Forecasting EMS Demand in NYC w/ Two Decades of Data
I've been on a kick of rehashing or finishing old work. Today's effort resulted in a draft research article comparing multiple ML/AI approaches for forecasting EMS demand. It's a much more robust approach than I've done in the past, and now I just need to go over the results with a fine-toothed comb.
Abstract:
Accurate forecasting of emergency medical services (EMS) demand is essential for optimizing ambulance deployment, crew scheduling, and station coverage in large cities, yet systematic comparisons across modeling paradigms on long-horizon, large-scale data remain absent. Here we analyze 28.7 million Fire Department of New York (FDNY) dispatch records spanning January 2005 to August 2025 and benchmark six forecasting models—ordinary least squares (OLS) regression, Lasso, Prophet, XGBoost, long short-term memory (LSTM) networks, and Temporal Fusion Transformers (TFT)—at daily citywide and hourly borough-level granularities. We engineer 42 predictive features capturing calendar, meteorological, epidemiological, and pandemic-related signals. At the daily scale (n ≈4,500), XGBoost achieves the lowest error (MAE = 117.9, MAPE = 2.7%) while OLS attains the highest explained variance (R2= 0.62), outperforming both deep learning architectures. At the hourly borough level (n > 160,000),
TFT and XGBoost converge at R2= 0.89, demonstrating that model complexity should be matched to data regime size. Feature importance analyses reveal that weekend effects, temperature, and autoregressive lags are the dominant demand drivers, with influenza surveillance rates providing an independent epidemiological signal. These results offer practical guidance for cities building forecasting infrastructure to support proactive ambulance redeployment and staffing optimization.