Scholarlink

<p>This study presents a new Time-Sentiment Interaction (TSI) feature designed to improve Consumer Price Index (CPI) forecasting by capturing dynamic relationships between news sentiment and inflation movements. News articles related to inflation were collected from five major online sources using the GDELT database, translated into English, and analyzed with the Economy Lexicon and FinVADER. Sentiment scores derived from article titles and filtered inflation-related content were aggregated into monthly indicators and incorporated into machine learning models along with the proposed TSI feature. A comprehensive empirical evaluation was conducted that compared classical univariate models with multivariate machine learning approaches, including Support Vector Regression&#xa0;(SVR), Random Forest, and extreme gradient boosting (XGBoost), under a rigorous rolling window validation scheme with hyperparameter optimization. The results show that the TSI feature provides substantial gains in predictive accuracy across all machine learning models. XGBoost with this feature achieved the best performance, with an RMSE of 0.299 and an MAE of 0.225, which represents more than 14.81% improvement over the sentiment-based version of the model and a clear advantage over established benchmarks such as ETS, ARIMA, and SARIMAX. Diebold Mariano tests confirmed the statistical significance of these improvements, while residual diagnostics indicated lower error variance and greater forecast stability for models that incorporated the TSI feature. These findings demonstrate that the proposed feature provides a more informative representation of news-driven behavioral signals and serves as an effective enhancement for inflation forecasting, offering improved accuracy and greater reliability for policymakers and analysts.</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

Time sentiment interaction features for enhanced forecasting of economic indicators

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Thi Mai Trang Nguyen,
  • Chi Hieu Dinh,
  • Quang Hung Do,
  • Quoc Khanh Tran

摘要

This study presents a new Time-Sentiment Interaction (TSI) feature designed to improve Consumer Price Index (CPI) forecasting by capturing dynamic relationships between news sentiment and inflation movements. News articles related to inflation were collected from five major online sources using the GDELT database, translated into English, and analyzed with the Economy Lexicon and FinVADER. Sentiment scores derived from article titles and filtered inflation-related content were aggregated into monthly indicators and incorporated into machine learning models along with the proposed TSI feature. A comprehensive empirical evaluation was conducted that compared classical univariate models with multivariate machine learning approaches, including Support Vector Regression (SVR), Random Forest, and extreme gradient boosting (XGBoost), under a rigorous rolling window validation scheme with hyperparameter optimization. The results show that the TSI feature provides substantial gains in predictive accuracy across all machine learning models. XGBoost with this feature achieved the best performance, with an RMSE of 0.299 and an MAE of 0.225, which represents more than 14.81% improvement over the sentiment-based version of the model and a clear advantage over established benchmarks such as ETS, ARIMA, and SARIMAX. Diebold Mariano tests confirmed the statistical significance of these improvements, while residual diagnostics indicated lower error variance and greater forecast stability for models that incorporated the TSI feature. These findings demonstrate that the proposed feature provides a more informative representation of news-driven behavioral signals and serves as an effective enhancement for inflation forecasting, offering improved accuracy and greater reliability for policymakers and analysts.