Scientific African 21 (2023) e01739

Contents lists available at ScienceDirect

Scientific African
journal homepage: www.elsevier.com/locate/sciaf

Comparison of random forest and support vector machine

regression models for forecasting road accidents
Antoine Gatera a,∗, Martin Kuradusenge b, Gaurav Bajpai c, Chomora Mikeka d,
Sarika Shrivastava e
a
Department of Education in Sciences, University of Technology and Arts of Byumba (UTAB), P.O. Box 25, Byumba, Rwanda
b
School of ICT, College of Science and Technology, University of Rwanda, P.O. Box 3900, Kigali, Rwanda
c
Kampala International University, P.O. Box 20000, Kampala, Uganda
d
Directorate of Science, Technology and Innovation, Ministry of Education, Lilongwe P/Bag 328, Malawi
e
Ashoka Institute of Technology & Management, Varanasi 221007 U.P., India

a r t i c l e i n f o a b s t r a c t

Article history: One of the world’s concerns today is the rate of road traffic accidents (RTA). The over-
Received 8 November 2021 whelming majority of these accidents occur in low and middle-income countries. RTA are
Revised 27 May 2023
one of the leading causes of death in Rwanda. RTA prediction is crucial for both trans-
Accepted 27 May 2023
portation management and intelligent transportation systems (ITS) development. This pa-
per adopted the use of two modelling techniques, Random forest (RF) and Support vector
Editor: DR B Gyampoh machine (SVM) for short-term road accident forecasting. The data used to evaluate the
models was obtained from the Police. The lower error indices of MAE, MSE, RMSE and
Keywords:
higher coefficient of determination (R2 ) were accuracy measures in comparing the models.
Road accidents
The RF model performed better than the SVM model as it revealed higher R2 = 0.91 com-
Random forest
Support vector machine pared to the SVM model that was with R2 = 0.86. Machine learning methods are promising
tools for the prediction of road accidents. Prediction positively influences safety enhance-
ments and regulation formulation to prevent future accidents. The appropriate prediction
method would help policymakers and healthcare providers adjust their contributions to
the accident management process.
© 2023 Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)

Introduction

Annually, approximately 1.3 million people died from traffic accidents and more than 50 million people sustain non-fatal
injuries worldwide. The low-and-middle income (LMI) countries account for 93% of all fatal RTA. In LMI, the average death
rate per 10 0,0 0 0 population is 27.5 while in the high-income (HI) countries, the average rate is 8.3 deaths per 10 0,0 0 0
population [1]. The African region has a casualty rate of 26.6 per 10 0,0 0 0 individuals, whereas the European region has 9.3
per 10 0,0 0 0 individuals. In Africa, RTAs constitute 25% of all injury-related deaths. In order to address this public health is-
sue, the United Nations (UN) has proposed global goals for sustainable development with which the aim was to have halved
road accidents that result from RTAs by 2020. Moreover, a decade of activity for street safety initiatives to extend awareness

∗
Corresponding author.
E-mail address: [email protected] (A. Gatera).

https://doi.org/10.1016/j.sciaf.2023.e01739
2468-2276/© 2023 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
A. Gatera, M. Kuradusenge, G. Bajpai et al. Scientific African 21 (2023) e01739

