Time Series Analysis Model Research Articles

Temporal Variations in Chemical Proprieties of Waterbodies within Coastal Polders: Forecast Modeling for Optimizing Water Management Decisions

In polder-type land, water dynamics are heavily influenced by the artificial maintenance of water levels. Polders are low-lying areas of land that have been reclaimed from the sea or from freshwater bodies and are protected from flooding by dikes or other types of flood-protection structures. The water regime in polders is typically managed using a system of canals, pumps, and sluices to control the flow of water in and out of the area. In this study, the temporal changes in water salinity in the polder-type agricultural floodplain within the Neretva River Delta (NRD), Croatia, were analyzed by applying multivariate statistics and forecast modelling. The main aim of the study was to test the model that can be used in practice to forecast, primarily, water suitability for irrigation in a coastal low-lying agricultural catchment. The specific aim of this study was to use hydrochemistry data series to explain processes in water salinity dynamics and to test the model which may provide accurate salinity prediction, or finally select the conditions in which the model can be applied. We considered the accuracy of the model, and it was validated using independent data sets. To describe different patterns of chemical changes in different water classes due to their complex hydrological connectivity, multivariate statistics (PCA) were coupled with time-series analysis and Vector Autoregression (VAR) model forecasting. The multivariate statistics applied here did not indicate a clear connection between water salinity of the surface-water bodies and groundwater. The lack of correlation lies in the complex hydrological dynamics and interconnectivity of the water bodies highly affected by the artificial maintenance of the groundwater level within the polder area, as well as interventions in the temporal release of freshwater into the drainage canal network. Not all individual water classes contributed equally to the dominant patterns of ionic species identified by PCA. Apparently, land use and agricultural management practices in the different polders lead to uneven water chemistry and the predominant contributions of specific ions, especially nutrients. After applying the Granger causality test to reveal the causal information and explain hidden relationships among the variables, only two surface-water and two groundwater monitoring locations displayed a strong causal relationship between water electrical conductivity (ECw) as an effect and sea level as a possible cause. The developed models can be used to evaluate and emphasize the unique characteristics and phenomena of low-lying land and to communicate their importance and influence to management authorities and agricultural producers in managing and planning irrigation management in the wider Mediterranean area.

Manufacturing conflict or advocating peace? A study of social bots agenda building in the Twitter discussion of the Russia-Ukraine war

ABSTRACT After the outbreak of the Russia-Ukraine war, social media become the major contested field for both sides. Social bots are widely involved in the Russia-Ukraine war discussion as communication actors. However, how social bots build the war agenda has not been well answered. This study explores the role of social bots through time series analysis and Structural Topic Model. The study finds that at the aggregate level there is a two-way agenda-setting between social bots and human accounts, and at the secondary level, social bots show different or even opposite agenda-setting effects in national/regional discussion. In addition, we find that social bots exhibit multiple positions in the conversation, reflecting the complexity of the manipulators behind them. This multiplicity of positions reveals that the role played by social bots is still based on the opposition of one side to the other, and the end result may exacerbate the conflict of online public opinion.

Combined effect of early diagnosis and treatment on the case fatality risk of COVID-19 in Japan, 2020

Japanese government initially enforced restrictions on outpatient attendances among febrile individuals suspected of having COVID-19, asking everyone to remain at home for at least 4 days from the onset of fever. This restriction was cancelled on 8 May 2020, and a new antiviral, remdesivir, was approved from 7 May 2020. To investigate how this policy change influenced the prognosis of people with COVID-19, we estimated the case fatality risk as a function of the date of illness onset from April to June 2020. We used an interrupted time-series analysis model with an intervention date of 8 May 2020, and estimated time-dependent case fatality risk by age group. The case fatality risk showed a decreasing trend in all groups, and models were favored accounting for an abrupt causal effect, i.e., immediate decline in fatality risk. The trend was estimated at − 1.1% (95% CI [confidence interval]: − 3.9, 3.0) among people aged 60–69 years, − 7.2% (95% CI − 11.2, − 2.4) among those aged 70–79 years, − 7.4% (95% CI − 14.2, 0.2) among those aged 80–89 years, and − 10.3% (95% CI − 21.1, 2.7) among those aged 90 and over. Early diagnosis and treatment greatly contributed to reducing the case fatality risk.

