Forecasting model based on information-granulated GA-SVR and ARIMA for producer price index

Xiangyan Tang, Liang Wang, Jieren Cheng, Jing ChenarXiv 2019 · arXiv

Introduction

The Producer Price Index (PPI) is a critical economic indicator that measures the average change over time in the selling prices received by domestic producers. It serves as an essential early warning signal for inflation and deflation. However, forecasting the PPI is notoriously difficult because the data is "chaotic," containing both stable linear trends and volatile non-linear fluctuations.

Traditional parametric models like ARIMA often struggle to capture non-linear complexities, while non-parametric machine learning models like Support Vector Regression (SVR) can be sensitive to data noise. This research proposes a hybrid intelligence system that leverages the strengths of both statistical and machine learning approaches to improve forecasting precision.

Methodology

The authors developed a three-stage pipeline designed to handle the imprecision and complexity of economic data. The process integrates fuzzy information granulation, Genetic Algorithm-optimized SVR, and ARIMA error correction.

Flowchart of the proposed GA-SVR-ARIMA hybrid forecasting model.

1. Fuzzy Information Granulation (FIG)

Instead of using raw monthly data, the model employs FIG to pre-process the PPI data. The timeline is divided into 3-month "windows," from which three fuzzy granules are extracted: the Minimum, Average (Mean), and Maximum values. This step reduces data dimensionality and helps the model manage the inherent "fuzziness" of economic shifts by smoothing seasonal volatility.

2. GA-SVR (Non-linear Modeling)

Support Vector Regression (SVR) is used to model the non-linear components of these granules. To optimize the model, a Genetic Algorithm (GA) is employed to automatically evolve the optimal hyperparameters, including the penalty parameter ($C$), loss function ($\epsilon$), and kernel parameters ($\gamma$), removing the guesswork from manual configuration.

3. ARIMA Error Correction (Linear Modeling)

To further refine the results, the residual errors—the difference between the GA-SVR prediction and the actual value—are processed through an ARIMA model. Since these residuals often contain linear patterns that the SVR might miss, ARIMA "rectifies" the forecast to produce the final estimate.

Data and Experimental Setup

The study utilizes a dataset of 192 monthly PPI data points from the China National Bureau of Statistics, spanning from January 2001 to December 2016.

  • Training Set: Data from January 2001 to September 2016.
  • Testing Set: Data from October 2016 to December 2016, used to validate out-of-sample performance.

Results and Discussion

The hybrid GA-SVR-ARIMA model significantly outperformed standalone models across all key metrics. The complementarity of the components proved essential: SVR excelled in non-linear spaces, while ARIMA captured the remaining linear residuals.

  • Accuracy: The hybrid model achieved a Mean Squared Error (MSE) of 0.1322.
  • Comparison: For the same period, the standalone ARIMA model had an MSE of 1.5496, and the standalone GA-SVR model (without ARIMA correction) had an MSE of 14.5109.
  • Error Reduction: The hybrid approach reduced the Mean Absolute Percentage Error (MAPE) to just 0.28%, compared to 1.04% for ARIMA and 2.98% for GA-SVR.

Conclusion

The proposed hybrid model provides a robust framework for high-precision PPI forecasting. By combining fuzzy granulation for denoising, GA-SVR for non-linear modeling, and ARIMA for linear error correction, the system captures complex economic patterns more effectively than traditional methods. Future work may involve integrating external economic variables, such as policy shifts or international trade fluctuations, to enhance long-term forecasting capabilities.

Summary updated

BibTeX
@misc{tang2019forecasting,
  author = {Xiangyan Tang and Liang Wang and Jieren Cheng and Jing Chen},
  title = {{Forecasting model based on information-granulated GA-SVR and ARIMA for producer price index}},
  year = {2019},
  eprint = {1903.12012},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1903.12012},
}
Figure