Dr Svetlana Borovkova: Words matter - News sentiment analysis for financial markets

8 mei 2020

By Dr Svetlana Borovkova, Probability & Partners and VU Amsterdam

Part II

In the previous column, we argued that news is an important driver of financial markets, but also noted that humans cannot quickly read and process all relevant news for a large diversified portfolio. Fortunately, modern AI techniques such as Natural Language Processing, powered by fast computers, can help us to quantify the sentiment, or “tone” of thousands of news items in real time.

However, even when all this clever news interpretation by NLP is done, the work of quantitative analysts is just beginning. News sentiment data is very noisy (due to diversity of opinions and/or interpretation error), there is still a huge volume of it - it is truly “big data”, due to high frequency of news occurrence, multiple news/media sources and large number of companies, commodities and other assets about which news appears. So the analysis task is to extract a clean, interpretable signal from all these noisy data and to aggregate sentiment data over many individual assets into sentiment signals that reflect the overall “mood” of the market participants about e.g., sectors of the economy, regional or worldwide stock and commodity markets. 

One such useful sentiment signal was developed by us after the financial crisis of 2007-2008. Its name is SenSR (Sentiment-based Systemic Risk indicator), and it measures media sentiment about the so-called Systemically Important Financial Institutions (SIFIs). All news about big, important banks is continuously monitored, combined with risk measures of these banks (such as leverage or size of their balance sheet) and aggregated into a single number: SenSR, which tells us how positive or negative news is about the global financial system. When this risk indicator was compared to other, traditional systemic risk indicators – such as VIX, Libor-OIS spread or famous S-RISK measure by Robert Engle – it turned out that it was able to signal the financial system’s distress up to twelve weeks before any other available indicator.

Equipped with this SenSR methodology, we can monitor sentiment for all other sectors and not just for the financial sector: for tech, energy, materials and others. It turns out that these “sector SenSRs” are also excellent in predicting collective price movements (and especially price declines) of stocks in these sectors. For example, a sector rotation strategy based on sector sentiment can dramatically decrease a portfolio’s volatility and enhance returns even in very good economic times by at least 2% per annum.    

Going one step further, we can now generate and monitor sentiment signals reflecting the “mood” about large stock markets such as those in the US, UK or EU. We monitor news sentiment related to these stock markets. At the end of January this year, we noticed and published online our observation of a consistent decline in sentiment, and, as of January 27th, our Probability Sentiment Indicator (PSI) for the S&P500 signalled a rapid and sustained decline, triggered by sentiment analysis of stock-related news. After that date, it took more than three weeks for the S&P500 index itself to collapse on February 19. After that, stock prices and sentiment continued to tumble down in sync.

Retrospective analysis of the PSI scores showed that this sentiment decline was related to the anticipated effect on business of coronavirus that impacted the PSI scores – however, in the sentiment analysis, no coronavirus-related or any other event-specific terminology is used. The decline in sentiment predated the actual plunge in S&P500 by more than three weeks, as shown in the graph below. The sentiment for the US stock market has hit its all-time low on 14^th April but has slightly recovered since, together with the recovery in stock markets.

Over the last decade, we have demonstrated that sentiment analysis can capture the “mood” and subsequent actions of market participants in a timely fashion. Sentiment indicators such as SenSR are effective early risk indicators for financial markets. The coronavirus shock provides further compelling evidence of the power of sentiment analysis in signalling significant market downturns. This is remarkable, since these indicators are based purely on “soft”, previously unquantifiable perception-based information rather than on “hard” financial measures, such as companies’ fundamentals.

It seems that a collective view of the news media (the “wisdom of crowds”) is a better indicator of the state of the financial markets (and an important driver) than fact-based measures and analysts’ reports. “Things do not pass for what they are, but for what they seem”, as the 17^th century philosopher Baltasar Gracian famously said.  Sentiment analysis allows us to better judge how things seem: based on more extensive inputs than any human could encompass – and then make our decisions accordingly.

Probability & Partners is a Risk Advisory Firm offering integrated risk management and quantitative modelling solutions to the financial sector and data-driven companies.