Comparative Analysis of Different Univariate Forecasting Methods in Modelling and Predicting the Romanian Unemployment Rate for the Period 2021-2022

Abstract

Unemployment has risen as the economy has shrunk. The coronavirus crisis has affected many sectors in Romania, some companies diminishing or even ceasing their activity. Making forecasts of the unemployment rate has a fundamental impact and importance on future social policy strategies. The aim of the paper is to comparatively analyze the forecast performances of different univariate time series methods with the purpose of providing future predictions of unemployment rate. In order to do that, several forecasting models (seasonal model autoregressive integrated moving average (SARIMA), self-exciting threshold autoregressive (SETAR), Holt-Winters, ETS (error, trend, seasonal), and NNAR (neural network autoregression)) have been applied, and their forecast performances have been evaluated on both the in-sample data covering the period January 2000-December 2017 used for the model identification and estimation and the out-of-sample data covering the last three years, 2018-2020. The forecast of unemployment rate relies on the next two years, 2021-2022. Based on the in-sample forecast assessment of different methods, the forecast measures root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percent error (MAPE) suggested that the multiplicative Holt-Winters model outperforms the other models. For the out-of-sample forecasting performance of models, RMSE and MAE values revealed that the NNAR model has better forecasting performance, while according to MAPE, the SARIMA model registers higher forecast accuracy. The empirical results of the Diebold-Mariano test at one forecast horizon for out-of-sample methods revealed differences in the forecasting performance between SARIMA and NNAR, of which the best model of modeling and forecasting unemployment rate was considered to be the NNAR model.

Keywords: ETS; Holt–Winters; Romania; SARIMA; SETAR; neural network autoregression; unemployment rate.

Figure 1

A diagrammatic representation of the NNAR(p,P,k)m model. Source: Touplan [69]. NNAR: neural network autoregression.

Figure 2

The seasonal pattern in the monthly ILO unemployment rate.

Figure 3

Autocorrelation and partial correlation plot of Romania’s monthly unemployment rate for the horizon 2000–2020.

Figure 4

The forecast of unemployment rate based on Holt–Winters (HW) models for the period 2021–2022.

Figure 5

The forecast of unemployment rate based on the results of ETS(M,N,M).

Figure 6

Autocorrelation and partial correlation plot of Romania’s monthly unemployment rate.

Figure 7

Forecasts from a neural network with one seasonal and non-seasonal lagged input and one hidden layer containing ten neurons.

Figure 8

Descriptive statistics of unemployment rate for the horizon 2000–2017.

Figure 9

Autocorrelation and partial correlation plot of Romania’s monthly unemployment rate (a) and first difference of the original time series (b).

Figure 10

The Romanian ILO unemployment rate for the period 2000M1–2020M12.

Figure 11

Autocorrelation and partial correlation plot of the first difference of the unemployment rate.

Figure 12

Diagnostic plot of SARIMA(0,1,6)(1,0,1)_12.

Figure 13

Forecasts of unemployment rate based on the results of ARIMA(0,1,6)(1,0,1)_12.

Figure 14

Partial autocorrelation plot of unemployment series.

Figure 15

Grid search method estimation of one threshold value.

Figure 16

Forecasts of unemployment rate based on the results of the SETAR(2,13,1) model.

Figure 17

Forecast combination of the Romanian unemployment rate.

Figure 18

The forecasts of unemployment rate for the period 2021–2022.