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.

from sklearn.ensemble import RandomForestRegressor

# Train RF on 25 years of country-level emissions data
rf = RandomForestRegressor(n_estimators=500, max_depth=12)
rf.fit(X_train, y_train)

# Extract and rank feature importances
importances = pd.Series(
    rf.feature_importances_, index=feature_names
).sort_values(ascending=False)

# Top predictor: energy_mix_fossil (0.31)
# R² = 0.89 on held-out test set

Citi Bike Expansion Optimization

Where should NYC put 20 new Citi Bike stations to maximize ridership while improving equity? I analyzed 87MB of trip data to model demand patterns, then built an optimization framework balancing projected ridership against access gaps in underserved neighborhoods.

So what: The equity-weighted model shifted 8 of 20 recommended stations to transit deserts in the outer boroughs — locations a pure ridership-maximizing model would have ignored.

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 →