Skip to content
Sayash RaajRésumé

Work

Quantitative modeling · Case study

Pre-Medicare healthcare cost simulation

Monte Carlo model of pre-Medicare healthcare costs, where right-skewed outcomes make point estimates misleading.
Organization
Fidelity Investments
Context
Lead, Financial Solutions, Quant Strategy
Period
2024 – 2026
Methods
Monte Carlo simulation · Scenario design · Distribution modeling
Tools
Python · NumPy · Snowflake

32%

reduction in expected policy costs

Target cohorts, scenario-design experiments

10,000

simulated scenarios

Note This page describes the modeling approach in general terms. The chart uses a toy process and synthetic numbers; it is not model output and contains no Fidelity data.

01Problem

In the US, Medicare eligibility generally begins at 65, and healthcare costs before then have to be covered in other ways. Those costs are uncertain and right-skewed: most years are ordinary, but occasional high-cost events dominate the total.

For a skewed quantity, the mean is not the typical outcome and the median understates what has to be funded. Planning decisions need the distribution, including its tail, and a way to compare decisions against that distribution.

02Method

  • Data: cost inputs assembled through Snowflake data pipelines.
  • Simulation: a vectorized Monte Carlo engine in Python/NumPy generating 10,000 scenarios, producing cost distributions under uncertainty rather than single estimates.
  • Scenario design: experiments comparing alternatives for target cohorts across the simulated distribution.
  • Monitoring: outputs integrated with operational monitoring dashboards.
Illustrative · synthetic data
Illustrative Monte Carlo fan chart of cumulative costAcross 2000 synthetic paths over 10 years, the median final cumulative cost is 100, the mean is 125, and the 95th percentile is 251.0501001502002503000246810YearCumulative cost (median at year 10 = 100)95th · 251Mean · 125Median · 1005th · 68Year-10 distribution+58 paths above 300
Why a distribution matters. Cumulative cost across simulated paths of a toy lognormal process with occasional high-cost events: the mean sits above the median, and the 95th percentile is far above both.

03Results

32%

reduction in expected policy costs

for target cohorts in scenario-design experiments

10,000

Monte Carlo scenarios

vectorized NumPy simulation

04Limitations

  • The 32% figure is for target cohorts under designed scenarios, not a population-wide average.
  • Monte Carlo output is only as good as the input distributions; tail behaviour is the hardest part to estimate and the part that matters most.