The US Federal Reserve began raising the federal funds rate in March 2022. Since then, nearly all asset classes have underperformed while the correlation between fixed-income assets and equities has rallied, rendering fixed income ineffective in its traditional role as a hedge.
With the value of asset diversification waning at least temporarily, achieving an objective and measurable understanding of the Federal Open Market Committee’s (FOMC) outlook is more important than ever.
This is where machine learning (ML) and natural language processing (NLP) come in. We applied Loughran-McDonald wordlists, BERT techniques, and XLNet ML for NLP to FOMC data to see if they expected changes in the fed funds rate and then examined whether our results correlated with stock market performance.
Loughran-McDonald Sentiment Word Lists
Before calculating emotion scores, we first constructed word clouds to visualize the frequency/importance of specific words in the FOMC data.
The Word Cloud: The March 2017 FOMC Statement
Word Cloud: July 2019 FOMC Statement
Although the Fed more March 2017 federal funds rate dropped In July 2019, the word cloud of the corresponding phrases looks the same. That’s because the FOMC statements generally contain many emotionless words with little bearing on the FOMC’s view. Thus, the drag word fails to distinguish the signal from the noise. But quantitative analyzes can provide some clarity.
Loughran-McDonald’s Sentiment Word Lists analyzes 10-K documents, earnings call transcripts, and other text by classifying words into the following categories: negative, positive, uncertainty, litigation, strong, weak, and limiting. We applied this technique to FOMC data, designating words as positive/hawkish or negative/pessimistic, while filtering out less important texts such as dates, page numbers, voting members, and explanations of monetary policy implementation. We then calculated sentiment scores using the following formula:
Feeling Score = (Positive Words – Negative Words) / (Positive Words + Negative Words)
FOMC Data: Loughran-McDonald Sentiment Results
As the previous chart shows, the FOMC’s statements grew more positive/hawkish in March 2021 and culminated in July 2021. After declining in the next 12 months, sentiment jumped again in July 2022. Although these moves may have been partially driven Recovering from the COVID-19 pandemic, they also reflect the FOMC’s increasing hawkishness in the face of rising inflation over the past year or so.
But the large fluctuations also indicate a flaw inherent in the Loughran-McDonald analysis: Sentiment scores only assess words, not sentences. For example, in the sentence “Unemployment has decreased,” both words are scored as negative/pessimistic even though the phrase indicates, as a sentence, an improving job market, which most interpret as positive/stressful.
To address this problem, we trained the BERT and XLNet models to analyze phrases on a sentence-by-phrase basis.
BERT and XLNet
Bidirectional encoding representations from transformers, or BERT, is a language representation paradigm that uses a two-way encoder instead of a one-way encoder to improve tuning. In fact, with the two-way encoder, we find that BERT is superior to OpenAI GPT, which uses a one-way encoder.
XLNet, meanwhile, is a generalized self-regressive pre-training method that also features a bidirectional encoder but not Masked Language Modeling (MLM), which feeds a sentence to BERT and optimizes the weights inside BERT to output the same sentence on the other side. Before we feed BERT the input sentence, we hide some tokens in MLM. XLNet avoids this, making it an improved version of BERT.
To train these two models, we divided the FOMC data into training data sets, test data sets, and out-of-sample data sets. We extracted training and test datasets from February 2017 to December 2020 and out-of-sample datasets from June 2021 to July 2022. We then applied two different labeling methods: manual and automatic. Using automatic classification, we assigned the sentences a value of 1, 0, or none based on whether they indicated an increase, decrease, or no change in the federal funds rate, respectively. Using manual labeling, we rated the sentences as 1, 0, or none depending on whether they were strict, pessimistic, or neutral, respectively.
Then we ran the following formula to generate the sentiment score:
Feeling score = (positive sentences – negative sentences) / (positive sentences + negative sentences)
performance of artificial intelligence models
Predicted sentiment score (auto-tagging)
Expected sentiment score (manual labeling)
The two charts above show that manual labeling better reflects the recent shift in the FOMC’s stance. Each statement includes hard (or pessimistic) sentences even though the FOMC ended up lowering (or increasing) the fed funds rate. In this sense, calling the sentence a sentence trains these ML models well.
Since ML and AI models tend to be black boxes, how to interpret their results is very important. One approach is to apply local interpretable model neutral interpretations (LIME). They apply a simple model to explain a more complex model. The two figures below show how XLNet interprets (with manual labeling) sentences from the FOMC data, where the first sentence reads as positive/hawkish based on a strengthening labor market and moderately expanding economic activities, and the second sentence as negative/pessimistic since consumer prices have fallen The inflation rate has fallen to less than 2%. The model’s judgment on both economic activity and inflationary pressure appears appropriate.
LIME Results: Strong economic sentence from the Federal Open Market Committee
LIME Results: Overall Weak Inflationary Pressures from the FOMC
By extracting sentences from the data and then assessing how they feel, these approaches have given us a better understanding of the FOMC’s policy perspective and have the potential to make it easier to interpret and understand central bank communications in the future.
But was there a relationship between changes in sentiment in the FOMC data and US stock market returns? The chart below plots the cumulative returns for the Dow Jones Industrial Average (DJIA) and the Nasdaq Composite (IXIC) along with the FOMC sentiment results. We investigated correlation, tracking error, excess return, and excess variability in order to detect system changes in ROEs, which are scaled by the vertical axis.
Equity returns and sensitivity scores of the FOMC statement
The results show that, as expected, our sentiment scores detect system changes, with stock market system changes and abrupt shifts in FOMC sentiment score occurring at roughly the same times. According to our analysis, the NASDAQ may be more responsive to the FOMC sentiment results.
Taken as a whole, this test hints at the huge potential of machine learning technologies for the future of investment management. Of course, in the final analysis, how these technologies stack up against human judgment will determine their ultimate value.
We would like to thank Yoshimasa Satoh, CFA, James Sullivan, CFA and Paul McCaffrey. Satoh organized and coordinated the AI study groups as moderator and reviewed and reviewed our report with thoughtful insights. Sullivan wrote Python code that converts PDF-formatted FOMC data into text, snippets, and related information. McCaffrey gave us great support in finishing this research report.
If you liked this post, don’t forget to subscribe to it Venture investor.
All posts are the opinion of the author. As such, it should not be construed as investment advice, nor do the opinions expressed necessarily reflect the views of the CFA Institute or the author’s employer.
Photo credit: © Getty Images / AerialPerspective Works
Professional learning for CFA Institute members
CFA Institute members are empowered to report self-earned and self-report Professional Learning (PL) credits, including content on Venture investor. Members can easily score credits using their online PL tracker.