Developing spatio-temporal approach to predict economic dynamics based on online news

Yuzhou Zhang, Hua Sun, Guang Gao, Lidan Shou, Dun WuScientific 2022 · DOI

TL;DR

This paper introduces a spatio-temporal approach to predict economic dynamics in China using online news data. By combining Generalized Linear Models (GLM) and Geographically Weighted Regression (GWR), the study demonstrates that online news sentiment can effectively forecast economic output, with GWR showing superior performance in capturing spatial variations. This work offers a novel data source for economic forecasting and a foundation for developing real-time economic surveillance systems.

What Problem Does This Paper Address?

Economic forecasting is crucial for government planning and policy-making. While traditional methods rely on statistical data, the advent of big data, particularly online news, presents a promising new avenue for more dynamic and timely economic predictions. However, existing research on leveraging online news for economic forecasting in China is limited, especially concerning the spatio-temporal dynamics – how economic patterns change across both space and time. This paper aims to fill this gap by exploring these spatio-temporal variations and developing models to predict economic output using online news data in Yinzhou, Ningbo, China.

Methods

The study employs two primary modeling techniques:

  • Generalized Linear Model (GLM): This statistical framework extends traditional linear models to accommodate non-normally distributed dependent variables and allows for a linear relationship between economic output and predictors through a specified link function. The researchers used GLM to quantify the relationship between economic output dynamics and online news patterns, considering industry policy as a confounding factor. A negative binomial distribution was chosen to handle potential over-dispersion.
  • Geographically Weighted Regression (GWR): Recognizing that spatial relationships can vary across a study area, GWR was introduced to address spatial non-stationarity. Unlike OLR, GWR estimates local regression coefficients for each spatial unit, allowing for a more nuanced understanding of how predictors influence economic dynamics at different locations. This method is particularly suited for analyzing spatio-temporal data by incorporating geographic coordinates into the regression.

The models were trained on data from 2018-2019 and validated on data from 2020-2021. Model performance was evaluated using metrics like R-squared, AIC, Pearson correlation, RMSE, and MAPE.

Datasets

  • Economic Output Data: Annual economic output values for 1180 manufacturing enterprises in Yinzhou, Ningbo, China, from 2018 to 2021. This data was sourced from the Yinzhou Economic Statistics Platform and represents over 80% of the total annual output value from these enterprises.
  • Online News Data: The number of positive and negative online news articles for each enterprise over the same period. This data was collected from several major Chinese online news websites (Baidu news, Jinritoutiao, Fengguangwang, Sina News, Souhu News, Wangyi News) using enterprise names as keywords.
  • Industry Policy Data: Transformed numeric data representing government policies (development policy: +1, restriction policy: -1) to control for policy influences on economic dynamics.

Key Insights and Ideas

  • Online News as a Predictive Signal: The study strongly suggests that online news sentiment is a valuable predictor of economic dynamics, offering a more immediate and granular insight than traditional data sources.
  • Spatio-Temporal Dynamics are Crucial: Economic activity is not uniform; it varies significantly across both geographic locations and time. Capturing these spatio-temporal variations is key to accurate forecasting.
  • GWR Outperforms GLM: The Geographically Weighted Regression model demonstrated superior performance in predicting economic dynamics, highlighting the importance of accounting for spatial heterogeneity. This suggests that local factors and relationships play a significant role in economic fluctuations.
  • Foundation for Surveillance Systems: The findings lay the groundwork for developing real-time, online news-based surveillance systems that can provide early warnings and insights into economic trends.

Key Results

  • Spatio-temporal Variations: The study confirmed significant spatio-temporal variations in economic output values within Yinzhou. Agglomerations of higher economic output were observed in the northwest and southeast regions.
  • GLM Performance: GLMs combining the number and percentage of positive online news showed good performance in predicting economic dynamics, with statistically significant positive contributions from online news sentiment. For instance, a one-unit increase in positive online news was associated with an increase of 50.86 units in economic output value.
  • GWR Performance: The GWR model generally exhibited better goodness-of-fit than GLM. For the training period (2018-2019), the GWR model for predicting economic output value achieved an R² of 0.82.
  • Predictive Accuracy: In the predictive period (2020-2021), the GWR models showed high Pearson correlations for predicting economic output value (0.96 for 2020, 0.94 for 2021) and annual change (0.97). MAPE values were also low (e.g., 1.26% for 2020 economic output value).
  • Online News Impact: The annual number of positive online news consistently showed higher regression coefficients in the GWR model, suggesting it's a strong predictor. Similarly, the percentage of positive online news was a key predictor for the change in economic output value.

Limitations and Future Work

  • Geographic Scope: The study was conducted in Yinzhou, Ningbo, China. The authors suggest extending the proposed model to larger geographic areas to validate its applicability and robustness.
  • Data Granularity: While using enterprise-level data, the analysis aggregated it into hexagonal areas. Future work could explore finer spatial resolutions.
  • News Sentiment Nuance: The study primarily focused on the number and percentage of positive news. More sophisticated sentiment analysis techniques could potentially yield even richer insights.
  • Policy Data: While industry policy was included, a more detailed and dynamic incorporation of policy changes could further refine the models.
  • Model Extension: The authors propose extending the spatio-temporal approach to other economic indicators and geographic regions, aiming to build a comprehensive online news-based economic surveillance system.

Summary updated

Abstract

Economic forecasting is a scientific decision-making tool, and it is one of the important basis for the government to formulate economic plans, predict the implementation of the plan, and guide the implementation of the plan. Current knowledge about the use of online news in the prediction of economic patterns in China is limited, especially considering the spatio-temporal dynamics over time. This study explored the spatio-temporal patterns of economic output values in Yinzhou, Ningbo, China between 2018 and 2021, and proposed generalized linear model (GLM) and Geographically weighted regression (GWR) model to predict the dynamics using online news data. The results indicated that there were spatio-temporal variations in the economic dynamics in the study area. The online news showed a great potential to predict economic dynamics, with better performance in the GWR model. The findings suggested online news combining with spatio-temporal approach can better forecast economic dynamics, which can be seen as a pre-requisite for developing an online news-based surveillance system The advanced spatio-temporal analysis enables governments to garner insights about the patterns of economic dynamics over time, which may enhance the ability of government to formulate economic plans and to predict the implementation of the plan. The proposed model may be extended to greater geographic area to validate such approach.
BibTeX
@inproceedings{zhang2022developing,
  author = {Yuzhou Zhang and Hua Sun and Guang Gao and Lidan Shou and Dun Wu},
  title = {{Developing spatio-temporal approach to predict economic dynamics based on online news}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-20489-w},
}
Figure