使用chatgpt增强gdp即时预报：pmi新闻发布的新应用

Gabe J. de Bondt, Yiqiao Sun AbstractThis study involves tasking ChatGPT with classifying an “activity sentiment score” based on PMI newsreleases. It explores the predictive power of this score for euro area GDP nowcasting. We find that the PMItext scores enhance GDP nowcasts beyond what is embedded in ECB/Eurosystem Staff projections andEurostat’s first GDP estimate. The ChatGPT-derived activity score retains its significance in regressionsthat also include the composite output PMI diffusion index. GDP nowcasts are significantly enhanced withPMI text scores even when accounting for methodological variations, excluding extraordinary economicevents like the pandemic, and for different GDP growth quantiles. However, the forecast gains from theenhancement of GDP nowcasts with ChatGPT scores are time dependent, varying by calendar years.Sizeable forecast gains of on average about 20% were obtained apart from the two most recent years dueto exceptionally low forecast errors of the two benchmarks, especially the first GDP estimate.KeywordsChat Generative Pre-training Transformer,text analysis,zero-shot sentiment analysis,Purchasing Managers’ Index (PMI), nowcasting GDPJEL ClassificationC8; E32; C22ECB Working Paper Series No 3063 1 Non-technical summaryThis study introduces a novel approach to economic forecasting by utilizing artificial intelligence (AI),specifically ChatGPT, to enhance predictions of euro area Gross Domestic Product (GDP). Traditionalmethods of economic forecasting primarily rely on numerical data, such as hard data on industrialproduction and retail sales as well as soft data such as Purchasing Managers’ Index (PMI) diffusion indices.Instead, this research explores the potential of integrating qualitative information - textual content fromPMI news releases - into existing forecasts. What makes this study unique is its focus on the narrative, tone,and anecdotes reported in PMI news releases. ChatGPT was employed to analyse PMI news releases andgenerate activity sentiment scores. These scores quantify the sentiment about activity expressed in thenarratives and anecdotes of the news release, ranging from strongly expanding activity to significantlycontracting activity. The study then integrates these ChatGPT-derived activity scores in traditional GDPnowcasts, i.e., forecasts of real GDP growth in the current quarter, to assess their predictive power. Theanalysis utilizes two notoriously hard-to-beat benchmarks for GDP nowcasting, setting a high standard foraccuracy, namely ECB/Eurosystem Staff projections and the first GDP estimate. The first benchmarkincludes judgment from experts and the second relies on statistical methods that fill a sizeable gap ofmissing statistical information for the first estimate of GDP.The main compelling result is that the enhancement of the PMI text scores to the two GDP nowcastbenchmarks significantly improves the accuracy of GDP nowcasts. Similar in-sample gains are not obtainedby adding the composite output PMI diffusion index. Ordinary least squares, robust least squares, and ridgeregressions all show that the diffusion index has no value added to the benchmark GDP nowcast or evencontributes counterintuitively negatively. The GDP nowcasting results enhanced with the ChatGPT-derivedscore holds even when accounting for methodological variations, excluding extraordinary economic eventslike the pandemic, and for different GDP growth quantiles. The out-of-sample forecast gains of enhancingGDP nowcasts with PMI text scores are on average about 20% apart from the two most recent years, butthey are strongly time-dependent, varying by calendar years. The study shows that the qualitative insightsfrom the PMI narratives and anecdotes provide valuable information that complements the numerical data,offering a more comprehensive assessment of real GDP growth.Our results imply the following. They confirm earlier findings that economic forecasting can be enhancedby integrating qualitative data sources into traditional models. A new element is that this research showsthat only two pages of text rather than for example millions of newspapers articles can be sufficient toenhance existing hard-to-beat benchmarks. Moreover, the robustness of these findings across differentmethodological adjustments underscores the potential of AI in economic forecasting. This study advancesthe field of AI-driven economic forecasting and provides a new practical tool for policymakers, financialanalysts, and economists to predict more accurately GDP. The success of ChatGPT opens new avenues forfurther research, such as applying similar techniques to other types of economic texts, including forecastreports of policy institutions. Additionally, this method could be explored for other regions or countries.ECB Working Paper Series No 3063 2 1 IntroductionThis study explores the integration of ChatGPT-based sentiment activity scores into existing GrossDomestic Product (GDP) nowcasts. We hypothesize th