Crash Risk Prediction Research Articles

Stock price crash risk prediction based on high-low frequency dual-layer graph attention network

The phenomenon of a stock price crash involves a rapid, significant decrease in stock prices, severely impacting the market, investors, and the economy. This study introduces the BiGAT-GRU model, which combines Graph Attention Networks (GAT) and Gated Recurrent Units (GRU) to predict stock price crash risk by analyzing multi-scale investor sentiment propagation using data from Baidu search index and public opinion texts. The model demonstrates superior performance in predicting crash risk, providing valuable insights for policymakers and investors.

Deep Q‐network learning‐based active speed management under autonomous driving environments

AbstractEfficient traffic safety management necessitates real‐time crash risk prediction using expressway characteristics. With the emergence of autonomous vehicles (AVs), the development and evaluation of variable speed limit (VSL) strategies, a key active traffic management technique, become crucial for enhancing safety and mobility in mixed traffic flows. This underscores the need for optimized VSL strategies to accommodate both conventional and AVs. This paper presents a study on the development of VSL control algorithms using deep reinforcement learning in a microscopic traffic simulation. As the rewards function, time‐to‐collision and speed were considered. To enhance traffic safety, VSL strategies were refined across various market penetration of connected AVs. Analysis revealed that safety and traffic density are improved by 53% and 59%, respectively, in market penetration rate (MPR) 50, marking significant safety improvements in congested and low MPR scenarios. These findings present the importance of developing and evaluating VSL strategies for mixed traffic flow, particularly in the context of increasing the prevalence of connected and AVs.

A crash occurrence risk prediction model based on variational autoencoder and generative adversarial network

Traffic crash risk prediction is pivotal for proactive traffic safety management but faces challenges due to the extreme imbalance between crash and non-crash data. This paper proposes integrating variational autoencoder into a generative adversarial network (VAE-GAN) for crash data augmentation without information loss. VAE-GAN generates higher-quality data due to its superior deep generative capabilities. For the crash occurrence risk prediction task, we utilize a convolutional neural network (CNN) trained on the balanced datasets generated by VAE-GAN. Two kinds of category determinations are tested for better prediction results by using prediction probability maximum and selecting threshold. The prediction model integrating variational autoencoder into a deep convolutional generative adversarial network (VAE-DCGAN) exhibits the best performance. Furthermore, we leverage Shapley additive explanations (SHAP) to interpret the key features and patterns impacting prediction results. This analysis gains insights into the underlying mechanisms of crash occurrences, and helps improve proactive traffic management and control strategies.

Assessing the predictability of surrogate safety measures as crash precursors based on vehicle trajectory data prior to crashes

This study aims to investigate the predictability of surrogate safety measures (SSMs) for real-time crash risk prediction. We conducted a year-long drone video collection on a busy freeway in Nanjing, China, and collected 20 rear-end crashes. The predictability of SSMs was defined as the probability of crash occurrence when using SSMs as precursors to crashes. Ridge regression models were established to explore contributing factors to the predictability of SSMs. Four commonly used SSMs were tested in this study. It was found that modified time-to-collision (MTTC) outperformed other SSMs when the early warning capability was set at a minimum of 1 s. We further investigated the cost and benefit of SSMs in safety interventions by evaluating the number of necessary predictions for successful crash prediction and the proportion of crashes that can be predicted accurately. The result demonstrated these SSMs were most efficient in proactive safety management systems with an early warning capability of 1 s. In this case, 308, 131, 281, and 327,661 predictions needed to be made before a crash could be successfully predicted by TTC, MTTC, DRAC, and PICUD, respectively, achieving 75 %, 85 %, 35 %, and 100 % successful crash identifications. The ridge regression results indicated that the predefined threshold had the greatest impact on the predictability of all tested SSMs.