over diverse nations was proposed with the objective to reduce at least half of the total fatalities and injuries from crashes
on the roads by 2030. Effective interventions incorporate the development of safer structures and the integration of road
safety features into road use and transportation planning, the enhancement of vehicle safety attributes, the enhancement
of post-crash care for road crash victims, the enforcement and implementation of key risk laws, and the encouragement of
public awareness are encouraged to reduce the extent of road accidents [2]. Road safety is a major concern for LMI coun-
tries, as it significantly costs their economies. However, the social costs of road crashes in LMI countries are presented to be
underreported, every year road accidents on average costs the countries between 3 and 5% of their gross domestic product
(GDP) [3,4].
The East African Community (EAC) is a regional intergovernmental organization consisting of seven member states.
Though road fatalities keep increasing in the region, some countries are taking measures such as setting maximum speed
limits to reduce the road fatality rate [5]. Rwanda is a member, the country is addressing RTA issues by implementing dif-
ferent measures such as the utilization of speed limits by both public buses and heavy vehicles and imposing road traffic
cameras on the roads countrywide to decrease the accidents rate. According to the statistics, the total accidents accounted
for 22,382 cases, where fatal accidents represented 11%, serious 14%, minor 32% and property damage 44% of the total ac-
cidents which occurred between 2016 and 019. Given these challenges, road accident reduction is an urgent issue to be
addressed by governments.
Forecasting the number of accidents is necessary for countries to develop safety systems and developing measures. Re-
gression models were applied to predict accidents. The models study the relationship between the dependent variable and
the independent variables [6,7]. However, ML models are the most preferred as they do not include drawbacks, such as
assuming linearity or correlations between exploratory factors and the target variable [8,9]. RF being one of the ML algo-
rithms can handle nonlinear and high-dimensional variables [10]. RF has been used to classify and predict occupational
accident types [11]. Four popular data-driven traffic forecasting methods, K-Nearest Neighbor, Decision Tree, RF, and SVM
were compared to determine the best prediction method based on the predictive performance [12]. Traditional statistical
methods have been employed in the majority of accident forecast studies. Time-series-based methods have been applied for
accident prediction [13,14]. Autoregressive Integrated Moving Average (ARIMA) model has been gradually used to study the
trends and patterns of a RTA by implementing the knowledge of autocorrelation analysis based on autoregressive integrated
moving average models [15,16]. Statistical and ML methods for crash severity are the most used prediction tools [12,17].
Related works have focused on real-time accident prediction based on causal factors such as speeding, leakless drive and
others [18].
Various studies have used the RF model as a predictive tool. RF is useful for a variety of datasets involving regression
and classification problems [19]. RF is used to rank variables based on their relative importance, assisting in the selection of
the most appropriate independent variables prior to applying other statistical models. RF model is an ensemble method for
classification in the field of ML model [20,21]. In particular, the RF has been reported to be an effective prediction tool in the
prediction of accidents, as it dramatically improves the prediction accuracy [22–24]. In agriculture, it was used to predict soil
organic carbon [25] and in mining RF were used together with the SVM in predicting coal temperature [26,27]. RF can limit
overfitting without affecting prediction accuracy [28], and the SVM is highly effective in solving non-linear problems. Traffic
accidents cause many deaths, injuries, fatalities and economic losses. This makes them the world’s most pressing issue.
Accurate models for predicting the severity of traffic accidents are critical for transportation systems. Analytic methods have
been utilized by researchers to characterize the information and variables of a traffic accident in order to reveal hidden
patterns, profile behaviours, generate rules and inferences. These patterns can be used to profile drivers’ behaviour on the
road, dangerous driving areas, generate classification rules using road accident data, perform variable selection in real-time
accident models, and select relevant variables to train other methods.
Various methods proposed have focused on causal factors modelling and accident severity prediction. Prediction of ac-
cident severity based on the observed crash change and growth over time-based on accidents contributing factors. There is
an increasing focus in studies that have focused on reducing the frequency and severity of traffic crashes. Many previous
studies looked into the impact of accident-causing factors on crash severity. In this paper, we compared the results of the
selected ML models using the measure of statistical significance, that indicates the extent to which variation in a dependent
variable can be explained by a factor that is independent in a regression model. Fatal, serious, minor and property damages
are considered to be the predictors. The remainder of the paper is organized as detailed below: Section 2 introduces the
description of the user database, and preliminaries of the models and discusses the model evaluation metrics. Section 3
covers the performance and comparison of proposed models. Section 4 the conclusions.

Material and method

Data

In order to evaluate the models, we have used the data from accident-recorded cases that occurred from 2016 to 2019.
The average number of accidents in a typical year was 5595 cases. Statistical data from a secondary source was used for
forecasting road accidents. The traffic department collects road traffic data related to traffic accidents, injuries and prop-
erty damage cases. They gather crash information from the scene, vehicle details, vehicle conditions, interviewing the crash
victims and non-participating observers. The recorded data contain information on road accident causal factors and crash

2
A. Gatera, M. Kuradusenge, G. Bajpai et al. Scientific African 21 (2023) e01739

Table 1
Descriptive statistics of road accidents.

Category Min Max Mean SD

Fatal 549 673 603 51

Serious 629 911 789 133
Minor 1485 1896 1765 193
Property damage 1584 3347 2437 733

Fig. 1. Number of road accident cases in Rwanda by category (2016–2019).

