Abstract
The application of multivariate time series is extensive, spanning economic, biological, and complex systems. This paper introduces the construction and structure of the ARIMAX model (Auto-Regressive Integrated Moving Average with explanatory variables) and its implementation using SAS. By analyzing China's tertiary industry with the real estate business as an input variable, the study demonstrates a co-integration relationship between these two series. The resulting ARIMAX model, fitted with historical data from 1978 to 2007, proves to be a high-precision tool for economic forecasting.
Introduction: Beyond Univariate Forecasting
Forecasting economic trends is a complex challenge because industries are deeply interconnected. Traditional ARIMA models, while useful, are univariate—they rely solely on a variable's own past to predict its future. This creates a "blind spot" regarding external influences. In this study, researchers address this limitation by applying the ARIMAX model to China’s tertiary (service) industry, incorporating real estate data as an explanatory variable to capture cross-sector dynamics.
Methodology
Building an effective ARIMAX model requires a rigorous multi-step process to ensure statistical validity:
- Stationarity and Transformation: Economic data typically exhibits exponential growth. To stabilize the variance and mean, the authors applied logarithmic transformations and first-degree differencing.
- Cointegration Analysis: Using Engle-Granger theory, the team verified a long-term equilibrium relationship between the service and real estate sectors, ensuring the observed correlation was not spurious.
- Pre-whitening: To isolate the true relationship between the two series, ARMA models were used to filter internal noise from the input series.
- Cross-Correlation and Lag Identification: Analysis revealed the specific temporal delay at which real estate trends impact the broader service industry.
- Parameter Estimation: Final model parameters were estimated via the Least Squares method within a SAS environment.
Data Overview
The study draws on historical data from the China Statistical Yearbook (1978–2007). The primary output variable is the production value of China's tertiary industry, while the input variable is the production value of the real estate sector.
Real estate was selected as the explanatory variable due to its foundational role in China's economic expansion during this 30-year period. As shown in Figure 1, both sectors exhibit strong upward trends, necessitating transformation for accurate modeling.
Key Insights: The Leading Indicator
A critical discovery was the one-year lag effect. Cross-correlation analysis showed that fluctuations in the real estate market significantly impact the tertiary sector with a one-year delay. This identifies real estate as a "leading indicator," providing valuable lead time for economic policy and planning.
Furthermore, the study emphasizes that while individual economic series may be non-stationary (trending), their residuals can be stationary. This "cointegration" allows for robust modeling of long-term relationships without losing the underlying trend information.
Key Results and Model Precision
The ARIMAX model demonstrated exceptional fit and predictive accuracy. Statistical tests confirmed that the influence of real estate on the service sector is highly significant, with a parameter estimate of approximately 0.90469 for the lagged input variable.
As shown in the visual validation in Figure 2, the model's predicted values track the actual historical data for the stationary component almost perfectly. This high precision confirms that ARIMAX is a superior tool for capturing the nuances of interconnected economic sectors compared to simpler ARIMA models.
Conclusion and Future Directions
While this research proves the efficacy of ARIMAX for the tertiary industry, it also opens doors for further refinement. Future models could incorporate multiple input variables—such as technological innovation or international trade metrics—to create an even more comprehensive forecasting tool. Additionally, applying these methods to more recent data would help determine if the established one-year lag relationship persists in the modern digital economy.
Summary updated