Application of wastewater-based surveillance and copula time-series model for COVID-19 forecasts

The objective of this study was to develop a novel copula-based time series (CTS) model to forecast COVID-19 cases and trends based on wastewater SARS-CoV-2 viral load and clinical variables. Wastewater samples were collected from wastewater pumping stations in five sewersheds in the City of Chesapeake VA. Wastewater SARS-CoV-2 viral load was measured using reverse transcription droplet digital PCR (RT-ddPCR). The clinical dataset included daily COVID-19 reported cases, hospitalization cases, and death cases. The CTS model development included two steps: an autoregressive moving average (ARMA) model for time series analysis (step I), and an integration of ARMA and a copula function for marginal regression analysis (step II). Poisson and negative binomial marginal probability densities for copula functions were used to determine the forecasting capacity of the CTS model for COVID-19 forecasts in the same geographical area. The dynamic trends predicted by the CTS model were well suited to the trend of the reported cases as the forecasted cases from the CTS model fell within the 99 % confidence interval of the reported cases. Wastewater SARS CoV-2 viral load served as a reliable predictor for forecasting COVID-19 cases. The CTS model provided robust modeling to predict COVID-19 cases.

Time series analysis and probabilistic model of the financial costs of major disasters in the USA

Time series analysis and probabilistic model of the financial costs of major disasters in the USA

A comparative framework for criticality assessment of strategic raw materials in Turkey

The rapid development of innovative technologies, population growth, shifting resource management patterns, and decarbonization challenges are expected to increase the demand for raw materials labeled as critical in the following years. The increasing concerns have led to evaluate the critical raw materials (CRMs) with high economic importance and supply risk. In this context, this paper aims to develop a conceptual methodology to identify CRMs in Turkey. The methodology integrates a multi-criteria decision-making method (MCDM) to provide the raw material criticality matrix regarding various criteria, the EU (European Union) method, which represents supply risk and economic importance indicators, and the time series analysis method, in which dynamic evaluation of materials over time instead of the static indicators. We first provide a Bayesian Best-Worst method (B-BWM) and TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) based model for the criticality matrix, which enables evaluations of participated experts to be combined to reveal the weights of the criteria and sub-criteria from a probabilistic point of view without loss of knowledge. The main and sub-criteria include supply risk and economic importance, which are frequently studied in the literature, but also environment, price, secondary production, and socio-economic indicators are evaluated in the criticality matrix. Then, exponential smoothing forecast, one of the Time Series Analysis models, which determine the element criticality in the perspective of future years with the data of import figures of Turkey, is superior to the other models. Finally, with some modifications, the EU criticality methodology is used to quantify the raw materials regarding each material associated with the sector applications and their value-added and supply disruptions. Thus, a comprehensive criticality list integrating the results of the three mentioned methods is important because Turkey's raw materials characterize many sectors. The findings should also interest practitioners and researchers as the detailed assessment of raw material criticality ensures a sufficient database for future studies.

Sales Prediction Based on ARIMA Time Series and Multifactorial Linear Model

Contemporarily, sales forecasting based on historical data is one of the key topics of research. Researchers have found that different analytical forecasting models can predict the future sales volume to a certain extent based on different actual conditions. However, it is also very important to compare the prediction results of these models, and a reasonable choice of models can help managers to make more scientific decisions. Therefore, in this paper, the historical sales data of a store on the Kaggle is utilized to make predictions based on different models, and investigate the impact of linear regression model and time series model on sales prediction. In this paper, multiple factorial linear regression model and ARIMA model in time series analysis will be adopted based on R to build prediction models to forecast sales volume, respectively. According to the analysis, the root mean square error of the prediction results under the multiple factorial linear regression model and the ARIMA model were 21.135 and 23.221, respectively. The study surface that there were differences in the degree of fit of the prediction results by different pairs of models. Therefore, the choice of the researcher needs to select the appropriate model for prediction according to the specific object of analysis, combined with the data characteristics to achieve the expected results. These results shed on guiding further exploration of differences in predictions from different models.

Reconstruction of Lake-Level Changes by Sedimentary Noise Modeling (Dongying Depression, Late Eocene, East China)

