A Comparative Analysis of Models for Forecasting Thailand’s Unemployment Rate, 2021-2024

Abstract

This study aims to compare the performance of unemployment rate forecasting models, including the Autoregressive Integrated Moving Average (ARIMA), Holt’s Winters,          K-Nearest Neighbors (K-NN), and Linear Regression models. The analysis is based on Thailand’s unemployment rate data obtained from the Department of Employment, consisting of provincial-level data from 77 provinces covering the period from 2021 to 2024. The objective is to evaluate the performance of each model and compare their forecasting accuracy. The results indicate that the ARIMA model yields the lowest Mean Absolute Error (MAE) of 6,658.515 and the lowest Root Mean Square Error (RMSE) of 8,578.801, reflecting lower numerical forecasting errors and greater stability. Although Holt’s Winters model achieves the lowest Mean Absolute Percentage Error (MAPE) at 4%, indicating high relative accuracy, its RMSE and MAE values remain higher than those of the ARIMA model. The K-Nearest Neighbors (K-NN) model shows a moderate MAPE of 6% but exhibits a very high RMSE of 62,036.073, suggesting instability in its forecasting results. Meanwhile, the Linear Regression model records a MAPE of 0%, but its RMSE and MAE values are abnormally high, indicating an overfitting problem and poor suitability for practical applications. Therefore, the findings conclude that the ARIMA model is the most appropriate approach for forecasting Thailand’s unemployment rate, as it provides the best balance between forecasting accuracy and numerical error.