Fig. 2. Monthly variation of accidents over the period (2016–2019).

severity. Road crashes are categorized into four categories: fatal, serious, minor, and property damage accidents. The Mann-
Kendall technique was applied to the data to analyse accident trends [29,30]. The cases of accidents were revealed to be
decreasing in 2018–2019. Table 1 depicts the descriptive statistics of accidents on average per year and by category. Descrip-
tive statistics and models were implemented in Python. Figure 1 depicts the variation of road accidents for the past four
years by type. The time series data collected were converted to comma-separated values (CSV) format for use as input data
in SVM and RF models.
Figure 2 reveals the variation of accidents by category over the past years. The statistics show that, although different
measures have been taken to reduce the total number of accidents, only minor and property damage accidents have been
reduced proportionally to total cases. On the other hand, fatal and serious injury accident cases showed an additive trend.

3
A. Gatera, M. Kuradusenge, G. Bajpai et al. Scientific African 21 (2023) e01739

Prediction methods

In this paper, two prediction methods, support vector machines and random forest regression for classification and pre-
diction were used for short-term road accident forecasting.

Support vector machine (SVM) for regression

The SVM, which is a supervised machine learning algorithm was introduced by Vapnik in 1995 to solve classification and
regression problems [31]. The SVM is one of the supervised learning models that investigate data and identifies data samples
used for classification. An SVM training algorithm constructs a model that labels a new data set into one class. Assuming
that road accident prediction is a regression problem X, fatal accidents (x1 ), severe accidents (x2 ), minor accidents (x3 ) and
property damage accidents (x4 ) are considered as input (predictors) for SVM. This was done in order to understand how
the response X depends simultaneously on the predictors in our context assumed to be (x1 ), (x2 ), (x3 ) and (x4 ). Given the
training data consisting of the input matrix X = [x1 , x2 , . . . , xn ] and an output vector Y = [y1 , y2 , . . . , yn ], the SVM construct
an optimized linear regression through mapping the input vector x. SVM in machine learning approaches includes a set of
learning methods and shows better results [32,33]. SVM performs linear and nonlinear classification. SVM supports linear
and nonlinear regression applications [34].

Random forest (RF) model

RF is a prediction method categorized as ensemble learning that integrates multiple decision trees. It is one of the data
mining tools used in the ML framework. It can be used for both classification, regression problems and time series forecast-
ing [10,35]. The RF classifier is a collection of decision trees derived from the random training set. It combines the votes
from various decision trees to determine the final class of the test object. It has a series of decision paths from the node
to the last leaf safeguarded by a sub-feature. Prediction is a sum of individual features and the mean value of the top most
region covered by the training set. Given a training dataset A = (xi , yi ), i = 1, 2, . . . , n, (X, Y ) ∈ Rm × R, the input matrix X
consists of n samples with m features, and the output Y is a target vector. The RF adopts the bootstrap resampling method
to form N tree sample sets (Sk , k = 1, 2, . . . , N tree) randomly from the original sample set S, the number of elements of
Sk is the same as that of S (where k is the current number of iterations). In bootstrap samples, approximately one-third of
the data in original sample set S which are called out-of-bag (OOB) data are not drawn, and the remaining data are called
in-bag data [36]. Random survival forest (RSF) being an extension of RF, to regular random forests independent bootstrap
sampling was used to handle multicollinearity while examining crash severity at public highway-rail [37].
Therefore, the number of trees in the forest and the number of variables utilized to develop each tree are the two
primary characteristics that influence the RF model’s capacity to estimate. The model’s mean square error calculation is cal-
culated by the OOB and this is the method for measuring the prediction error and computing the variable importance [38].
Equation (1) is used to calculate the error:

1
n
MSEOOB = (Oi − PiOOB )2 (1)
n
i=1

where n is the observation number and PiOOB is the average of the OOB’s predictions across all the trees.

Dataset size
In this paper, time series data spanning from January 2016 to December 2018 (36 months) of reported accidents
were selected to develop the SVM and RF models when training the models. The training dataset comprised approxi-
mately 70% (17,729 cases) and the testing dataset (January 2019 to December 2019) consists of 30% (4653) of the total
recorded cases (22,382 cases).