The late Eocene succession of the Dongying Depression forms a highly productive hydrocarbon source. However, due to lack of an unambiguous fine chronostratigraphic framework for the late Eocene stratigraphy, it is challenging to understand the paleolake’s evolution and the driven mechanism of lake-level variation, a limitation which hinders hydrocarbon exploration. In this work, high-resolution gamma-ray logging data were analyzed to carry out the cyclostratigraphic analysis of the third member (Es3) of the Shahejie Formation in the Dongying Depression. Significant 405-kyr eccentricity cycles were recognized based on time series analysis and statistical modeling of estimated sedimentation rates. We abstracted ~57 m cycles of the GR data in the Es3 member, which were comparable with the long eccentricity cycles (~405-kyr) of the La2004 astronomical solution, yielding a 6.43 Myr long astronomical time scale (ATS) for the whole Es3 member. The calibrated astronomical age of the third/fourth member of the Shahejie Formation boundary (41.21 Ma) was adopted as an anchor point for tuning our astrochronology, which provided an absolute ATS ranging from 34.78 ± 0.42 Ma to 41.21 ± 0.42 Ma in Es3. According to the ATS, sedimentary noise modeling for the reconstruction of lake-level changes was performed through the late Eocene Es3. The lake-level changes obtained based on sedimentary noise modeling and spectrum analysis reveal significant ~1.2 Myr cycles consistent with global sea level variations which were related to astronomical forcing. Potential driven mechanisms of marine incursion and/or groundwater table modulation were linked to explain the co-variation of global sea level changes and regional lake level changes. Our results suggest global sea level fluctuations may have played an important role in driving the hydroclimate and paleolake evolution of the late Eocene Dongying Depression.

Overwintering aggregation patterns of European catfish Silurus glanis

Animal aggregation, particularly in large-bodied species, is both a fascinating and intriguing phenomenon. Here we analyzed the overwintering behavior of the European catfish, Silurus glanis Linnaeus, 1758, the largest freshwater fish in Europe. By tracking 47 subadults and adults in a shallow lake in southeastern France, we reported a consistent aggregative behavior across four successive winters. By implementing time series analysis and Cox proportional hazard models, we investigated the dynamics of these aggregations (formation, stability, dislocation), and the factors that govern it, whether external (temperature, time of the day) or specific to the fish (size, key individuals). These aggregations lasted 1.5–2 months and mainly took place in a single small 4 m-deep area whose environmental conditions (temperature, oxygen, substrate) did not differ from other parts of the lake. In some periods during winter, all tagged fish were aggregated, which suggests that a large proportion of the lake population gathered there. Low temperatures (below 9 °C) triggered the formation of aggregations. They became more stable with decreasing temperatures, while individuals more frequently left the aggregation, preferentially at dusk and at night, when temperatures increased. The largest individuals swam more frequently back and forth to the aggregation. Irrespective of their size, some individuals consistently arrived earlier in the aggregation in winter and left later. This predictable seasonal grouping of individuals and, more generally, the knowledge provided by such studies on how species use space have important operational value and are useful for species conservation as well as for species control.

Gold Price Forecast by Different Models

In recent years, the gold investment market has a dramatic effect on the world. So being able to predict gold price accurately has become an urgent demand for economic development. This paper gives three different models including the regression analysis model, back propagation neural network (BP neural network) model, and time series analysis model to realize the prediction of gold price. Firstly, this paper verifies that the data applies to each model respectively. Secondly, through training the models, the corresponding gold price prediction formulas are output. Finally, through the test of the models, it can be concluded that the three models all have high reliability and accuracy, which means that they are suitable for the prediction of gold prices. In the end, this paper compares the prediction effects of the three models, and finds that compared with the regression analysis model, the BP neural network model and the time series analysis model have certain errors and limitations for long-term prediction, while they are more suitable for the short-term forecast. Since there are certain risks in gold investment, grasping the trend and fluctuation of the gold price accurately is the concern of both all gold investors and academic circles.

Wordle data analysis based on time series analysis model

Using LSTM time series analysis and forecasting is an important guide for Wordle's game development direction planning and economic revenue visualization. Accurate game report data prediction is of great significance for game development, economic investment, post-game planning, and improving player experience. As Wordle's game becomes more and more popular, it is essential to make predictions and projections about the future of the game as well as collate the data. In order to accurately predict the data reported by Wordle players in the future, based on the theory of time series analysis, combined with the extensive collection and screening of retrieval data, and the advantages of LSTM model and linear regression equation in the direction of prediction, a multi-dimensional prediction model for big data was established. With this prediction model, the development of Wordle games can be predicted according to a variety of prediction dimensions. After the accurate prediction of big data, the influential factors behind the data can be analyzed, which can simplify people's understanding of data to a certain extent, and successfully realize the transition from sophisticated technology to service-oriented demand.

Construction and validation of a COVID-19 pandemic trend forecast model based on Google Trends data for smell and taste loss.

