Research on the Scale Prediction of the Digital Economy Integrating Statistical Models and Machine Learning

Xueying Han, Yan Wang, Guoqing Chen, Zhao TianwenDEVELOPMENT 2025 · DOI

TL;DR

This paper proposes a hybrid forecasting framework that combines traditional statistical models (ARIMA) with machine learning (BPNN/LSTM) to predict the scale of the digital economy. By utilizing a nonlinear error correction strategy and Quantum Particle Swarm Optimization (QPSO), the researchers achieved a predictive accuracy improvement of up to 73.8% over traditional standalone methods.

What Problem Does This Paper Address?

Predicting the growth of the digital economy is notoriously difficult due to its high volatility and the complex interplay of internal and external factors. The authors identify a "dual dilemma" in current forecasting:

  1. Traditional Statistical Models (e.g., ARIMA): While interpretable and theoretically sound, they struggle to capture the nonlinear patterns inherent in modern economic shifts.
  2. Machine Learning Models (e.g., BPNN, LSTM): While excellent at capturing nonlinearities, they often suffer from overfitting when dealing with "high-dimensional small samples"—a common scenario in economic datasets where data points are limited but variables are numerous.

Additionally, existing data often suffers from missing values, inconsistent dimensions, and feature redundancy, making raw data difficult to use effectively.

Methods

The authors developed a multi-stage combinatorial forecasting pipeline:

  • Indicator System & Preprocessing: They constructed a system based on four dimensions: digital infrastructure, industrialization, digitization, and innovation. They used linear regression for missing data imputation and Z-score standardization.
  • Feature Reduction (PCA): Principal Component Analysis was employed to eliminate multicollinearity and reduce dimensionality, selecting top components that account for over 85% of the variance.
  • Nonlinear Error Correction: Instead of a single model, they used a "residual-based" approach. An ARIMA model makes a preliminary linear prediction. The residuals (errors) from this model are then fed into a neural network (BPNN or LSTM) to capture the remaining nonlinear patterns. The final forecast is the sum of the ARIMA prediction and the predicted error.
  • Meta-Heuristic Optimization (QPSO): To avoid the "blind" selection of neural network parameters (like weights and thresholds), they introduced Quantum Particle Swarm Optimization. QPSO uses quantum bit encoding to perform a more robust global search than standard PSO, preventing the model from getting stuck in local optima.

Datasets

The study utilized a specific regional dataset to validate the model:

  • Source: Digital economy-related indicator data from Jiangxi Province, China.
  • Timeframe: 2011 to 2020.
  • Metrics: The dataset includes variables related to digital infrastructure, industrialization, and innovation capabilities.

Key Insights and Ideas

  • The Power of Hybridization: The most significant "aha moment" is that ML models perform better when tasked with predicting the errors of a statistical model rather than the raw data itself. This anchors the prediction in established economic theory while allowing ML to handle the "noise" and complexity.
  • Quantum-Inspired Tuning: Moving from standard PSO to QPSO (Quantum PSO) significantly improved convergence speed and parameter optimization, which is critical for complex models like LSTMs.
  • Dimensionality Matters: The paper emphasizes that for economic forecasting, "less is more." Using PCA to strip away redundant features was essential to preventing the neural networks from overfitting on the small sample size.

Key Results

The PCA-BPNN model emerged as the top performer across all metrics:

  • Accuracy: Improved predictive accuracy by 30% to 73.8% compared to traditional methods.
  • Performance Metrics:
    • MAE (Mean Absolute Error): 240.018 (the lowest among tested models).
  • MAPE (Mean Absolute Percentage Error): 0.0233.
  • $R^2$ (Coefficient of Determination): 0.976 (indicating an extremely high fit).
  • Global Forecast: The model predicts worldwide digital trade will reach **$7.39 trillion in 2024**, with China ranking 4th globally at approximately $765.898 billion.

Limitations and Future Work

The authors acknowledge several hurdles for future research:

  • Dynamic Indicators: The digital economy evolves rapidly; the current indicator system may need frequent updates to stay relevant.
  • Computational Complexity: The combined model, especially with meta-heuristic optimization, is computationally expensive and may struggle with real-time prediction requirements.
  • Data Granularity: There is a persistent lack of detailed regional and high-frequency data, which hinders the development of more refined, localized models. Future work will focus on increasing data openness and optimizing algorithm efficiency.

Summary updated

Abstract

The digital economy has become an important component of our country's modern economic system, but its operating conditions are influenced by multiple factors and face significant uncertainties during its development. This paper addresses the issue of forecasting the scale of the digital economy by proposing a combinatorial forecasting method that integrates statistical models and machine learning. The study first constructs an index system of influencing factors suitable for the digital economy in China, addressing issues such as missing data, inconsistent dimensions, and feature redundancy; furthermore, it proposes a combinatorial model framework based on nonlinear error correction strategy and meta-heuristic algorithm optimization. Empirical research shows that the combinatorial model significantly outperforms individual models in terms of convergence speed, parameter optimization, and predictive accuracy, with the PCA-BPNN model's predictive accuracy improving by 30% to 73.8% compared to traditional methods. The research findings provide a scientific basis for government departments in formulating digital economy policies and offer theoretical support for regional digital economy development strategies.
BibTeX
@inproceedings{han2025research,
  author = {Xueying Han and Yan Wang and Guoqing Chen and Zhao Tianwen},
  title = {{Research on the Scale Prediction of the Digital Economy Integrating Statistical Models and Machine Learning}},
  booktitle = {DEVELOPMENT},
  year = {2025},
  doi = {10.47297/wspdecwsp2515-797321.20250908},
}
Figure