Algorithms Over Analysts | How Prattle Predicted The BOE Hold

  • Friday, July 22, 2016

In July 2016 the BOE held rates, surprising analysts and the market (75% priced in a cut).

Traders using Prattle were not surprised. The sentiment trend data leading up to the meeting pointed to one and only one thing: despite what almost every human analyst was saying, the Bank of England was going to hold rates.

Figure 1: BOE Sentiment Around July 14 Meeting

BOEJuly18_2months_annotated

But how is it possible for a machine to analyze such nuanced, dense text better than the best analysts in the world?

Algorithms over analysts

The basis of Prattle’s correct call was the sentiment data produced by our algorithm.

Prattle has improved on more common sentiment analysis methods by not using a standard dictionary of preset terms with associated positive/negative values to produce its sentiment scores. Instead, Prattle’s lexicon of positive or negative expressions is the outcome of a patent-pending two-step process.

First, Prattle selects reference documents based on historical market reactions, then employs subject-matter experts to hand-pick a subset of those documents for scaling the lexicon appropriately. These reference documents are representative of market response to three types of central bank sentiment: hawkishness (the economy is overheating and contractionary policy might be needed), dovishness (the economy is doing poorly and stimulus might be needed), and neutrality (the economy is steady and no action is necessary).

Second, Prattle uses an algorithm to associate specific expressions–words, phrases, sentences, etc.–within these reference documents with the corresponding market reactions. The outcome: a lexicon of scored expressions built on the historical relationship between central bank language and the markets.

The algorithm uses this lexicon to parse subsequent documents and assign each a numerical score representative of the document’s sentiment in terms of potential market reaction. The scores are normalized around zero and range between -2 and 2, negative numbers indicating dovishness and positive numbers indicating hawkishness. Machine learning is then used to identify, score, and store new expressions in a constantly-expanding lexicon unique to each bank.

Producing comprehensive, unbiased, quantitative, and real-time evaluations, Prattle’s methodology holds several key advantages over human analysis. In the case of the BOE’s surprise rate hold, the unbiased nature of Prattle’s process–as opposed to human analysis–was particularly noteworthy.

A “Surprising” Hold?

One of the key factors in setting market expectations for the July meeting was a speech delivered by BOE Governor Mark Carney on July 5. To many analysts, the speech indicated that a cut (or some other form of stimulus) was imminent.

Prattle’s algorithm saw things differently. The system scored Carney’s speech as fairly neutral to marginally hawkish (0.35) when compared to BOE communications as a whole–and marginally dovish (-0.45) when compared to other communications by Carney.

Why? While Prattle’s algorithm scored the conclusion and the section on financial stability as significantly dovish, every other section of the speech scored neutral…or slightly positive. If you only read the first and/or last portions of the speech (or only remembered those portions), it would be easy to interpret the communication as dovish. But, when considered in its entirety–without bias–the speech appears far more measured.

For instance, many analysts were likely caught up in the dovishness of the following portion…

“The concerns that the historically large current account deficit could be vulnerable to sudden shifts in foreign capital and sharp adjustments in sterling appear to have been borne out. Portfolio flows into UK equities and corporate 2 debt appear to have slowed, and sterling experienced its largest two-day fall against the dollar since floating exchange rates were re-introduced almost half a century ago….In particular, there is growing evidence that uncertainty about the referendum has delayed major economic decisions, such as business investment, construction and housing market activity.”

….while, primed by the previous negative sentiment, overlooked what directly followed it:

“More positively, sterling’s sharp depreciation should, for given foreign demand, provide support to UK exporters, and the sharp fall in gilt yields has meant that all-in corporate borrowing costs actually fell modestly over the course of last week. In addition, financial markets have managed the volatility around the referendum well and have not added to stress.”

Furthermore, analysts were projecting a rate cut long before they read Carney’s speech. Brexit had surprised the world, and there was a reasonable, widespread belief that stimulus would be the logical next move for a BOE coping with the fallout from the referendum. As a MarketWatch article published on June 28th states, “investors are expecting a rate cut at the next Bank of England meeting in two weeks after the surprise vote in the U.K. last week to leave the European Union sent shock waves through global financial markets.” An analyst operating under these assumptions could have taken the dovish portions of Carney’s speech as a definitive easing signal.

This is a classic case of confirmation bias: an analyst’s pre-existing conclusion(s) skew his or her interpretation of new evidence. It may be evident in retrospect, but confirmation bias is devilishly difficult to detect in the moment. While this error is a constant threat to the integrity of human evaluation, it does not, however, plague well-engineered programmatic analysis. As detailed above, Prattle’s algorithm evaluates every central bank communication in light of the entire history of that bank’s language as it relates to market reaction, and, therefore, Prattle scores are an unadulterated, unemotional, and unbiased assessment of that document’s sentiment.

Such analytical objectivity is a clear advantage in any scenario, but it represents a particular necessity to investors. Choosing a restaurant based on qualitative–likely bias–reviews is a non-issue. Choosing which million-dollar position to take on a currency is an entirely different story. Simply put, when it comes to trading millions of dollars, can you afford any but objective analysis?

Trading on Prattle Data

While those relying on market consensus would have missed out, traders using Prattle data could have optimally positioned themselves to profit from this surprise policy decision. Figure 2 illustrates the immediate currency market response the hold caused–and the potential gains Prattle users could have made.

Figure 2: Currency Market Response

GBPUSD_July14_annotated

The announcement caused a strong, 1.36% spike in the pound within 2 minutes of release–a fluctuation Prattle users had a much higher chance of anticipating. By carefully monitoring central bank trends in Prattle data, users can expect to gain similar insights ahead of future market-moving releases.

Conclusion

The shortfalls of traditional human analysis became glaringly apparent when the BOE’s rate hold shocked financial markets. Such analysis was once the only resource of its kind, but Prattle has created an entirely unique–and obviously critical–alternative. Comprehensive, quantitative, and unbiased, Prattle scores allow traders to cut through emotion and empty opinion and make their investments based on hard data–instead of hearsay.

The Signal

Downdload the pdf below:

White Paper Download