To explore the role of smell and taste changes in preventing and controlling the COVID-19 pandemic, we aimed to build a forecast model for trends in COVID-19 prediction based on Google Trends data for smell and taste loss. Data on confirmed COVID-19 cases from 6 January 2020 to 26 December 2021 were collected from the World Health Organization (WHO) website. The keywords "loss of smell" and "loss of taste" were used to search the Google Trends platform. We constructed a transfer function model for multivariate time-series analysis and to forecast confirmed cases. From 6 January 2020 to 28 November 2021, a total of 99 weeks of data were analyzed. When the delay period was set from 1 to 3 weeks, the input sequence (Google Trends of loss of smell and taste data) and response sequence (number of new confirmed COVID-19 cases per week) were significantly correlated (P < 0.01). The transfer function model showed that worldwide and in India, the absolute error of the model in predicting the number of newly diagnosed COVID-19 cases in the following 3 weeks ranged from 0.08 to 3.10 (maximum value 100; the same below). In the United States, the absolute error of forecasts for the following 3 weeks ranged from 9.19 to 16.99, and the forecast effect was relatively accurate. For global data, the results showed that when the last point of the response sequence was at the midpoint of the uptrend or downtrend (25 July 2021; 21 November 2021; 23 May 2021; and 12 September 2021), the absolute error of the model forecast value for the following 4 weeks ranged from 0.15 to 5.77. When the last point of the response sequence was at the extreme point (2 May 2021; 29 August 2021; 20 June 2021; and 17 October 2021), the model could accurately forecast the trend in the number of confirmed cases after the extreme points. Our developed model could successfully predict the development trends of COVID-19. Google Trends for loss of smell and taste could be used to accurately forecast the development trend of COVID-19 cases 1-3 weeks in advance.

Stopping prehospital chlorhexidine skin wash does not increase wound morbidity after incisional hernia repair: results of a 4-year quality improvement initiative.

Prehospital chlorhexidine gluconate (CHG) skin washes are used to prevent wound complications, but little evidence supports this practice in hernia surgery. A propensity-matched retrospective review published by our group in 2016 found that prehospital CHG was associated with an increased risk of surgical site occurrences (SSO) and surgical site infections (SSI) after ventral hernia repair. Prehospital CHG was, therefore, abandoned by three of five surgeons at the Cleveland Clinic Foundation (CCF) by April 2017. We aimed to determine if discontinuation of prehospital CHG affected wound morbidity rates after incisional hernia repair. The Abdominal Core Health Quality Collaborative was queried for all patients who underwent open, clean incisional hernia repairs with 30-day follow-up from 2014 to 2019. Using an interrupted time series (ITS) analysis model adjusted for group and mean propensity score, wound morbidity before and after April 1, 2017 (start of Q2) was compared between three groups: CCF surgeons who abandoned prehospital CHG (Group 1), CCF surgeons who continued using prehospital CHG (Group 2), and non-CCF surgeons using prehospital CHG (Group 3). Outcomes included rates of SSOs, SSIs, and surgical site occurrences requiring procedural intervention (SSOPI) at 30days. In total, 4276 patients were included in the analysis (Group 1: 339 before Q2 vs 673 after Q2; Group 2: 211 before Q2 vs 175 after Q2; Group 3: 1312 before Q2 vs 1566 after Q2). Rates of SSO, SSIs, and SSOPIs at 30days were similar across all three groups before and after prehospital CHG discontinuation. Stopping prehospital CHG wash did not result in increased wound morbidity after open, clean, incisional hernia repair. We have abandoned CHG use in this context.

A data-driven lifetime prediction method for thermal stress fatigue failure of power MOSFETs

As one of the core power electronic devices that undertake power conversion and control tasks in electrical systems, power MOSFETs are widely used in key fields such as transportation, industrial drives, and aerospace. At present, the traditional method improves the reliability of power electronic devices by new material/ structure/ process, redundancy, and derating operation, which is becoming increasingly difficult to meet the requirements of rapidly developing power conversion. Based on the needs of reliability research, a data-driven lifetime prediction method for thermal stress fatigue failure of power MOSFETs is proposed. The main work is reflected in two aspects: (1) The thermal stress fatigue failure mechanism of the power MOSFETs is analyzed. On-state resistance is selected as the failure precursor parameter for evaluating the health status of power MOSFETs. (2) Autoregressive Integrated Moving Average (ARIMA) model of Time-Series Analysis is applied to realize data-driven lifetime prediction. Compared with the model-based lifetime prediction method using nonlinear regression algorithm, The data-driven method has higher prediction accuracy and better prediction stability.

Sequences of Events from the Electronic Medical Record and the Onset of Infection.

