News media versus FRED‐MD for macroeconomic forecasting

Abstract

Using a unique dataset of 22.5 million news articles from the Dow Jones Newswires Archive, we perform an in depth real-time out-of-sample forecasting comparison study with one of the most widely used data sets in the newer forecasting literature, namely the FRED-MD dataset. Focusing on U.S. GDP, consumption and investment growth, our results suggest that the news data contains information not captured by the hard economic indicators, and that the news-based data are particularly informative for forecasting consumption developments.

Real‐time macroeconomic projection using narrative central bank communication

“…A value less than 1 in the relative RMSE score indicates that the textaugmented forecast has a lower RMSE relative to the mean of the PF. Unlike the findings based on US data (Bianchi et al, 2022;Ellingsen et al, 2021), all four text-augmented predictions for GDP growth rate outperform the PF in our exercise. The differences in predictive performance are also statistically significant according to the DM test.…”

“…More concretely, we apply HDMR to construct the text‐based predictors from the PBC communication texts and then embed them into a MIDAS‐type model that can learn from mixed‐frequency macro and textual data, as well as produce the forecast to the target of interest. Accordingly, our results are related to the recent forecasting literature that uses text as data, for example, Kelly et al (2021) and Ellingsen et al (2021). In contrast, our work extends these applications by studying an entirely different corpus, central bank communication texts.…”

“…To meet the time‐varying characteristics of texts (Lima et al, 2020), we conduct a recursive window to estimate and update our text‐based PBC communication indices, which makes it more efficient to use all data when the parameters do not change substantially over time. It is also widely used in a number of forecasting studies (Andreou et al, 2013; Ellingsen et al, 2021; Stock & Watson, 2006).…”

“…A dictionary-based approach cannot make full use of the raw information in texts, nor can it accommodate the high dimensionality and sparsity of textual data. However, although LDA can address the high dimensionality of texts, it is computationally heavy for real-time estimation(Ellingsen et al, 2021;Thorsrud, 2020).…”

Making text count: Economic forecasting using newspaper text

“…Here, we assess algorithms that do not learn (for instance, dictionary and Boolean methods) as well as supervised machine learning as our sample size is large enough for it to be effective in forecasting the (continuous) economic variables that most concern policymakers. More recently, Ellingsen et al (2021) have shown that news text data contain information not captured by hard economic indicators, including information that is particularly useful in forecasting consumption.…”