Abstract
This paper explores the use of modern time series techniques to understand and forecast urban traffic patterns in Los Angeles. By leveraging an ensemble of statistical models, the authors demonstrate a robust methodology for predicting monthly traffic accidents. The study utilizes an Online Gradient Descent (OGD) ensemble that combines ARIMA, Exponential Smoothing, and Fourier-based models. The OGD approach achieved state-of-the-art accuracy and was the only model to successfully pass all statistical diagnostic tests for residual independence, providing a reliable framework for urban safety planning.
Introduction
Traffic accidents represent a significant public safety and urban planning challenge, particularly in a car-centric metropolis like Los Angeles. While traffic data often exhibits clear trends and seasonal patterns, it is also subject to significant noise and complex, non-linear fluctuations. Traditional individual models, such as standard ARIMA, often fail to capture the full complexity of the data, leaving residual patterns that indicate a lack of model fit. This paper seeks a modeling strategy that truly captures the underlying data-generating process to improve forecasting reliability.
Methods
The methodology employs a "wisdom of the crowd" approach, integrating several classical and modern time series models.
- Data Preprocessing: To stabilize variance and achieve stationarity, the researchers applied Box-Cox transformations and first-order differencing.
- Candidate Models: The study tested ARIMA/IMA models, Exponential Smoothing (ETS) for handling trends and seasonality, and Dynamic Harmonic Regression (DHR) with TBATS, which utilize Fourier terms to model complex, multi-period seasonality.
- Ensemble Forecasting: The core innovation is the use of Online Gradient Descent (OGD) to dynamically weight the predictions of the candidate models. Unlike simple averaging, OGD learns which models are performing best over time and adjusts their influence accordingly.
Data Analysis
The study utilizes a dataset of traffic collisions reported to the Los Angeles Police Department (LAPD) from 2010 to 2019, comprising 113 monthly data points.
The data exhibits a clear increasing trend over the decade and significant seasonality, with accidents typically peaking in October and dipping in February. Analysis revealed that Fourier terms were essential for capturing nuanced seasonal patterns that standard seasonal ARIMA models missed.
Key Results
The OGD Ensemble outperformed all individual component models across all evaluation metrics on the validation set. It achieved a Root Mean Square Error (RMSE) of 114.14, significantly lower than the 153.12 achieved by the best ARIMA model. Additionally, the ensemble's Symmetric Mean Absolute Percentage Error (sMAPE) was 2.32%, indicating high precision.
Crucially, the OGD model was the only one where residuals passed both the Shapiro-Wilk test for normality and the Ljung-Box test for independence. This confirms that the ensemble successfully extracted almost all available information from the historical data, leaving only random white noise.
Conclusion and Future Work
While the OGD ensemble provides a powerful tool for traffic accident prediction, the authors noted a slight tendency to underestimate actual accident counts in the final forecast period. Future improvements could involve incorporating external variables such as weather events, economic shifts, or city policy changes. Furthermore, moving beyond Gaussian distributions to Poisson or Negative Binomial regression could better account for the discrete nature of accident counts, potentially further increasing model accuracy.
Summary updated