Headline risk is increasingly a source of short-term volatility, how well can sentiment analysis manage this risk? Chris Hall investigates.
We live in angry, volatile times. Twitter offers an outlet for these feelings; septuagenarian fantasy author Philip Pullman’s recent tweet regarding a prime minister swinging from a lamppost is indicative of a more general sentiment in politics today.
However, we do not need to rely on anecdotal evidence to support it, we can quantify changes in sentiment toward companies, currencies or indeed governments, as expressed in news and social media. Using analytical tools that deploy natural language programming (NLP) methods to interpret the meaning of written content and compile aggregated scores or indexes, reported sentiment is a measurable dynamic.
Sentiment analytics have been used by quantitative hedge funds for more than a decade and are enjoying increased demand from the wider buy-side community. This is partly driven by investments in data science by fundamental asset management firms, seeking new sources of alpha to regain market share from passive investment products.
“The increased recruitment of data scientists on the buy side has mixed results overall,” says Tom Doris, chief data scientist at Liquidnet, which recently bought Prattle, a developer of NLP-based analytics. “It’s not easy to generate systematic signals that can make a practical difference to a portfolio manager running a mandate across hundreds of stocks and billions of dollars.”
Sentiment analysis is a good fit for those dipping their toe in the alternative data pool because of the sheer size of the available data set and the maturity of the solutions, which have been regularly refined to improve their ability to generate actionable signals. For example, analytics provider MarketPsych allows users to track a variety of emotions beyond positive and negative changes in overall sentiment (e.g. anger, fear, optimism) due to their differing impact on asset prices.
These emotion-based indices certainly confirm and indeed quantify rising anger levels expressed in news and social media. MarketPsych’s measures of UK and US anger have trended gently up since the firm’s records begin in 1998. But more recently the firm’s CEO, Richard Peterson, noted patterns in anger levels toward particular stocks. Specifically, the shares of firms that attracted the ire of investors and the wider public, e.g. for causing environmental damage, would quickly become so toxic that many investors would dump or cut their stakes without consideration to longer-term value. This created opportunities for other investors to pick up cheap stock in anticipation of the strong rebound. The ‘anger reversal’ phenomenon could be observed in almost all major equity markets.
“We saw that there was a global anger upsurge (against certain stocks), but wanted to harvest the profits that could be made from over-reactions by feeding it back into causes that could have a positive impact, such as mental health charities,” explains Peterson. “We started a small pilot fund in January and by July it was beating the S&P 500 by 3%, up 21% overall”.
Calm and considered reactions
For others, over-reaction to anger is a hazard best avoided. Empirical Research Partners, a New York-based boutique that mixes quantitative tools with fundamental analysis, is a long-term investor and user of sentiment analysis.
“Sentiment can be a useful addition to a quantitatively focused investment process,” says Empirical Research Partners’ portfolio strategist, Rochester Cahan. “Whether driven by fundamental or quantitative analysis, investors spend a lot of time focused on numerical inputs but their models can lack ‘read’ on the mood surrounding a stock. For us, news sentiment is an obvious place to try to add some colour to quantitative data points.”
The firm combines inputs from a number of analytics providers to produce a proprietary ‘Media Sentiment Indicator’, which applies weights to inputs to generate a signal geared toward longer-term, low-turnover investors. The indicator does not ignore social media or newsflashes, but these are not as heavily weighted as other inputs, notably ‘full story sentiment’ and ‘earnings sentiment’.
Nevertheless, Cahan asserts that sentiment analysis can offer long-term investors an earlier signal than traditional models in several circumstances, including identifying genuine turnaround stories before other investors. “Sentiment can be a useful guide for deciding when to get into a value stock. It doesn’t matter how long a stock has been cheap; what matters is buying as sentiment turns positive. Sentiment moves before momentum so is a better catalyst for jumping into a value stock,” he explains. Conversely, a stock which has all the hallmarks of impending failure – e.g. diminishing earnings quality, high capex levels, counter-productive M&A activity, etc – is still too early to short if sentiment still has not turned against it.
Peter Hafez, chief data officer of sentiment and big data analytics provider RavenPack, suggests market-moving headlines are an inevitable function of wider trends and as such can be managed. “In certain markets, notably US equities, asset prices are more headline-driven than by the information that you need to dig out of the body of an article. This does not necessarily change the way in which we address news. We have always put a lot of emphasis on headlines in our analysis, but we do allow users to distinguish between headline and story body if they wish,” he explains.
Selecting the right Input
Indeed, some users have a preference to consume only headlines. “It’s a question of building efficient models. For certain more prominent asset classes or companies, you can get a much cleaner signal from the headline versus the body. But there is always a trade-off between the strength and the clarity of the signal,” says Hafez, noting that a bigger challenge for providers is to ensure both context and sentiment are accurately reflected in the outputs of analytics tools, in response to the need to incorporate a wider range of non-premium, web-based sources with minimal levels of human curation.
Peterson acknowledges the potential pitfalls of incorporating sentiment signals into different types of investment process, adding that MarketPsych is employing machine learning to improve the firm’s understanding of interaction between sentiment and price movement.
“Impacts on prices are subtle and the results might not be as systematic as you would need them to be across an entire portfolio. Sentiment often works best at the extremes or when there are dramatic changes, but it does not necessarily work consistently every day,” he says.
Liquidnet’s Doris says exploration of NLP to augment investment processes is still in its early stages. “We’re only scratching the surface,” he says. “One area with a lot of potential is using NLP to analyse analysts’ earnings calls with CEOs for tone or language in those interactions that could identify an inflection point in sentiment before it is reflected in any analyst’s report. NLP offers the ability to systematise and scale those kinds of alpha-generating insights.”
The catalyst for Philip Pullman’s now-deleted tweet was the announcement on 28 August of the planned suspension of parliament by the UK government, loudly denounced as an ‘outrageous’ affront to Britain’s unwritten constitution. On the morning of the announcement, sterling fell sharply to 1.2155 against the US dollar before recovering to close at US$1.2209, moderately down on its opening price (US$1.2288). In parallel, MarketPsych’s UK news and social media sentiment index fell sharply to -0.2252 (the index’s full range is 1 to -1) having been in positive territory the previous weekend as UK premier Boris Johnson attended his first G-7 summit in Biarritz. The markets, it seems, do not always over-react to headlines.
“The anger and disappointment about Brexit is already priced in for the most part,” says Peterson. “There’s potential for a gradual GBP rally once Brexit gets underway, since this is a ‘sell on the rumour, buy on the news’ type of pattern.”