We present a novel model of time-series analysis to learn from electronic health record (EHR) data when infection occurred in the intensive care unit (ICU) by translating methods from proteomics and Bayesian statistics. Using 48,536 patients hospitalized in an ICU, we describe each hospital course as an 'alphabet' of 23 physician actions ('events') in temporal order. We analyze these as k-mers of length 3-12 events and apply a Bayesian model of (cumulative) relative risk (RR). The log2-transformed RR (median=0.248, mean=0.226) supported the conclusion that the events selected were individually associated with increased risk of infection. Selecting from all possible cutoffs of maximum gain (MG), MG>0.0244 predicts administration of antibiotics with PPV 82.0 %, NPV 44.4 %, and AUC 0.706. Our approach holds value for retrospective analysis of other clinical syndromes for which time-of-onset is critical to analysis but poorly marked in EHRs, including delirium and decompensation.

Detecting and mitigating simultaneous waves of COVID-19 infections

The sudden spread of COVID-19 infections in a region can catch its healthcare system by surprise. Can one anticipate such a spread and allow healthcare administrators to prepare for a surge a priori? We posit that the answer lies in distinguishing between two types of waves in epidemic dynamics. The first kind resembles a spatio-temporal diffusion pattern. Its gradual spread allows administrators to marshal resources to combat the epidemic. The second kind is caused by super-spreader events, which provide shocks to the disease propagation dynamics. Such shocks simultaneously affect a large geographical region and leave little time for the healthcare system to respond. We use time-series analysis and epidemiological model estimation to detect and react to such simultaneous waves using COVID-19 data from the time when the B.1.617.2 (Delta) variant of the SARS-CoV-2 virus dominated the spread. We first analyze India’s second wave from April to May 2021 that overwhelmed the Indian healthcare system. Then, we analyze data of COVID-19 infections in the United States (US) and countries with a high and low Indian diaspora. We identify the Kumbh Mela festival as the likely super-spreader event, the exogenous shock, behind India’s second wave. We show that a multi-area compartmental epidemiological model does not fit such shock-induced disease dynamics well, in contrast to its performance with diffusion-type spread. The insufficient fit to infection data can be detected in the early stages of a shock-wave propagation and can be used as an early warning sign, providing valuable time for a planned healthcare response. Our analysis of COVID-19 infections in the US reveals that simultaneous waves due to super-spreader events in one country (India) can lead to simultaneous waves in other places. The US wave in the summer of 2021 does not fit a diffusion pattern either. We postulate that international travels from India may have caused this wave. To support that hypothesis, we demonstrate that countries with a high Indian diaspora exhibit infection growth soon after India’s second wave, compared to countries with a low Indian diaspora. Based on our data analysis, we provide concrete policy recommendations at various stages of a simultaneous wave, including how to avoid it, how to detect it quickly after a potential super-spreader event occurs, and how to proactively contain its spread.

A time series analysis model of the relationship between psychoacoustic parameters of urban soundscape spatial sequences and emotional changes.

A listening test was conducted with 32 participants to obtain data on emotional changes in response to three types of urban soundscape spatial sequences. By establishing a time series model, the relationship between psychoacoustic parameters of the sequence and changes in the two dimensions of emotion was determined. Results showed that psychoacoustic parameters can explain 44% and 40%-49% of the changes in the pleasantness and arousal dimensions of emotion, respectively. Roughness and fluctuation have the highest correlation with emotional changes, while loudness and articulation index have the lowest correlation with emotional changes. This research verified the lags between psychoacoustic changes in the soundscape and the associated perceived emotion. First, there was a 3-4 s lag between psychoacoustic parameters and emotional changes. Second, changes in roughness and loudness could cause synchronous changes in emotions, while other parameters could cause delayed changes in emotions. Finally, the lag of emotion had a strong and stable explanatory power for emotional changes. This research proves the effectiveness of the time series analysis technology in establishing the dynamic relationship between the acoustic parameters of soundscape sequences and the second-by-second perceived emotions and provides a new data analysis method for in-depth study of soundscape sequence perception.