A variable speed limit control approach for freeway tunnels based on the model-based reinforcement learning framework with safety perception

To improve the traffic safety and efficiency of freeway tunnels, this study proposes a novel variable speed limit (VSL) control strategy based on the model-based reinforcement learning framework (MBRL) with safety perception. The MBRL framework is designed by developing a multi-lane cell transmission model for freeway tunnels as an environment model, which is built so that agents can interact with the environment model while interacting with the real environment to improve the sampling efficiency of reinforcement learning. Based on a real-time crash risk prediction model for freeway tunnels that uses random deep and cross networks, the safety perception function inside the MBRL framework is developed. The reinforcement learning components fully account for most current tunnels' application conditions, and the VSL control agent is trained using a deep dyna-Q method. The control process uses a safety trigger mechanism to reduce the likelihood of crashes caused by frequent changes in speed. The efficacy of the proposed VSL strategies is validated through simulation experiments. The results show that the proposed VSL strategies significantly increase traffic safety performance by between 16.00% and 20.00% and traffic efficiency by between 3.00% and 6.50% compared to a fixed speed limit approach. Notably, the proposed strategies outperform traditional VSL strategy based on the traffic flow prediction model in terms of traffic safety and efficiency improvement, and they also outperform the VSL strategy based on model-free reinforcement learning framework when sampling efficiency is considered together. In addition, the proposed strategies with safety triggers are safer than those without safety triggers. These findings demonstrate the potential for MBRL-based VSL strategies to improve traffic safety and efficiency within freeway tunnels.

A spatiotemporal deep learning approach for pedestrian crash risk prediction based on POI trip characteristics and pedestrian exposure intensity

Pedestrians represent a population of vulnerable road users who are directly exposed to complex traffic conditions, thereby increasing their risk of injury or fatality. This study first constructed a multidimensional indicator to quantify pedestrian exposure, considering factors such as Point of Interest (POI) attributes, POI intensity, traffic volume, and pedestrian walkability. Following risk interpolation and feature engineering, a comprehensive data source for risk prediction was formed. Finally, based on risk factors, the VT-NET deep learning network model was proposed, integrating the algorithmic characteristics of the VGG16 deep convolutional neural network and the Transformer deep learning network. The model involved training non-temporal features and temporal features separately. The training dataset incorporated features such as weather conditions, exposure intensity, socioeconomic factors, and the built environment. By employing different training methods for different types of causative feature variables, the VT-NET model analyzed changes in risk features and separately trained temporal and non-temporal risk variables. It was used to generate spatiotemporal grid-level predictions of crash risk across four spatiotemporal scales. The performance of the VT-NET model was assessed, revealing its efficacy in predicting pedestrian crash risks across the study area. The results indicated that areas with concentrated crash risks are primarily located in the city center and persist for several hours. These high-risk areas dissipate during the late night and early morning hours. High-risk areas were also found to cluster in the city center; this clustering behavior was more prominent during weekends compared to weekdays and coincided with commercial zones, public spaces, and educational and medical facilities.

Integrating a non-gridded space representation into a graph neural networks model for citywide short-term crash risk prediction

Current applications of Graph Neural Networks in citywide short-term crash risk prediction have been limited by a gridded representation of space, which restricts the network’s capability to effectively capture the spatial and temporal dependency of crash occurrences. In addition, a grided representation does not match most geographic units used for administrative purposes, limiting the use of crash risk predictions by practitioners. This paper applies a gated localised diffusion graph neural network (GLDNet) model to compare the use of two alternative geographic units, Mesh Block (MB) and grid, to forecast locations where crashes are likely to occur in a future time window. The GLDNet relies on a graph-based representation of geographic units and a weighted loss function to address the sparsity of crash occurrences. The tests are performed using crash data from the City of Melbourne, Australia, over a period of one year. The predictions are made at six-hour intervals, and the results show that the GLDNet consistently outperforms baseline methods, with differences in prediction accuracy from 10% to 21% in relation to historical average and benchmark deep learning models. In terms of geographic units, the MB-based GLDNet performed better than its grid counterpart, with differences in prediction accuracy of up to 12.3%. The better performance stems from the underlying information attached to the MB units (i.e., land use) and the network properties (i.e., degree of centrality), which enhance the GLDNet capability to identify crash risk in both central and peripherical areas. Regarding its applicability, the MB-based GLDNet directly integrates with other data sources, which provides contextual information about crash hotspots that helps decision-makers develop police patrolling and rescuing strategies.

