Forecasting Medical Charges (ARIMA/ARX) + Optimization (LP)

Coursework project combining predictive analytics (time-series forecasting) and prescriptive analytics (linear programming) on the Kaggle Medical Cost dataset.

This repo reproduces the workflow described in the report:

• Part A: ARIMA vs ARX forecasting on a BMI-sorted “pseudo-time” series for the most common age group (age 18).
• Part B: Budget allocation between a smoking-cessation program and a weight-reduction program to maximize net savings.
• Part C: A forecast-to-decision pipeline that uses ARX forecasts as inputs to the optimization.

Key results (from the report)

Part A (Age 18, BMI-sorted pseudo-time)

• ARIMA(1,0,0): AIC 1474.21, RMSE 10,099.25
• ARX(1,0,0) with exogenous variables (BMI, sex, smoker, region): RMSE ~5,281 and improved AIC (see report)

Part B (Optimization, static averages)

• Budget: $120,000
• Costs: $500 per smoker program participant, $400 per obese program participant
• Reported optimal allocation: 137 smokers, 128 obese, spend $119,700
• Reported savings (Part B section): gross ~$1.57M, net ~$1.45M

Part C (Forecast-to-decision)

• Uses ARX(1,0,0) with BMI as exogenous to produce one-step-ahead forecasts for smokers and obese subgroups.
• Reported optimal allocation remains 137 / 128, with higher projected savings (see report’s Executive Summary).

How to run

1. Install dependencies

pip install -r requirements.txt

2. Run each part

python -m src.part_a_forecasting
python -m src.part_b_optimization
python -m src.part_c_forecast_to_decision

Data

• data/raw/medical_cost_sample.csv (sample for repo) — Kaggle “Medical Cost Dataset” (see report for citation/source).

Repo structure

• src/part_a_forecasting.py — ARIMA vs ARX on age-18 BMI-sorted pseudo-time series
• src/part_b_optimization.py — LP + heuristic on static averages
• src/part_c_forecast_to_decision.py — ARX subgroup forecasts → LP decision
• docs/ForecastingMedicalCharges.docx — original report

Notes

• This dataset is cross-sectional; the “time series” is a pseudo-time ordering (BMI-sorted), used for demonstrating ARIMA/ARX mechanics.
• Some reported values (e.g., AIC) can vary slightly by statsmodels version and default settings.

Getting the full dataset

To reproduce the exact report results, download the Medical Cost Dataset from Kaggle (see report) and place it at data/raw/medical_cost.csv.