A Model Selection Approach for Time Series Forecasting: Incorporating Google Trends Data in Australian Macro...

Area of Science:

• Economics
• Data Science
• Computational Social Science

Background:

• Accurate forecasting of macroeconomic indicators is crucial for economic policy and planning.
• Traditional time series models may not fully capture complex real-world dynamics influencing economic variables.

Purpose of the Study:

• To investigate the utility of Google Trends data for forecasting Australian macroeconomic indicators.
• To compare the forecasting performance of traditional (SARIMA), machine learning (SVR), and deep learning (CNN) models.
• To assess the impact of forecasting horizon and data-driven feature selection on model accuracy.

Main Methods:

• Utilized Google Trends data alongside historical macroeconomic data for Australia.
• Implemented and compared Seasonal Autoregressive Integrated Moving Average (SARIMA), Support Vector Regression (SVR), and Convolutional Neural Network (CNN) models.
• Employed a multi-step approach to evaluate out-of-sample forecasting performance across various horizons.

Main Results:

• Incorporating Internet search behavior data from Google Trends significantly improved forecasting accuracy for both the unemployment rate and visitor numbers.
• The effectiveness of Google Trends data varied depending on the forecasting horizon and the specific modeling technique employed.
• The study identified optimal model and feature combinations for accurate macroeconomic forecasting.

Conclusions:

• Google Trends data offers a valuable, supplementary source for enhancing macroeconomic forecasts.
• The choice of forecasting technique (SARIMA, SVR, CNN) and forecasting horizon is critical for maximizing accuracy.
• This research provides a data-driven framework for selecting efficient forecasting models for economic indicators.