Using Aggregated Fine Geo-Resolution Vehicle Telemetric Data to Predict Crash Occurrence

Being able to predict motor vehicle crashes which are a major public health concern would greatly improve traffic safety. The prevalence of mobile sensing platforms now allows spatially and temporally rich driving data to be collected relatively easily. Research efforts have been devoted to predicting crashes from such data. This paper seeks in particular to assess the feasibility and performance of using aggregated fine geo-resolution vehicle telemetric data for crash risk prediction. We acquired vehicle telemetric data from Geotab Inc., which recorded the frequency of hard acceleration, hard braking, harsh cornering, and the average magnitude of those harsh events among its registered commercial vehicles for every 150 × 150 m2 roadway segment within Columbus, Ohio between January and April 2018. We aggregated the data, obtained the crash history from the Ohio Police Accident Report, and leveraged three machine learning–based algorithms to predict the crash risk. The results suggest that aggregated vehicle telemetric data could provide acceptable predictions for crash risk at a roadway segment level. Our models’ predictive performances were further improved and maximized by including in the models both vehicle telemetric data and roadway geometric characteristics. Several factors, such as the aggregated count of hard accelerations and the presence of an intersection, were shown to be the factors that potentially made the greatest contribution to crash occurrence. We concluded that vehicle telemetric data could provide complementary and valuable information about crash likelihood monitoring, which may enable the police and city planners to implement proactive safety interventions. Yet the nature of traffic crashes is still complex and multi-dimensional.

Studying the predictability of crash risk caused by manual takeover of autonomous vehicles in mixed traffic flow

ABSTRACT This study explores how to reduce the cost of prediction as much as possible while ensuring the prediction accuracy of a real-time crash risk model. The extreme gradient enhancement (XGBoost) algorithm was used to predict the crash risk of autonomous vehicles in different sections of highway. The results show that the prediction performance of the model is the best when the threshold value is 0.05. Choosing two variables to predict can ensure high accuracy and simultaneously reduce the cost of prediction when the accuracy of crash risk prediction of the three sections can reach 73%, 62%, and 70%. However, when only one variable can be selected due to sensor or system failure, the speed difference between the takeover car and the front car can be chosen to achieve the greatest benefit. These findings could provide a reference for technicians to design safer and more economical autonomous vehicles.

Real-time crash risk prediction in freeway tunnels considering features interaction and unobserved heterogeneity: A two-stage deep learning modeling framework

Real-time prediction of crash risk is an effective method for enhancing traffic safety, but it is not fully explored in freeway tunnels. A two-stage deep learning modeling framework comprising a preliminary exploration stage and a prediction and analysis stage is proposed for real-time crash risk prediction in freeway tunnels. A random parameters logit model with heterogeneity in means and variances is used in the preliminary exploration stage to investigate the unobserved heterogeneity and influence mechanism of precursors on real-time crash risk. In the prediction and analysis stage, a random deep and cross network model considering feature interactions and unobserved heterogeneities is developed to predict and analyze real-time crash risk, which is interpreted by the shapley additive explanations approach. The multi-source fusion dataset, collected from the Caltrans performance measurement system and the weather information website, is used to validate the proposed framework for exploring real-time crash risk in freeway tunnels. Results reveal that: (1) the random parameters logit model with heterogeneity in means and variances outperforms the traditional logit model in terms of the model fitting, providing a reference for deep learning modeling that may be able to improve model performance by addressing heterogeneity; (2) the important crash precursors such as the average difference in speed between detectors of tunnel entrance and exit are discovered based on the marginal effect analysis of the random parameters logit model with heterogeneity in means and variances; (3) the random deep and cross network model yields the best prediction performance compared to its counterparts (some other data-driven models), demonstrating the superior performance of deep learning models for real-time risk prediction tasks. It also indicates that considering feature interaction and heterogeneity in deep learning modeling can improve prediction performance; and (4) the important precursors found in the random deep and cross network model using the shapley additive explanations approach are close to those discovered in the statistical model, indicating that the proposed deep learning model can capture the similar effects of precursors as the statistical models, and the precursor interactions and heterogeneities also can be observed by the shapley additive explanations approach.

Robust spatiotemporal crash risk prediction with gated recurrent convolution network and interpretable insights from SHapley additive explanations

