Fed communication is both a source of guidance about the path of monetary policy, and a direct driver of economic conditions and financial markets. Therefore, it’s a crucial source of information for investors, as US rates exert an influence well beyond US borders.

Over the past 30 years, the central bank has increasingly delivered guidance and policy signals to financial markets via its communication: post-policy decision statements; economic forecasts; ‘dot plot’ interest-rate projections; policy-meeting minutes; and press conferences.

However, interpreting these communications has traditionally been very subjective. The sheer volume, as well as the need to account for the different views of the various individual policy makers, make things even more tricky.

Amid rapid advances in technology, and given the importance of getting it right, we decided to use natural language processing (NLP) techniques to scan and rate US central bank communications.

These techniques provide a consistent way of interpreting the available data.

Read the full paper

What’s natural language processing?

According to IBM, natural language processing ‘refers to the branch of computer science – and more specifically, the branch of artificial intelligence or AI – concerned with giving computers the ability to understand text and spoken words in much the same way human beings can’.

Using NLP techniques, we developed an analytical tool that categorises Fed communications on a scale from hawkish (a bias to tighten monetary policy) to dovish (a bias to ease monetary policy). We’ve called this the ‘H-D indicator’.

We believe the H-D indicator represents a more objective and systematic way of spotting the turning points in Fed policy, as well as anticipating movements in financial markets.

Natural language processing in practice

• Prepare texts: remove common words (e.g. ‘it’ or ‘and’); shrink remaining words to common linguistic root; define a set of ‘hawkish’ and ‘dovish’ words based on established machine learning literature.
• Scan texts: take the 200 documents with most ‘hawkish’ and ‘dovish’ scores (the labelled set); manually confirm; randomly split the labelled set into ‘test’ and ‘train’.
• Build ‘sentiment classifier model’: ‘train’ on the labelled speeches set using machine learning algorithms; model assigns ‘p-values’ – the confidence that each of the remaining documents is either ‘hawkish’ or ‘dovish’.
• P-values form the basis of H-D indicator: individual words treated differently depending on context; manual intervention where appropriate; analysis of history of individual speakers as part of context.

Why it works

We inputted more than 2,000 Fed communication texts – including Federal Open Market Committee (FOMC) statements, policy-setting meeting minutes, and Board of Governors’ speeches since 1996 – into our tool and ran an analysis.

The H-D indicator picked out key moments in recent economic history. For example, hawkish policy before the 2007/8 global financial crisis, that turned dovish as fears gathered; dovish policy in response to the pandemic, which turned hawkish earlier this year on inflation concerns.

The indicator also anticipated changes in market interest rates – with a hawkish tone generally driving rates higher and vice versa for a dovish tone. We found it works best with shorter-dated government bonds, such as two-year Treasury paper.

Where are US rates heading?

Using the H-D indicator, alongside our proprietary US Financial Stress Index, we can say with some conviction that investors shouldn’t expect the Fed’s tone to get dovish any time soon.

This is despite the fact that we’re currently in a period of heightened financial stress – a time when you would normally expect the central bank to ease monetary conditions.

However, this bout of financial stress is being deliberately caused by the Fed in its fight to bring down high inflation.

Indeed, our indicator isn’t flagging a meaningful Fed pivot in monetary policy in the near term. That’s why we can expect US interest rates to be ‘higher-for-longer’ – with all the associated implications for global economies and financial markets.

Final thoughts

Over recent decades, central bank communication has become an independent and important policy tool in its own right.

Communication that changes market expectations around the future path of policy amounts to a change of policy today.

That’s why systematically classifying the tone of such communication is crucial to understanding how the policy environment is evolving.

We think we’ve found a way to do so by using machine learning. The process reduces the degree of subjectivity in analysis, as well as handles the large volume of material available.

As a result, we feel confident that we’ve found a reliable way to help investors navigate the waymarks that indicate impending shifts in US monetary policy.