Scholarlink

<p>Gross Domestic Product (GDP) is one of the most important factors to determine the economic state and performance of a country. The correct prediction of GDP is crucial in the development of the fiscal policy, investment and in socio-economic planning. This study makes the following contribution: the first attempt to adopt a lightweight neural network architecture with a unique design to predict GDP is compared to a number of other well-known regression models. When carrying out the study, one starts with thorough data preprocessing, which refers to removing the noise, the treatment of missing values, and normalizing it. In order to be able to make some assumptions about the patterns and relationships between features, statistical summaries and data visualizations were conducted as part of Exploratory Data Analysis (EDA). As the benchmark models, multiple regression models were created and optimized, such as Linear Regression, Ridge Regression, Lasso Regression, Decision Tree Regressor, XGBoost Regressor and etc., All models were tested with a set of performance metrics in terms of accuracy, i.e., Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), R2 Score, Mean Absolute Percentage Error (MAPE) and Root Mean Square Logarithmic Error (RMSLE). The lightweight neural network proposed and fine-tuned with the purpose of achieving a higher training accuracy as well as the best generalization performance, performed outstandingly on all assessment measures. The model combines residual learning, batch normalization, dropout based regularization and kernel constraints to improve generalization while keeping the architecture simple. A five-fold cross-validation is used to ensure reliability and to avoid the possibility of overfitting. Also, the resilience of the model was measured quantitatively by a robustness test at structural noise perturbations, which is the case of non-stationary economic dynamics. The findings have shown a high predictive performance with a moderate R2 score and a stability ratio of 1.02, which is a good indicator of high resilience to changes in inputs. The proposed model was found to have a better predictive performance and significantly lower error scores for MAE, RMSE and computational cost than traditional regression methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A lightweight neural network approach for predicting national Gross Domestic Product (LightNet-GDP) with regression benchmarks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月29日

查看全文
  • Ch. V. Raghavendran,
  • K. Chandra Mouli,
  • S. Bhargavi Latha,
  • Ashwinkumar U. Motagi

摘要

Gross Domestic Product (GDP) is one of the most important factors to determine the economic state and performance of a country. The correct prediction of GDP is crucial in the development of the fiscal policy, investment and in socio-economic planning. This study makes the following contribution: the first attempt to adopt a lightweight neural network architecture with a unique design to predict GDP is compared to a number of other well-known regression models. When carrying out the study, one starts with thorough data preprocessing, which refers to removing the noise, the treatment of missing values, and normalizing it. In order to be able to make some assumptions about the patterns and relationships between features, statistical summaries and data visualizations were conducted as part of Exploratory Data Analysis (EDA). As the benchmark models, multiple regression models were created and optimized, such as Linear Regression, Ridge Regression, Lasso Regression, Decision Tree Regressor, XGBoost Regressor and etc., All models were tested with a set of performance metrics in terms of accuracy, i.e., Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), R2 Score, Mean Absolute Percentage Error (MAPE) and Root Mean Square Logarithmic Error (RMSLE). The lightweight neural network proposed and fine-tuned with the purpose of achieving a higher training accuracy as well as the best generalization performance, performed outstandingly on all assessment measures. The model combines residual learning, batch normalization, dropout based regularization and kernel constraints to improve generalization while keeping the architecture simple. A five-fold cross-validation is used to ensure reliability and to avoid the possibility of overfitting. Also, the resilience of the model was measured quantitatively by a robustness test at structural noise perturbations, which is the case of non-stationary economic dynamics. The findings have shown a high predictive performance with a moderate R2 score and a stability ratio of 1.02, which is a good indicator of high resilience to changes in inputs. The proposed model was found to have a better predictive performance and significantly lower error scores for MAE, RMSE and computational cost than traditional regression methods.