The vigorous progress in artificial intelligence and the widespread availability of computing power and big data have resulted in remarkable achievements in applying deep learning methods to traffic engineering. However, the complexity of deep learning models poses a significant challenge to their interpretability, despite their scalability to big data. As a result, academia has shown a keen interest in interpreting these intricate models. To address these limitations, this study introduces the Gated Recurrent Convolution Network (GRCN) for predicting spatiotemporal crash risks. Moreover, to transform the opaque nature of the GRCN into a transparent framework, SHapley Additive Explanations (SHAP) was employed, a game theory-based approach. The application of SHAP analysis not only enhances the explainability of the GRCN model but also reveals complex causal relationships regarding crash risks, using extensive datasets encompassing taxi trips data, law enforcement, and weather conditions. The study incorporates both local and global interpretability analyses at the grid level. Additionally, it indicates that intersection red light violations tend to elevate crash risks. The findings further demonstrate that adverse weather conditions, such as low visibility and strong crosswinds, negatively affect driving behavior and correlate with higher crash risks. However, the presence of cloudy skies is associated with a decrease in crash risks. By shedding light on the intricate dynamics of crash risk factors, this study contributes to the field of traffic engineering and emphasizes the importance of interpretability in deep learning models. The insights gained from this research can inform the development of effective interventions and policies aimed at reducing crash risks and improving overall transportation safety.

Design of a Novel Convolutional Deep Network Model for Car Accident Prediction

Real-time collision risk estimation is thought to be essential to a sophisticated traffic management system. To swiftly determine accident probability is the goal of real-time crash risk prediction. However, due to the complex traffic situation on urban arterials, urban arterials were rarely included in previous studies, which mostly focused on highways. This paper suggests using Convolutional Deep Network model (CDNM) to forecast the probability of vascular accidents in real time. This model has the benefit of being able to use both LSTM and CNN. CNN retrieves the time-invariant characteristics, while LSTM captures the data's long-term dependability. To estimate the likelihood of an accident, many sorts of data are used, including weather, traffic, and signal timing data. There are also many other data preparation methods employed. The problem of data imbalance is also addressed by normalization which oversamples the crash cases. Using a variety of measures, the CDNM is enhanced on the training data and assessed on the test data. Five more benchmark models are constructed for model comparison. K-NN, ISVM, ANN, CNN, CNN-EVT and GAN are some of the models in this group. Experimental findings show that the proposed CDNM beats the competition in terms of sensitivity, specificity, accuracy, AUC and G-mean value. The findings of this paper demonstrate that CDNM can real-time prediction of crash risk at arterials.

On the prediction of stock price crash risk using textual sentiment of management statement

PurposeSince stock return and volatility matters to investors, this study proposes to incorporate the textual sentiment of annual reports in stock price crash risk prediction.Design/methodology/approachSpecific sentences gathered from management discussions and their subsequent analyses are tokenized and transformed into numeric vectors using textual mining techniques, and then the Naïve Bayes method is applied to score the sentiment, which is used as an input variable for crash risk prediction. The results are compared between a collection of predictive models, including linear regression (LR) and machine learning techniques.FindingsThe experimental results find that those predictive models that incorporate textual sentiment significantly outperform the baseline models with only accounting and market variables included. These conclusions hold when crash risk is proxied by either the negative skewness of the return distribution or down-to-up volatility (DUVOL).Research limitations/implicationsIt should be noted that the authors' study focuses on examining the predictive power of textual sentiment in crash risk prediction, while other dimensions of textual features such as readability and thematic contents are not considered. More analysis is needed to explore the predictive power of textual features from various dimensions, with the most recent sample data included in future studies.Originality/valueThe authors' study provides implications for the information value of textual data in financial analysis and risk management. It suggests that the soft information contained within annual reports may prove informative in crash risk prediction, and the incorporation of textual sentiment provides an incremental improvement in overall predictive performance.

Exploring key spatio-temporal features of crash risk hot spots on urban road network: A machine learning approach

Traffic safety is a critical factor that has always been considered in policy making for urban transportation planning and management. Accurately predicting crash risk hot spots allows urban transportation agencies to better implement countermeasures towards enhancing traffic safety. Considerable efforts have been devoted to investigate crash risk hot spots in many previous studies. We hereby identify three research gaps that remain to be resolved: first, the effects of spatio-temporal features surrounding hot spots are often ignored; second, false discovery rates tend to be higher when applying local spatial indices to identify hot spots; and third, the spatio-temporal correlations and heterogeneity of crash-related features in a subject spot and its neighboring spots have not been well captured in most crash risk prediction models. To fill these gaps, we propose an urban crash risk identification model by integrating space-time cubes and machine learning techniques. The spatio-temporal correlations and heterogeneity of crash-related features are represented by using statistical descriptions of neighboring cubes. Three space-time cube risk datasets collected from Manhattan in New York City are used to validate the proposed model in the case study. The eXtreme Gradient Boosting (XGBoost) classifier is employed to predict the risk patterns (hot spots, normal spots, and cold spots) of each cube due to its satisfactory prediction performance. The validation results suggest that our proposed model attains lower false discovery rates and higher crash risk prediction accuracy as compared to conventional methods. As the results of the feature selection are empowered by machine learning, we found that most key features are inherent to the features of spatial neighboring cubes, which manifests the importance of considering the features of neighboring spots when identifying crash risk hot spots. Moreover, SHapley Additive exPlanations (SHAP) are employed to interpret the effects of key features on hot spots, upon which the contributions of the features related to urban facilities, public transit, and land use are discussed. Based on the feature interpretation, several policy recommendations could be made to enhance urban road traffic safety in the future.

