Machine Learning and FOMC Statements: What’s the Sentiment?

The US Federal Reserve began raising the federal funds rate in March 2022. Since then, just about all asset classes have performed poorly, while the correlation between fixed income assets and stocks has increased sharply, rendering fixed income securities ineffective of their traditional role as a hedging tool.

As the worth of asset diversification has diminished, not less than temporarily, it has turn out to be increasingly essential to realize an objective and quantifiable understanding of the Federal Open Market Committee’s (FOMC) outlook.

This is where machine learning (ML) and natural language processing (NLP) come into play. We applied Loughran-McDonald sentiment word lists and BERT and XLNet ML techniques for NLP to FOMC statements to see in the event that they anticipated changes within the federal funds rate, after which examined whether our results had any connection to stock market performance had.

Loughran-McDonald mood word lists

Before calculating sentiment scores, we first created word clouds to visualise the frequency/importance of specific words in FOMC statements.

Word Cloud: March 2017 FOMC Statement

Word Cloud: July 2019 FOMC Statement

Although the Fed set the federal funds rate in March 2017 and July 2019, the word clouds of the 2 corresponding statements look similar. That’s because FOMC statements generally contain quite a lot of sentimental language that has little bearing on the FOMC’s outlook. Therefore, the word clouds couldn’t distinguish the signal from the noise. But quantitative evaluation can provide clarity.

Loughran-McDonald’s sentiment word lists analyze 10-K documents, telephone call transcripts, and other texts by grouping words into the next categories: negative, positive, uncertain, contentious, strong modal, weak modal, and limiting. We applied this system to FOMC statements, labeling words as positive/hawkish or negative/devisive while filtering out less essential text comparable to dates, page numbers, voting members, and explanations of monetary policy implementation. We then calculated the sentiment scores using the next formula:

Sentiment Score = (Positive Words – Negative Words) / (Positive Words + Negative Words)

FOMC Statements: Loughran-McDonald Sentiment Scores

As the chart above shows, FOMC statements became more positive or hawkish in March 2021 and peaked in July 2021. After weakening over the next 12 months, sentiment rose again in July 2022. However, these moves might be partly attributable to the recovery attributable to the COVID-19 pandemic, and in addition they reflect the FOMC’s growing hawkish stance amid rising inflation during the last yr or so.

However, the massive fluctuations also point to an inherent flaw within the Loughran-McDonald evaluation: the sentiment scores only evaluate words, not sentences. For example, within the sentence “unemployment fell,” each words would register as negative/relaxing, despite the fact that the statement as a sentence indicates an improving labor market, which most would interpret as positive/harmonious.

To address this issue, we trained the BERT and XLNet models to research statements sentence by sentence.

BERT and XLNet

Transformers Bidirectional Encoder Representations (BERT) is a language representation model that uses a bidirectional encoder reasonably than a unidirectional encoder for higher fine-tuning. In fact, we discover that BERT, with its bidirectional encoder, OpenAI outperforms GPT, which uses a unidirectional encoder.

XLNet, alternatively, is a generalized autoregressive pre-training method that also has a bidirectional encoder, but doesn’t have Masked Language Modeling (MLM), which feeds a sentence to BERT and optimizes the weights inside BERT to output the identical sentence on the opposite side. However, before we feed BERT with the input set, we mask some tokens within the MLM. XLNet avoids this, making it a sort of improved version of BERT.

To train these two models, we divided the FOMC instructions into training datasets, test datasets and out-of-sample datasets. We extracted training and testing datasets from February 2017 to December 2020 and out-of-sample datasets from June 2021 to July 2022. We then applied two different labeling techniques: manual and automatic. Using automatic labeling, we gave the rates a price of 1, 0, or none depending on whether or not they indicated a rise, a decrease, or no change within the federal funds rate. Using manual labeling, we categorized the sentences as 1, 0, or none, depending on whether or not they were restrictive, reserved, or neutral.

We then ran the next formula to generate a sentiment rating:

Sentiment Score = (Positive Sentences – Negative Sentences) / (Positive Sentences + Negative Sentences)

Performance of AI models

Predicted sentiment rating (automatic labeling)

Predicted sentiment rating (manual labeling)

The two charts above show that manual labeling higher reflects the recent FOMC policy change. Each statement accommodates restrictive (or dovish) rates, although the FOMC ultimately lowered (or raised) the federal funds rate. In this sense, sentence-by-sentence labeling trains these ML models well.

Since ML and AI models are typically black boxes, how we interpret their results is incredibly essential. One approach is to use Local Interpretable Model-Agnostic Explanations (LIME). These apply an easy model to elucidate a rather more complex model. The following two figures show how XLNet (with manual labeling) interprets sentences from FOMC statements, with the primary sentence being interpreted as positive/harmonic attributable to the strengthening labor market and moderately expanding economic activity and the second sentence as negative/loose since the Consumer prices fell and inflation was below 2%. The model’s assessment of each economic activity and inflationary pressures appears appropriate.

LIME results: FOMC rate on strong economy

LIME results: FOMC rate on weak inflation pressures

Diploma

By extracting sentences from the statements after which assessing their sentiment, these techniques allowed us to raised understand the FOMC’s policy perspective and have the potential to make central bank communications easier to interpret and understand in the long run.

But was there a connection between changes within the sentiment of the FOMC statements and US stock market returns? The chart below shows the cumulative returns of the Dow Jones Industrial Average (DJIA) and the NASDAQ Composite (IXIC) together with the FOMC sentiment readings. We examined correlation, tracking error, excess return, and excess volatility to discover regime changes in stock returns measured on the vertical axis.

Stock returns and FOMC statement sensitivity readings

The results show that our sentiment scores detect regime shifts as expected, with stock market regime shifts and sudden changes within the FOMC sentiment rating occurring at roughly the identical time. According to our evaluation, NASDAQ could react much more strongly to the FOMC sentiment rating.

Overall, this research points to the large potential of machine learning techniques for the long run of investment management. Of course, how these techniques are paired with human judgment will ultimately determine their ultimate value.

If you enjoyed this post, do not forget to subscribe .

Photo credit: ©Getty Images/ AerialPerspective Works