Model validation

The performances of the models were assessed by using the usual metrics for model performance, mean absolute er-
ror (MAE), mean squared error (MSE), root mean squared error (RMSE) and coefficient of determination (R2 ) [39–41]. The
MAE, MSE, RMSE, and the R2 were the indicators used to determine the performance of the SVM and RF models [42,43]. The
accident modelling scenario used in this paper is considered as a regression problem, which is a set of statistical processes
for estimating the relationships between the response variable (total accidents) and predictor variables referred to as fatal
accidents, minor accidents, severe accidents, and property damage accidents. The MAE given by the Eq. (2) is a risk metric
corresponding to the expected value of the absolute error.

1   
n
MAE = yi − yˆi  (2)
n
i=1

The MSE is the average of a set of errors and it is given by the Eq. (3).

1  2
n
MSE = yi − yˆi (3)
n
i=1

4
A. Gatera, M. Kuradusenge, G. Bajpai et al. Scientific African 21 (2023) e01739

Fig. 3. Pearson correlation.

Table 2
Prediction model results.

Model MAE MSE RMSE R2

Testing data
Random forest 11.925 268.386 16.382 0.918
Support vector machine 14.175 440.268 20.982 0.866
Training data
Random forest 6.119 67.682 8.226 0.983
Support vector machine 0.001 1.240 0.001 0.999

The RMSE which is the standard deviation of the residuals (prediction errors) is defined by Eq. (4).
  
n
1 yi − yˆi
RMSE = (4)
n
The R2 is a statistical metric used to measure how much of the outcome to be expected (accidents). The R2 values range from
zero to one [0, 1]. Zero (0) illustrates that the accidents can not be predicted based on the historically recorded accidents
cases, while One (1) implies the perfect prediction and is given by Eq. (5).
n  2
2 i=1 yi − yˆi
R =1− n (5)
i=1 (yi − ȳi )2
where yˆi is the predicted value of the ith sample and yi is the corresponding true value for the total n sample.

Results and discussion

Collinearity of the data

In the assessment of the correlation between the four categories assumed to be the variables and they were used in the
prediction. The collinearity of categories was computed to check how they are associated with each other and also with
the total number of accidents. Figure 3 shows the correlation of the variables. Property damage and minor accident cases
presented a higher correlation whereas fatal and serious injuries presented no significant correlation.

Comparative analysis of the models

The two models were estimated using modelling training and validation datasets to predict road traffic accidents. To
verify the prediction capability of the proposed models, the evaluation of their results is shown in Table 2.
The accuracy of the RF model had a significant improvement over the SVR model. Compared to the SVM model, it has
the highest R2 and lowest MSE, RMSE, and MAE compared to the SVM model. The average prediction error is the sum of
all test case differences between the actual and predicted values. A low MAE indicates that the prediction is appropriate.
The square root of the MSE is used to calculate the RMSE, which is a measure of the difference between the predicted and
the real numbers. A lower value for the RMSE signifies greater forecasting precision. The RF-based prediction model showed

5
A. Gatera, M. Kuradusenge, G. Bajpai et al. Scientific African 21 (2023) e01739

the highest prediction accuracy with an R2 = 0.918, MAE = 11.925, MSE = 268.386 and RMSE = 16.382, compared to SVM,
which has an R2 of 0.866, MAE = 14.175, MSE = 440.268 and RMSE = 20.982.

Conclusion

The paper details the feasibility of using ML prediction models for accident prediction. The predictive models of ML
algorithms SVR and RF, were applied to secondary data recorded by police for short-term accident prediction. A comparison
between the metrics mainly the R2 were used for assessing the models’ predictive performance. The results showed that
the RF outperformed the SVM with the highest performance in terms of R2 and lower error rates. The findings show that
the RF model has outstanding prediction performance. Understanding how traffic crashes fluctuate and expand over time
is essential to devise efficient prevention strategies. The research of algorithms and data analysis techniques to address
the challenge of developing a secured mobility environment is a field of current scientific knowledge open to innovation.
Accidents prediction could be one of the proactive methodologies to help road authorities in planning for safety before
accidents and problems arise. Appropriate policies are required in every sector to reduce the increase in accidents. Accident
prediction can be used as a potential measure to reduce traffic accidents. Time series analysis is one of the most effective
methods for forecasting future events and demonstrating the significance of certain variables. There is an opportunity to
extend it by applying more regression models.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

CRediT authorship contribution statement

Antoine Gatera: Conceptualization, Formal analysis, Methodology, Software, Writing – original draft, Writing – review
& editing. Martin Kuradusenge: Formal analysis, Methodology, Software. Gaurav Bajpai: Supervision, Validation, Writing
– review & editing. Chomora Mikeka: Supervision, Validation, Writing – review & editing. Sarika Shrivastava: Validation,
Writing – review & editing.