Modeling of Freeway Real-Time Traffic Crash Risk Based on Dynamic Traffic Flow Considering Temporal Effect Difference

With the development of traffic detection facilities technology, it is currently possible to obtain high-resolution traffic flow data. Due to the particular driving characteristics of vehicles on freeways, once traffic crashes occur, they are generally with serious consequences, and hence traffic safety issues on freeways have always been popular topics. In order to better realize the change from static analysis after the crash to dynamic analysis before the crash toward freeway safety, as well as explore the relationship between dynamic traffic flow characteristics and real-time traffic crash risk under different temporal conditions, this research constructed a real-time traffic crash risk prediction model considering the temporal effect difference. First, traffic crash information and the matched big data of high-resolution traffic flow located on the section of milepost 100–130 of Interstate 5 (I-5) in Washington State, were extracted. In terms of temporal dimension, the research object was divided into weekdays and weekends, and the traffic state was divided into unsaturated and saturated. The random forest (RF) algorithm was introduced to identify the traffic flow variables of crash precursors, and support vector machine (SVM) was applied to build the traffic crash risk prediction model under the condition of temporal difference. A confusion matrix, receiver operating characteristic (ROC) curve, and area under curve (AUC) values were used to evaluate the accuracy of the model performance. Furthermore, the prediction performance of the proposed model was tested via constructing the risk model without consideration of temporal effect and traffic state difference. Finally, the rationality of variable screening was verified by inputting the data set without variable screening into the constructed model. The results showed that the occurrence mechanism of dynamic traffic crashes under different temporal effect conditions varies; the AUC values of the constructed prediction model were all between 0.7 and 0.9, indicating that the recommended model has good prediction accuracy. In conclusion, the real-time freeway traffic crash risk prediction model considering the temporal effect difference has certain advantages compared with the conventional model, and its performance is better than the prediction model without screening of important traffic flow variables. This approach can provide theoretical guidance for dynamic traffic safety management toward freeway under temporal difference conditions.

A proactive crash risk prediction framework for lane-changing behavior incorporating individual driving styles

Driving style may have an important effect on traffic safety. Proactive crash risk prediction for lane-changing behaviors incorporating individual driving styles can help drivers make safe lane-changing decisions. However, the interaction between driving styles and lane-changing risk is still not fully understood, making it difficult for advanced driver-assistance systems (ADASs) to provide personalized lane-changing risk information services. This paper proposes a personalized risk lane-changing prediction framework that considers driving style. Several driving volatility indices based on vehicle interactive features have been proposed, and a dynamic clustering method is developed to determine the best identification time window and methods of driving style. The Light Gradient Boosting Machine (LightGBM) based on Shapley additive explanation is used to predict lane-changing risk for cautious, normal, and aggressive drivers and to analyze their risk factors. The highD trajectory dataset is used to evaluate the proposed framework. The obtained results show that i) spectral clustering and a time window of 3 s can accurately identify driving styles during the lane-changing intention process; ii) the LightGBM algorithm outperforms other machine learning methods in personalized lane-changing risk prediction; iii) aggressive drivers seek more individual driving freedom than cautious and normal drivers and tend to ignore the state of the car behind them in the target lane, with a greater lane-changing risk. The research conclusion can provide basic support for the development and application of personalized lane-changing warning systems in ADASs.

