Traffic crashes remain a major public health and socioeconomic challenge in Greece and worldwide, underscoring the need for data-driven methodologies to support effective prevention strategies and policy design. Despite recent improvements in infrastructure and law enforcement, the frequency of traffic crashes continues to impose a significant societal burden and needs further research. The present study aims to add to current knowledge and thus carries out a comprehensive analysis and forecasting of crashes and fatalities in Greece, by utilizing an extensive time series dataset spanning 1996–2022, examined at both weekly and monthly resolutions. The utilized dataset was enriched with socioeconomic indicators, seasonal patterns as well as policy-related variables to capture temporal fluctuations and structural changes in crash dynamics. Prior to modelling, the data underwent preprocessing and feature engineering to create lag variables, cumulative indicators and binary policy flags, thereby enhancing and better representing temporal dependencies. The proposed methodological framework combines traditional statistical models, such as the Auto Regressive Integrated Moving Average (ARIMA) and its seasonal extension (SARIMA) with advanced Deep Learning architecture based on Long Short-Term Memory (LSTM) neural networks. The main objective was to assess and compare the predictive performance of traditional linear approaches with nonlinear recurrent models capable of learning long term dependencies. Each time-series model was tuned via grid-based parameter search and afterwards its performance was reported on a fixed out-of-sample holdout period. Evaluation was conducted using quantitative performance metrics, including the Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE), so as to allow for comparative analysis. The results demonstrate that the LSTM model performs slightly better than the ARIMA and SARIMA models across all configurations, particularly in short-term forecasts of traffic crashes and fatalities. The reason behind these findings, is that the neural network effectively captured nonlinear relationships, seasonal patterns and exogenous influences, such as economic and legislative changes which conventional statistical models were not capable of adequately representing. The present analysis also revealed that crash occurrence exhibits strong seasonal trends, characterized by pronounced peaks during summer and holiday periods, as well as moderate correlations with selected socioeconomic variables. Overall, the study highlights the capabilities of deep learning methods when modelling complex and non-stationary crash data and demonstrates the potential of machine learning for advancing predictive analytics in transport safety research in Greece, where related studies are relatively limited. Lastly, the findings of our study provide a solid foundation for proactive decision-making and the development of targeted, evidence-based road safety strategies in Greece and beyond.
| ID | pc675 |
| Presentation | |
| Full Text | |
| Tags |
