全球汽車新聞情緒趨勢訊號與月度銷售滯後效應之實證研究

Global Automotive News Sentiment Trend Signals and Lagged Effects on Monthly Sales: An Empirical Investigation

本研究探討全球汽車產業新聞情緒與全球月度汽車銷售量之間的關聯性，分析期間涵蓋2016年1月至2025年12月，共120個月。研究以GDELT全球事件資料庫為新聞來源，共計13，104篇汽車業新聞，並採用三種均以yiyanghkust／finbert-tone為底層模型的長文本處理策略進行比較：（1）前段截斷模型、（2）全文分段聚合模型，以及（3）滑動視窗分段模型。全球月度銷售資料取自MarkLines資料庫，涵蓋64個國家及5個地區。研究以月度情緒指數（MSI）對月銷售成長率進行落後期OLS回歸。
實證結果顯示，在原始月度情緒規格下，三模型之最佳落後期均為12個月，但單月情緒係數因噪訊過多均未達統計顯著水準（p > 0.10），且模型資訊準則（AIC）較基準模型有所惡化。然而，經平滑處理之三個月移動平均指標（MSI_MA3）展現出穩健的解釋力，使Adjusted R2顯著提升至0.219 ~ 0.225，且AIC均優於基準模型。結果表明，全球汽車新聞情緒並非高頻的即時預測因子，而是一種具備中長期輔助預測價值的低頻趨勢訊號。

This study investigates the relationship between automotive industry news sentiment and global monthly auto sales over the period January 2016 to December 2025 (120 months). Using 13,104 automotive news articles sourced from the GDELT database, we construct Monthly Sentiment Indices (MSI) via three long-text processing strategies sharing the same base model yiyanghkust/finbert-tone: (1) front-512 truncation, (2) full-text chunk averaging, and (3) sliding-window chunk averaging. Global monthly sales data are obtained from MarkLines, covering 64 countries across 5 regions.
OLS regressions show that while the optimal lag is 12 months across all models, the coefficients of the raw monthly MSI are statistically insignificant (p > 0.10) due to high temporal noise, and the model fit (AIC) deteriorates compared to the baseline. Crucially, however, the smoothed 3-month moving average specification (MSI_MA3) robustly improves model fit, yielding an increased Adjusted R2 of 0.219–0.225 and lower AIC values across all three strategies. These findings suggest that automotive news sentiment operates not as a high-frequency shock, but as a low-frequency macroeconomic trend signal with long-term supplementary predictive value.