Highway Traffic Crash Risk Prediction Method considering Temporal Correlation Characteristics

Crash risk analysis and prediction are considered the premise of highway traffic safety control, which directly affects the accuracy and effectiveness of traffic safety decisions. A highway traffic crash risk prediction method considering temporal correlation characteristics is proposed in this research. Firstly, the case-control sample analysis method is used to extract 6 time series sample data composed of crash traffic flow data and corresponding non-crash traffic flow data for crash risk analysis and prediction. Secondly, the multiparameter fusion clustering analysis method is used to indicate that the sample data of different time series have different effects on the crash risk. Then, the random forest model is used to screen several traffic flow variables that affect the highway crash risk. Thereafter, the downstream mean speed (ASD1D2), the upstream mean occupancy (AOU1U2), and the speed difference (DSU1D1) on the nearest detector were determined as the explanatory variables of the crash risk prediction model. Finally, based on the three variables, the dynamic Bayesian network model for highway traffic crash risk prediction is proposed. The overall prediction accuracy of this model is 84.9%, the crash prediction accuracy is 60.8%, and the non-crash prediction accuracy is 92.3%. Also, the prediction results show that the dynamic Bayesian model has better prediction effect than the static Bayesian model for the same sample data.

Road crash risk prediction during COVID-19 for flash crowd traffic prevention: The case of Los Angeles

Road crashes are a major problem for traffic safety management, which usually causes flash crowd traffic with a profound influence on traffic management and communication systems. In 2020, the sudden outbreak of the novel coronavirus disease (COVID-19) pandemic led to significant changes in road traffic conditions. In this paper, by analyzing crash data from 2016 to 2020 and new COVID-19 case data in 2020, we find that the average crash severity and crash deaths during this period (a rapid increase of new COVID-19 cases in 2020) are higher than those in previous four years. Hence, it is necessary to exploit a novel road crash risk prediction model for such an emergency. We propose a novel data-adaptive fatigue focal loss (DA-FFL) method by fusing fatigue factors to establish a road crash risk prediction model under the scenario of large-scale emergencies. Finally, the experimental results demonstrate that DA-FFL performs better than the other typical methods in terms of area under curve (AUC) and false alarm rate (FAR) for imbalanced data. Furthermore, DA-FFL has better prediction performance in convolutional neural networks-long short-term memory (CNN-LSTM).

Attention-Based Lane Change and Crash Risk Prediction Model in Highways

Lane change and crash risk prediction are critical technologies for autonomous driving. An attention-based LSTM model is proposed in this paper for lane change behavior prediction in highways, considering both the surrounding vehicles’ information around the target vehicle at the current time and the historical trajectory data of the vehicle, and it shows higher accuracy and better interpretability than other models. There are two sections in this prediction model, including a pre-judgment model based on the C4.5 decision tree and bagging ensemble learning, and a multi-step lane change prediction model of LSTM with attention mechanism. In addition, the model is applied in the NGSIM datasets to test practicability and accuracy, and the precision of left lane change and right lane change is 98% and 94% respectively 1 second before the vehicle changes lane. Finally, to judge the safety of vehicles in the driving process, a crash risk prediction model based on Time-To-Collision is proposed, which further verifies the effectiveness of the method.

Wasserstein Generative Adversarial Network to Address the Imbalanced Data Problem in Real-Time Crash Risk Prediction

Real-time crash risk prediction models aim to identify pre-crash conditions as part of active traffic safety management. However, traditional models which were mainly developed through matched case-control sampling have been criticised due to their biased estimations. In this study, the state-of-art class balancing method known as the Wasserstein Generative Adversarial Network (WGAN) was introduced to address the class imbalance problem in the model development. An extremely imbalanced dataset consisted of 257 crashes and over 10 million non-crash cases from M1 Motorway in United Kingdom for 2017 was then utilized to evaluate the proposed method. The real-time crash prediction model was developed by employing Deep Neural Network (DNN) and Logistic Regression (LR). Crash predictions were performed under different crash to non-crash ratios where synthetic crashes were generated by Wasserstein Generative Adversarial Network (WGAN), Synthetic Minority Over-sampling Technique (SMOTE) and Adaptive Synthetic (ADASYN) sampling respectively. Comparisons were then made with algorithmic-level class balancing methods such as cost-sensitive learning and ensemble methods. Our findings suggest that WGAN clearly outperforms other oversampling methods in terms of handling the extremely imbalanced sample and the DNN model subsequently produces a crash prediction sensitivity of about 70% with a 5% false alarm rate. Based on the findings of this study, proactive traffic management strategies including Variable Speed Limit (VSL) and Dynamic Messing Signs (DMS) could be deployed to reduce the probability of crash occurrence.