25 Jun 2024 4 min read

Beige Book reloaded

By Rameet Gulsin

Last year, I wrote about how AI helped me analyse the Beige Book, the Federal Reserve's summary of current economic conditions. This year, and with a little bit of python programming, I’ve supercharged my analysis with quantitative measures of word intensity and AI-derived sentiment scores.


Economic data readings such as GDP are released infrequently and with significant lags. Extracting more timely information from the Beige Book can provide insight into the current state of the economy.

The Minneapolis Fed maintains an archive of the entire history of Beige Books. I scrape, copy and organise this archive by date and representative district, before splitting the raw text into a format for processing. After a bit of data cleaning, I have a database that makes analysing historical trends in the Beige Book a little bit easier.

Intensity of ‘inflation’ and ‘recession’

Topics mentioned in the Beige Book are driven by economic forces reflecting conditions at the time of writing. These can vary from the mundane to war, drought or pandemics. This also means the size of the document is not fixed. Measured by total word count, the average Beige Book today is larger than previous publications.

Therefore, one approach for extracting word intensity and analysing historical trends is to consider a statistic known as the term frequency-inverse document frequency (TF-IDF). It takes the pure word count, controls for changes in the size of a document over time, and scales down the impact of more frequently occurring words that may be empirically less informative. I plot the TF-IDF for ‘inflation’ and ‘recession’ below.


The positive relationship between inflation-intensity and CPI (adjusted for inflation targeting) shouldn’t be too surprising. The more inflation is above target, the more important inflation becomes to economic agents who are questioned for the Beige Book.

This is most apparent in the 1970s to mid-1990s, after which inflation hovers around target and inflation-intensity subsides. However, this correlation breaks down after the global financial crisis (GFC), when inflation-intensity begins to rise even as inflation remains below or at target.

The reasons for this are currently unclear. Districts may be starting to report more on worryingly low inflation, or there may be concentrated pockets of inflationary pressure within districts, causing an increase in the total intensity score. More apparent is the post-COVID-19 spike to inflation and inflation-intensity, which have since peaked and fallen back. The latest Beige Book, published at the end of May, shows that inflation intensity has started to rise once again, suggesting a resurgence of inflation worries.

Recent observations of recession intensity are a more welcoming read. Recession intensity peaked to its highest level in almost 30 years and has since faded to almost zero. On this occasion recession fears proved wrong. However, this highlights continued uncertainty in a post-pandemic world.

Sentiment trends: A FinBERT perspective

Another approach to extracting historical trends is to use FinBERT, a pre-trained large-language-model (LLM) based on a framework published by Google. Fine-tuned specifically for applications related to finance, I use FinBERT to classify Beige Book sentiment on a sentence-by-sentence basis and compute an aggregated Beige Book sentiment score, which I plot below.

Using LLMs to calculate sentiment and extract information from unstructured text is not new. I did so in my last blog post and for more examples see, using sentiment as a predictor for NBER dated recessions, measuring the effect of supply-chain bottlenecks and analysing the impact of Fed transparency on policy expectations. Data scientists and economists will continue to push this research envelope, and I have it under good authority that the Fed is also exploring ways to fine-tune their very own FedBERT.


I find a statistically significant and positive relationship between sentiment and cyclical GDP. Isolating observations occurring during NBER-dated recessions, we can identify a threshold sentiment score of -0.25, below which recessions become more likely. Curiously, since mid-2023, sentiment has hovered around this threshold, and yet the US economy shows continued resilience (see a previous blogs here and here for a discussion of why).

Overall, the relationship between sentiment and cyclical GDP looks to have weakened slightly 1) post-GFC, with sentiment tending to overestimate GDP and 2) post-pandemic, with sentiment tending to underestimate GDP. More recently, the data look to have converged with sentiment recovering close to neutral, suggesting US growth in line with potential.

Rameet Gulsin

Quantitative Economist

Ram first joined LGIM in 2015 as part of the Client Distribution team. After leaving his role to pursue research and teaching at the University of Kent, he was welcomed back into the Asset Allocation team as a Quantitative Economist in 2021. You’ll often find Ram reading from his favourite econometrics textbook. He’s recently discovered its hidden power of sending his new-born daughter straight to sleep.

Rameet Gulsin