A temporal-window framework for modelling and forecasting time series

Neto, Paulo S. G. de Mattos; Cavalcanti, George D. C.; Firmino, Paulo Renato Alves; Silva, Eraylson G.; Filho, Sérgio R. P. Vila Nova

doi:10.1016/j.knosys.2020.105476

Cited by 8 publications

(10 citation statements)

References 41 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…SEIR, SIR, and SIS), and usual time series (i.e. ARIMA and exponential smoothing - ETS) models: MSE; Mean Absolute Percentage Error (MAPE); Average Relative Variance (ARV); Index of Disagreement (ID); Theil’s U (Theil); Wrong Prediction on Change of Direction (WPOCID); Intercept of the linear fit between

and u t (Reg_Intercept); Slope Coefficient of the linear fit between

and u t (Reg_Slope); Indeterminacy Coefficient of the linear fit between

and u t (WR2) [41] , [53] ; and an Aggregate Performance Metric (APM). See Eqs (4) , (6) - (13) .…”

Section: Covid-19 Experimentsmentioning

confidence: 99%

“…In turn, WR 2 =

as well as Reg_Intercept and Reg_Slope are related to the linear model adjusted to the pairs u t e

via minimal squared estimation. In this way, one can consider the general equation

[41] . Thus, Reg_Intercept and Reg_Slope coefficients represent the additive and multiplicative errors of the forecasts

of u t , respectively.…”

Section: Covid-19 Experimentsmentioning

confidence: 99%

“…Thus, Reg_Intercept and Reg_Slope coefficients represent the additive and multiplicative errors of the forecasts

of u t , respectively. In this case, there is a constant error Reg_Intercept, independent from the forecast, and a proportional error Reg_Slope related to the prediction [41] . In turn, R 2 , the determination coefficient, reflects the performance of the model in capturing the variability of the time series [41] .…”

Section: Covid-19 Experimentsmentioning

confidence: 99%

“…In this case, there is a constant error Reg_Intercept, independent from the forecast, and a proportional error Reg_Slope related to the prediction [41] . In turn, R 2 , the determination coefficient, reflects the performance of the model in capturing the variability of the time series [41] . R 2 is defined as

in which

is the average of the observed series.…”

Section: Covid-19 Experimentsmentioning

confidence: 99%

“…Specially for the cases in Phase (i), the prediction of the periods in which the virus incidence time series will achieve Phases (ii) and (iii) is a hard task. Common time series formalisms, like ARIMA, ETS, support vector regression, and artificial neural networks [41] are not usually able to predict the plateau-decreasing phases when trained by data sets belonging to Phase (i), only. Further, these approaches suffer when performing relatively middle-long-step-ahead predictions in the light of small sized training data sets.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

A non-central beta model to forecast and evaluate pandemics time series

Firmino

Sales

Júnior³

et al. 2020

Chaos, Solitons & Fractals

View full text Add to dashboard Cite

Government, researchers, and health professionals have been challenged to model, forecast, and evaluate pandemics time series (e.g. new coronavirus SARS-CoV-2, COVID-19). The main difficulty is the level of novelty imposed by these phenomena. Information from previous epidemics is only partially relevant. Further, the spread is local-dependent, reflecting a number of social, political, economic, and environmental dynamic factors. The present paper aims to provide a relatively simple way to model, forecast, and evaluate the time incidence of a pandemic. The proposed framework makes use of the non-central beta (NCB) probability density function. Specifically, a probabilistic optimisation algorithm searches for the best NCB model of the pandemic, according to the mean square error metric. The resulting model allows one to infer, among others, the general peak date, the ending date, and the total number of cases as well as to compare the level of difficult imposed by the pandemic among territories. Case studies involving COVID-19 incidence time series from countries around the world suggest the usefulness of the proposed framework in comparison with some of the main epidemic models from the literature (e.g. SIR, SIS, SEIR) and established time series formalisms (e.g. exponential smoothing - ETS, autoregressive integrated moving average - ARIMA).

show abstract

and u t (Reg_Intercept); Slope Coefficient of the linear fit between

and u t (Reg_Slope); Indeterminacy Coefficient of the linear fit between

and u t (WR2) [41] , [53] ; and an Aggregate Performance Metric (APM). See Eqs (4) , (6) - (13) .…”

Section: Covid-19 Experimentsmentioning

confidence: 99%

“…In turn, WR 2 =

as well as Reg_Intercept and Reg_Slope are related to the linear model adjusted to the pairs u t e

via minimal squared estimation. In this way, one can consider the general equation

[41] . Thus, Reg_Intercept and Reg_Slope coefficients represent the additive and multiplicative errors of the forecasts

of u t , respectively.…”

Section: Covid-19 Experimentsmentioning