‘The Algo Is Agnostic’: Here’s What Happens When You Run NLP On 30 Years Of Fed Communications

In the post-crisis world, market participants have been forced to become experts at deciphering economist doublespeak – traders are compelled to perfect the art of parsing central bank communications.

The two-way communication loop between markets and central banks defines the contours of reflexive policymaking in a world where every turn of phrase matters. As Jerome Powell learned the hard way on October 3 when he uttered the words “long way from neutral”, one errant phrase is all it takes to start tipping dominoes.

Forward guidance has never been more important as a policy tool than it is now. With developed market central banks running woefully low on ammunition and the global economy teetering, guiding the market with credible pledges of accommodation can help alleviate the need to fire scarce bullets (i.e., to the extent it’s credible, forward guidance can substitute for rate cuts and asset purchases).

Of course, from the market’s perspective, this is tasseography (tea leaf reading). Often, central bankers will attempt to convey one thing, only to have the market read it another way, forcing policymakers to clarify in subsequent remarks. Usually, consensus is hard to come by in real time when it comes to deciding what Mario Draghi (for instance) is actually saying during a press conference, or what John Williams (to use Thursday’s dubious misstep as an example) is attempting to convey in a speech.

Simply put, parsing central bank communications is more art than science. Or at least it was.

In a note dated Friday, BNP applies natural language processing to decades of Fed communications in an effort to transform a tedious, imprecise qualitative exercise into an efficient, quantitative process.

“Scanning through decade’s worth of unstructured text documents is laborious for humans, who generally read at a pace of 250-500 words per minute”, the bank notes, adding that “since 1993, the average word count of each of the Fed Minutes is roughly 50,000 words [and] with 213 Fed Minutes to scan through, it would take a (very fast) reader over 350 continuous hours to read these documents”.  By contrast, BNP’s algo can analyze the same documents and score them all in under a minute.

How does it work? Well, in simple terms, the bank says the algorithm “identifies key words and phrases associated with positive or negative sentiment, and then aggregates sentence and phrase-level text into an overall score of sentiment”.

Below, find a visual which shows the series produced by the bank’s natural language processing algo when it was applied to Fed minutes. As you can see, the bank annotated the chart with key events (readings less than zero are outright dovish and the more positive the reading, the more hawkish):

(BNP)

What you want to note about this approach is that while some will invariably suggest it’s just another example of analysts with too much time and too much computing power on their hands, it’s potentially useful if you’re in the camp that believes carbon-based market participants have a tendency to read their own biases into policymaker communications. “The algorithm is agnostic”, BNP writes, noting that “natural language processing does not begin with any subjective preconception about what Fed communication should say, nor does it fit a subjective and evolving view”.

In addition to that rather important attribute, it’s also far more efficient and is capable of generating a time series, which the bank calls “a particularly useful addition when trying to model the impact of the Fed’s sentiment on asset performance”.

One finding from the bank’s initial stab at this is that the algo spits out the most dovish score since July 2016 when it’s applied to the December 2018 Fed minutes. You might recall that when those minutes were released, many observers suggested they conveyed a far more dovish tone than what came across in Powell’s December presser. BNP’s algo puts a score on that.

(BNP)

“In December 2018, markets sold off sharply, partly in response to the Fed rate hike and dots projecting continued tightening through the year”, BNP recounts. “One month later, the December FOMC meeting Minutes struck a more dovish [and] we noted a ‘shadow change’ in the balance of risks”, the bank continues. The algo quantifies things.

In the full note, BNP attempts to roll all of this up into a trading model.

As ever, the x-factor is probably the reflexivity inherent in central banks’ communications with the market. You can capture the tone of a speech or account of a meeting and you can assign a score, but given that subsequent market behavior will invariably impact the evolution of future communications which will, in turn, affect markets, the continuous feedback loop is quite difficult to model.

As Deutsche Bank’s Aleksandar Kocic put it back in 2015, “the Fed’s communication strategy… is an equivalent of what in theater context is referred to as Removing the fourth wall whereby the actors address the audience to disrupt the stage illusion — they can no longer have the illusion of being unseen”. When the fourth wall comes down, the audience is no longer watching a self-contained narrative with an unalterable course. Rather, the market is forced to observe itself from another angle, “as an observer of the observer of the observers”. Run natural language processing on that and see what comes out.