Yesterday’s infographic caused a minor stir when it seemed to indicate, as opposed to most assessments, that the BOE’s latest communication was hawkish:
This apparent misinterpretation, understandably, prompted some skepticism from our audience:
“Bloomberg & the FT (among others) highlighted in particular this sentence from BoE’s Carney: ‘Monetary policy must continue to balance two fundamental forces–domestic strength and foreign weakness’. All (human) analysts whom I read the comments, concluded about a dovish shift from the BoE, most likely following the downward revision of the inflation forecast in the Quarterly Report. Tracking with algos what matters in central banks subtle communication is a heavy task, and most likely the standard bag-of-words approach can’t apply. Happy to further discuss this thematic.”
Nicolas Boitout’s sharp comment brings up a number of important topics central to what we do at Prattle, but, before discussing those, we’d like to clarify what the latest infographic captured.
The infographic we posted on Thursday was a weekly average of BOE’s communication that was created before the latest communication, and, consequently, whatever mood that was indicated by that communication did not factor into our score.
If that is the case, how did we score it? Slightly dovish:
The prevailing mood from the BOE of late had been relatively hawkish–a trend that Thursday’s marginally dovish communication reversed. In short, our algorithm scored the BOE’s recent release similarly to the consensus evaluation.
But beyond the surprising score, Nicholas brought up another point worthy of addressing here: if the evaluation of central bank communication is a heavy task–and it is–can algorithms really be expected to approach these texts with the appropriate sophistication?
Interestingly, this is precisely the question that catalyzed the creation of Prattle’s methodology. Orthodox sentiment analysis, while advantageous in some respects to human analysis, often operates using rather simplistic dictionaries of positive and negative buzzwords created by data scientists to evaluate texts. When presented with the challenge of interpreting documents replete with highly sophisticated, nuanced language–like those produced by central banks–such a mechanism cannot be expected to produce accurate evaluations.
Seeing the limitations of current techniques–and the importance of central bank communications to the global economy–Prattle’s founders began work on a new method of assessing these market moving documents. We dubbed the finished product our “Central Bank Sentiment Indexes,” and they are built from the ground up to properly interpret the language of central banks.
Here’s how it works.
Prattle’s Indexes are rooted in reference texts. These texts are central bank communications that have led to particular, identifiable market reactions–the type and level of which allow the texts to be expertly scored.
For central banks, this score is an indication of the “hawkishness” or “dovishness” of a central bank’s position on the economy. A hawkish central bank views the economy as strong and growing and, because of this perception, will soon implement contractionary monetary policy–raising interest rates to ensure credit is less available–in an attempt to keep the market from overheating. A dovish central bank believes the economy is struggling and takes the corresponding strategy: lowering interest rates to encourage growth through a climate of easier credit. Directly connected to varying degrees of market reaction, these reference documents are firmly rooted in history–making them an excellent foundation of comparison.
Using these reference texts, Prattle has mathematically linked specific words, phrases, sentences, paragraphs and whole communications to specific market reactions. Expressions linked to hawkish policy and corresponding market reaction are awarded positive numbers based on the level of the response. Conversely, dovish terms are awarded negative numbers. This lexicon of hawkish and dovish expressions is the backbone of our methodology.
With a lexicon in place, it now becomes possible to accurately evaluate current central bank communications.
Aggregating text from all the streams of a given central bank’s communication within whatever time frame is desired, our process then generates a score for the sample in light of the pool of hawkish and dovish communications. The score generated is the average rating of all the hawkish and dovish expressions embedded within the selected documents for a given time period.
This score represents the central bank’s “mood”–i.e. their inflation expectations. We call each bank’s mood their “Index,” and these signals are the only comprehensive, unbiased, and quantitative data on the economic outlook of central banks in existence.