not
- βk(beta) the coefficient the Cox model estimates for flag k
- HRkthe hazard ratio: above 1 the claim closes faster, below 1 it drags on
- pthe number of flags in the model
Hazard ratios here are illustrative, chosen to show the mechanics. They are not Chubb figures.
How to read this. S(t) = exp(−λ0t · ∏ HRkxk) is the share of claims like this one still open after t days, where λ0 (lambda) is the baseline closing rate and ∏ means "multiply together". The dashed curve is a claim with no flags. Flags with a ratio below 1, like litigation, push the curve right: the claim stays open longer.
A plain model sees only 0s and 1s, so a litigation flag looks no different from any other flag. KOPPA multiplies each column by its hazard ratio, so every flag carries its own weight. The first row is the claim you built.
Accuracy rose from 18% to 50% once the flags carried hazard-ratio weights: a lift of about 30 percentage points. The target was claim duration, grouped into buckets.
Checking the assumption. Cox models assume each hazard ratio stays constant over the life of a claim. This is tested with Schoenfeld residuals, which measure how far each flag's observed risk drifts from what the model expects over time.
Where I fit
I am open to any role in insurance risk or finance, and I am especially drawn to quantitative risk work.
- (a)Insurance risk and ERMCapital modelling, reserving analytics, asset-liability management
- (b)Actuarial and pricing analyticsClaims frequency and severity, experience studies, predictive pricing
- (c)Market and credit riskVaR and stress testing, credit exposure, model validation
- (d)Quantitative analyticsDerivative and asset valuation, fixed income, time-series modelling
University of Toronto, Department of Statistical Sciences · 2026 – expected Fall 2027
Mathematical Theory of Finance · Insurance Risk Management · Applied Time-Series Analysis
Vishwakarma University, Pune · 2019 – 2023 · IBM Best Project award
Football, tennis, astronomy, theatre and board games. Speaks English, Hindi, Marathi and Tamil.