Opinions and emotions expressed in online content, from news articles, through blogs, forum posts, social media such as Facebook or LinkedIn, to tweets, can provide a sense of market sentiment that can reinforce or even anticipate fundamental indicators, thereby helping make investment decisions, argued Robertson, NN Investment Partners senior PM, in a conversation with FSA.
“Everyone is poring over the same financial statements, macro forecasts and central bank policies – there’s no real edge to be gained there,” he said. “By incorporating new data sources and new ways of looking into data we can give ourselves an edge.”
In order to incorporate this insight, NNIP has established a partnership with the US company Marketpsych, which specialises in analysing the emotions of financial markets (the co-founder and CEO is a psychiatrist).
Marketpsych collects the continuous stream of data from social media, news feeds, blogs and other online content. It uses artificial intelligence algorithms to analyse the sentiment each tweet, post or article expresses around its subject. It determines whether it is positive or negative, and also assesses its strength.
The process of training the algorithms – machine learning – is laborious. Robertson recalls that, initially, more than 60,000 articles needed to be assessed by human analysts in order to teach the software to recognise patterns in the data.
The work is never finished. “It has been an ongoing process of assessment of the algorithm, tweaking it when necessary and feeding that back into it so that it can learn to do a better job,” he said.
Data from sentiment
Marketpsych compiles the sentiment data into a variety of indices on a range of subjects, such as regions, countries, currencies, equity sectors, individual companies and many others. NNIP’s multi-asset team uses these indices as inputs into its own decision-making process.
“We create our own indices based on this data,” said Robertson. “The signals from both the fundamental side and from the big data side are combined into an investment scorecard.”
The firm uses various scorecards for different purposes, such as asset class or geography.
The weight of the sentiment data in these scorecards varies between 15% and 20%, according to Robertson.
Human judgment is necessary in the process to take into account events that cannot be easily modeled, like geopolitical developments. “You can never model for [whether] North Korea will launch a ballistic missile test this weekend,” said Robertson.
In practice
“We’ve seen that the big data we’ve incorporated is very good at picking out turning points and extremes,” he said.
In early 2017, the team’s scorecards were generally positive on commodity markets, despite some signals of credit tightening in China, said Robertson.
As part of its sentiment analysis, the team monitored political risk and emotional sentiment indices around commodities.
“Around mid-April we saw a really sharp deterioration in both of these signals in our scorecards, at the same time the overall scorecard was still giving us a positive view,” said Robertson.
The sentiment shift reinforced the misgivings the team had already had, based on their fundamental analysis, and a decision was made to reduce commodity exposure, according to Robertson. The Bloomberg Commodity Index subsequently fell from 86.3 on 13 April to 82.0 on 9 May.
Today, the data is collected across the English-language internet space. Robertson agreed that China’s enormous social media space was an obvious blind spot.
Future developments might incorporate more different types of data, such as satellite photography, but more can be done with the data already available. NNIP is looking to partner with researchers at academic institutions to explore that avenue, he said.
Performance of NN First Class Multi Asset Premium Fund v the sector
All fund NAVs have been converted to US dollars. Note that funds in this chart may be denominated in currencies other than the US dollar.