Short Term Forecast of COVID-19 cases in Japan Using Time Series Analysis Models

The new strain of coronavirus (COVID-19) was found to have started in Wuhan, China in late December 2019. The virus has spread to countries all over the world including Japan. The World Health Organization (WHO) declared COVID-19 as a pandemic on 11 March 2020 due to the increasing number of confirmed cases and deaths daily. The COVID-19 outbreak has impacted the nation of Japan adversely and the number of confirmed cases in Japan continues to increase day by day. On 7 April 2020, Japan declared a state of emergency to prevent the pandemic from worsening. This study is conducted to forecast new daily confirmed cases of COVID-19 in Japan over a short-term period. Four univariate time series models were applied: the Naïve Model, Mean Model, Autoregressive Integrated Moving Average (ARIMA) Model and Exponential State Space Model. This study analyses daily data from 22 January to 10 April 2020 collected from the Our World in Data website. The prediction involves five phases of data analysis and five different partitions of estimation and evaluation parts in every model to ensure the accuracy of forecast values. R and R Studio software were used in this study to analyze the data. The results reveal that Naïve model with 99 percent of estimation part and 1 percent evaluation part produces the lowest value of error measures for Mean Error (ME), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Mean Absolute Scaled Error (MASE).

COVID-19 Pandemic and Hospital Efficiency in Iran: Insight from an Interrupted Time Series Analysis and Pabon Lasso Model.

Limited evidence exists on the impact of the Coronavirus Disease 2019 (COVID-19) pandemic on hospital efficiency worldwide. This study aimed to investigate the impact of the COVID-19 pandemic on public hospitals in Iran. In this quasi-experimental study, monthly data on the average length of stay (ALoS), bed occupancy rate (BOR), bed turnover rate (BTR) and bed turnover interval (BTI) were collected for 58 months (36 months before and 22 months after the COVID-19 outbreak in Iran) from the health information systems of all 18 public hospitals affiliated with Kermanshah University of Medical Sciences in Kermanshah province, Iran. We used interrupted time series analysis and the Pabon Lasso model to investigate the impact of the COVID-19 pandemic on hospital efficiency. The monthly average ALoS, BOR, BTR and BTI before (after) the COVID-19 pandemic was 3.30 (3.48) days, 70.14% (49.37), 6.78 (4.81) patients per bed and 1.15 (2.88) days, respectively. The study indicated that a statistically significant decrease in ALoS of 0.29 and BOR of 25.09 in the first month following the COVID-19 pandemic. Compared with the before pandemic period, we observed a significant increase in the monthly trend of ALoS (coefficient = 0.021; p = 0.015), BOR (coefficient = 1.30; p = 0.002), and BTR (coefficient = 0.08, p = 0.012). We found a significant decrease in the monthly trend in BTI (coefficient = -0.11, p = 0.009) after the COVID-19 pandemic when compared with before the pandemic. Based on the Pabon Lasso model, before (after) the pandemic, 29.4% (29.4%) of the hospitals were located in zone 1 as an inefficient area, and 17.6% (35.3%) of hospitals were located in zone 3 as an efficient area. The study demonstrated that the BOR and BTR decreased substantially after the outbreak of COVID-19. In contrast, the ALoS and BTI have significantly increased following the COVID-19 pandemic. We also found that hospitals' performance in both periods was poor, and only 30% of hospitals were located in the efficient zone (zone three) based on the Pabon Lasso model. Further studies aimed at identifying the main factors affecting lower efficiency among hospitals in Iran are recommended.

Forecasting time trends of fatal motor vehicle crashes in Iran using an ensemble learning algorithm

Objective This study aimed to introduce the random forest (RF) method as a valuable tool for short-term crash frequency prediction. Besides, the study compares the forecast efficiency of the RF model with the classical seasonal autoregressive integrated moving average (SARIMA) model in the multivariate time-series analysis of crash counts. Methods To this end, fatal accidents reported by the police and intercity traffic flow extracted from the loop detectors were aggregated in intercity highways at the country’s level monthly from Farvardin 1395 to Mordad 1400. The first 55 data points were used as the training sample, and the remaining ten months were considered the test sample. The Box-Jenkins and random forest machine learning methods were employed for short-term crash frequency prediction. The mean absolute percentage error (MAPE) criterion was utilized to compare the forecast accuracy of the developed models. Results The performance of the random forest model (MAPE = 2.6) with the exogenous variables of traffic flow, crash year, and month outperformed the best SARIMA ( 1 , 0 , 0 ) ( 1 , 0 , 0 ) 12 model (MAPE = 5.7) with traffic flow as the regressor. Conclusions This study suggests that the random forest as an ensemble learning algorithm is a better crash prediction tool compared to the classical Box-Jenkins method, accounting for the non-linear dependencies in crash count time-series. Besides, the results illustrate that the multivariate SARIMA (SARIMAX) model significantly outperforms its univariate counterpart, accounting for the simultaneous impacts of exogenous variables.