References

[1] WHO, Global Status Report on Road Safety 2018, Technical Report, World Health Organization, 2018.
[2] WHO, Global Launch: Decade of Action for Road Safety 2011–2020, Technical Report, World Health Organization, 2011.
[3] S. Chen, M. Kuhn, K. Prettner, D.E. Bloom, The global macroeconomic burden of road injuries: estimates and projections for 166 countries, Lancet
Planet. Health 3 (9) (2019) e390–e398.
[4] W. Wijnen, H. Stipdonk, Social costs of road crashes: an international analysis, Accid. Anal. Prev. 94 (2016) 97–106.
[5] WHO, The global Health Observatory: Road Traffic Deaths, Technical Report, 2021.
[6] G. Prati, L. Pietrantoni, F. Fraboni, Using data mining techniques to predict the severity of bicycle crashes, Accid. Anal. Prev. 101 (2017) 44–54.
[7] B.K. Mohanta, D. Jena, N. Mohapatra, S. Ramasubbareddy, B.S. Rawal, Machine learning based accident prediction in secure IoT enable transportation
system, J. Intell. Fuzzy Syst. 42 (2) (2022) 713–725.
[8] L.-Y. Chang, W.-C. Chen, Data mining of tree-based models to analyze freeway accident frequency, J. Saf. Res. 36 (4) (2005) 365–375.
[9] S. Alagarsamy, M. Malathi, M. Manonmani, T. Sanathani, A.S. Kumar, Prediction of road accidents using machine learning technique, in: 2021 5th
International Conference on Electronics, Communication and Aerospace Technology (ICECA), IEEE, 2021, pp. 1695–1701.
[10] L. Breiman, Random forests, Mach. Learn. 45 (1) (2001) 5–32.
[11] K. Kang, H. Ryu, Predicting types of occupational accidents at construction sites in Korea using random forest model, Saf. Sci. 120 (2019) 226–236.
[12] J. Zhang, Z. Li, Z. Pu, C. Xu, Comparing prediction performance for crash injury severity among various machine learning and statistical methods, IEEE
Access 6 (2018) 60 079–60 087.
[13] S. Jomnonkwao, S. Uttra, V. Ratanavaraha, Forecasting road traffic deaths in thailand: applications of time-series, curve estimation, multiple linear
regression, and path analysis models, Sustainability 12 (1) (2020) 395.
[14] F. Zolala, A.A. Haghdoost, T. Ahmadijouybari, A. Salari, A. Bahrampour, M.R. Baneshi, A. Razzaghi, Forecasting the trend of traffic accident mortality in
West Iran, Health Scope 5 (3) (2016) 1–5.
[15] C.C. Ihueze, U.O. Onwurah, Road traffic accidents prediction modelling: an analysis of Anambra state, Nigeria, Accid. Anal. Prev. 112 (2018) 21–29.
[16] N. Igissinov, A. Aubakirova, G. Orazova, G. Akpolatova, S. Urazova, D. Tarzhanova, A. Zhantureyeva, Y. Kuandykov, Prediction mortality rate due to the
road-traffic accidents in Kazakhstan, Iran. J. Public Health 49 (1) (2020) 68.
[17] A. Iranitalab, A. Khattak, Comparison of four statistical and machine learning methods for crash severity prediction, Accid. Anal. Prev. 108 (2017)
27–36.
[18] A. Theofilatos, C. Chen, C. Antoniou, Comparing machine learning and deep learning methods for real-time crash prediction, Transp. Res. Rec. 2673 (8)
(2019) 169–178.
[19] A. Müller, Introduction to Machine Learning with Python, O’reilly Media, 2016.
[20] T. Prajwala, A comparative study on decision tree and random forest using R tool, Int. J. Adv. Res. Comput. Commun. Eng. 4 (1) (2015) 196–199.
[21] T.K. Ho, The random subspace method for constructing decision forests, IEEE Trans. Pattern Anal. Mach. Intell. 20 (8) (1998) 832–844.
[22] J. Choi, B. Gu, S. Chin, J.-S. Lee, Machine learning predictive model based on national data for fatal accidents of construction workers, Autom. Constr.
110 (2020) 102974.
[23] X. Zhou, P. Lu, Z. Zheng, D. Tolliver, A. Keramati, Accident prediction accuracy assessment for highway-rail grade crossings using random forest algo-
rithm compared with decision tree, Reliab. Eng. Syst. Saf. 200 (2020) 106931.
[24] L. Wahab, H. Jiang, A comparative study on machine learning based algorithms for prediction of motorcycle crash severity, PLoS One 14 (4) (2019)
e0214966.
[25] H. Zhang, P. Wu, A. Yin, X. Yang, M. Zhang, C. Gao, Prediction of soil organic carbon in an intensively managed reclamation zone of eastern China: a
comparison of multiple linear regressions and the random forest model, Sci. Total Environ. 592 (2017) 704–713.

