Predict & Simulate

Predicting Global CO2 Emissions

Which countries are on track to blow past their emissions targets, and why? I trained Random Forest and XGBoost models on 25 years of country-level data to predict per-capita CO2 trajectories — then used feature importance analysis to identify the socioeconomic and energy-mix variables that actually drive national emissions profiles.

So what: The model revealed that energy mix composition and GDP growth rate explain more emissions variance than population or industrialization level — suggesting that energy policy, not demographic change, is the key lever.

Citi Bike Expansion in Upper Manhattan

Where should NYC add Citi Bike capacity in upper Manhattan to maximize ridership and climate co-benefits? Working from a full year of 2023 Citi Bike trip records (12 GB across 12 months), I scored every existing station in Inwood, Washington Heights, and the Upper West Side — then ran a pairwise optimization that ranks candidate expansions by demand and by CO₂ and PM2.5 avoided per trip diverted from cars.

So what: A handful of nodes — Broadway & W 185 St, Seaman Ave & Isham St, and Dyckman St & Staff St — anchor most of the top 20 recommended pairs. Together those expansions project ≈100,000 new bike-miles, 57 kg of CO₂ avoided, and 4 kg of PM2.5 emissions prevented annually — concentrated in the neighborhoods where the new capacity would land.

Air Quality & PM2.5 Reduction Scenarios

What if 20% of subway commuters switched to micromobility? I modeled the air quality and climate co-benefits of mode-shift scenarios, estimating PM2.5 exposure avoided and CO2 reductions per trip.

So what: Even a modest 20% shift produces measurable health benefits — equivalent to removing thousands of car-trips from the most polluted corridors. The findings support micromobility subsidy arguments.

Transit Ridership Prediction

Can you predict how many people will ride transit in a given neighborhood based on how the system is built around them? I built OLS and Random Forest models across 2,181 NJ census tracts, engineering spatial features from 31,000+ bus stops and 165 rail stations.

So what: Bus stop density alone accounts for 55% of the Random Forest's predictive power — more than all demographic variables combined. Access, not demographics, drives ridership.

# Engineer transit accessibility features per tract
for tract in gdf.geometry:
    # Count bus stops within 2-mile buffer
    buffer = tract.buffer(3218)  # meters
    stops_nearby = bus_stops[bus_stops.within(buffer)]
    gdf.loc[idx, 'bus_density_2mi'] = len(stops_nearby)

    # Distance to nearest rail station
    gdf.loc[idx, 'rail_dist_m'] = tract.distance(
        rail_stations.unary_union
    )

Embodied Carbon Scenario Modeling

How much carbon could NYC save if it mandated low-carbon concrete in all new construction? I built scenario models estimating embodied carbon reductions under four policy pathways — procurement reform, subsidies, regulation, and a hybrid approach.

So what: The hybrid scenario projects 57 million tCO₂e in savings over 20 years and reaches market viability in 5 years. Procurement reform alone saves 19M — doing nothing costs us all of it.

Read the full thesis deep-dive →