6
A. Gatera, M. Kuradusenge, G. Bajpai et al. Scientific African 21 (2023) e01739

[26] C. Lei, J. Deng, K. Cao, Y. Xiao, L. Ma, W. Wang, T. Ma, C. Shu, A comparison of random forest and support vector machine approaches to predict coal
spontaneous combustion in gob, Fuel 239 (2019) 297–311.
[27] Z. Li, P. Liu, W. Wang, C. Xu, Using support vector machine models for crash injury severity analysis, Accid. Anal. Prev. 45 (2012) 478–486.
[28] D. Wielenga, Identifying and overcoming common data mining mistakes, in: SAS Global Forum, 2007, pp. 1–20.
[29] Y. Li, Analysis and forecast of global civil aviation accidents for the period 1942–2016, Math. Probl. Eng. (2019) 1–12.
[30] G.M. Ngwira, B. Bolaane, B.P. Parida, Investigating the trend of road traffic fatalities in Malawi using Mann–Kendall statistic, Heliyon 9 (2) (2023)
e13700.
[31] C. Cortes, V. Vapnik, Support-vector networks, Mach. Learn. 20 (3) (1995) 273–297.
[32] V. Kecman, Support vector machines–an introduction, in: Support Vector Machines: Theory and Applications, Springer, 2005, pp. 1–47.
[33] Y. Lin, R. Li, Real-time traffic accidents post-impact prediction: based on crowdsourcing data, Accid. Anal. Prev. 145 (2020) 105696.
[34] V. Vapnik, The Nature of Statistical Learning Theory, Springer Science & Business Media, 2013.
[35] N. Dogru, A. Subasi, Traffic accident detection using random forest classifier, in: 2018 15th Learning and Technology Conference (L&T), 2018, pp. 40–45.
[36] I.A. Ibrahim, T. Khatib, A novel hybrid model for hourly global solar radiation prediction using random forests technique and firefly algorithm, Energy
Convers. Manag. 138 (2017) 413–425.
[37] A. Keramati, P. Lu, A. Iranitalab, D. Pan, Y. Huang, A crash severity analysis at highway-rail grade crossings: the random survival forest method, Accid.
Anal. Prev. 144 (2020) 105683.
[38] S. Janitza, G. Tutz, A.-L. Boulesteix, Random forest for ordinal responses: prediction and variable selection, Comput. Stat. Data Anal. 96 (2016) 57–73.
[39] C. Van Hinsbergen, J. Van Lint, F. Sanders, Short term traffic prediction models, in: Proceedings of the 14th World Congress on Intelligent Transporta-
tion Systems (ITS), Beijing, October 2007, 2007.
[40] J. Van Lint, C. Van Hinsbergen, Short-term traffic and travel time prediction models, Artif. Intell. Appl. Crit. Transp. Issues 22 (1) (2012) 22–41.
[41] E.I. Vlahogianni, M.G. Karlaftis, J.C. Golias, Short-term traffic forecasting: where we are and where we’re going, Transp. Res. Part C 43 (2014) 3–19.
[42] D. Dancer, A. Tremayne, R-squared and prediction in regression with ordered quantitative response, J. Appl. Stat. 32 (5) (2005) 483–493.
[43] T. Chai, R.R. Draxler, Root mean square error (RMSE) or mean absolute error (mae)?—Arguments against avoiding RMSE in the literature, Geosci. Model
Dev. 7 (3) (2014) 1247–1250.