- A new BIS working paper uses deep learning methods to assess the impact of US-China trade sentiment on up to 60 different global equity markets.
- Overall, a worsening of trade sentiment triggers an equity sell-off in the majority of countries, especially those with close trade links to China and the US.
- The stock prices of IT-related firms in China are most responsive to any trade sentiment deterioration.
- Stocks of US-based firms are less affected, but tech firms, as well as firms with a larger revenue from China, are more negatively impacted by a fall in trade sentiment.
Former US President Donald Trump significantly escalated, if not began, the US-China trade war with two executive orders on 31 March 2017. Since then, tensions have caused volatility in global equity markets, but their exact effects are difficult to quantify. A new BIS working paper creates a measure of US-China trade sentiment between 2017 and 2019. By examining the tone of Chinese media outlets, they estimate the impact of trade sentiment changes on global stock prices.
- US-China trade sentiment was at its worst in Q1 2018, while there have been four occasions on which sentiment significantly improved.
- Changes in US-China trade sentiment can explain 10% of the stock price variability of countries closely integrated into the US-China value chain between January 2018 and June 2019.
- In the US, financial, IT, consumption discretionary and industrial sectors were most susceptible to worsening trade relations.
- The tone of posts on social media can explain the majority of stock price volatility attributed to US-China trade relations, rather than traditional media outlets such as newspapers.
Overall, the paper is a strong advocate for using deep learning methods when quantifying trade sentiment using media-based indices. Not only does it allow the researcher to gather information on the frequency of keywords, but it can also capture the keywords’ tone.
A New Take on Media-Based Sentiment Indices
The paper aims to quantify the contribution of trade sentiment in explaining equity price movements. Generally, four types of indices can be used to measure investors’ sentiment: (i) market-based (e.g., trading volumes), (ii) survey-based (e.g., consumer confidence surveys), (iii) search-based (e.g., Google Trends), and (iv) media-based indices (e.g., keywords in newspapers).
The Trade Sentiment Index (TSI) used in the paper is a media-based sentiment measure. Traditional media-based indices focus solely on the frequency of keywords and cannot qualify whether a word has positive or negative meaning. But the TSI uses deep learning techniques to capture the media’s ‘tone’. From this, they are able to determine the extent to which media coverage on US-China trade tensions can explain movements in stock prices around key events.
The TSI is a daily index based on textual analysis of Chinese media outlets between 2017 and 2019. The analysis disentangles information contained within 3.5bn articles from 74,000 media sources. These media sources are divided into newspapers (‘traditional media’) and social media. Social media outlets are then subdivided into web-based sources, WeChat posts and forums. The data is collected from Wisers.
Building the Trade Sentiment Index (TSI)
Media-based indices using counts of keywords face a common problem. Statements such as ‘the trade war intensifies’ versus ‘the trade war is unlikely to intensify’, or ‘the trade war will end soon’ versus ‘the trade war sees no signs of ending any time soon’ are interpreted as the same. Yet their underlying tone and therefore expected impact on financial markets are completely opposed.
In order to evaluate sentiment in the sense of a positive or negative tone, the paper uses a Multi-Channel Convolutional Neural Network (MCCNN) approach. CNNs initially were used in image classification tasks such as object detection and face recognition. But they have since been applied to text classification (see Appendix).
Similar to traditional media-based indices, the authors initially establish which keywords they wish to capture. Then they count the number of articles that contain at least one of these keywords (e.g., trade war, trade conflict, trade tension etc.). This step forms the basis of their first measure, the Trade Count Index (TCI), which simply estimates the frequency at which words related to trade tensions are featuring in the media.
To uncover the tone, the authors apply the MCCNN to each article that contains a keyword. The method returns an estimate for the sentiment ‘polarity’ for each article. The polarity scores can either be positive (implying the keyword(s) used in the article describes an improvement of US-China trade tensions), neutral or negative. The daily sums of the sentiment polarities over all articles form the TSI measure.
The correlation between the TSI and the traditional media-based measure (the TCI) is high across the full sample. Mapping the scores on the same graph, it is clear that US-China trade sentiment in the Chinese media was mainly negative 2018-2019 (Chart 1). Sentiment was at its worst on 23 March 2018, a day after the US imposed steel and aluminium tariffs. Positive sentiment spiked on four occasions, one of which was a day after the former US treasury secretary, Steven Mnuchin, announced an end to the trade war on 20 May 2018.
Source: Paper, page 32
Stock Market Responses to the TSI
The authors evaluate the impact of changes in the TSI on the daily returns at a country, sectoral and firm level. The impact of TSI is statistically significant in more than half of the 60 stock markets they study. This is consistent with the idea that a worsening in sentiment triggers a significant selloff in equity markets for those countries with trade linkages with US and China (Chart 2).
Asian equity markets are particularly exposed. Chinese stocks respond most to changes in sentiment: a one standard deviation deterioration on in the trade sentiment index leads to a drop in Shanghai equity prices by 0.4%. The results also show that US-China trade relations impact Hong Kong and Japanese stocks. Interestingly, trade sentiment impacts US equities by just a sixth of the magnitude that it affects China. This may be partly because the TSI represents the tone in Chinese media, not US media.
The main driver of stock price susceptibility is a country’s role in the China-US value chain (provided in the OECD’s Inter-Country Input-Output database). The more important a country’s role is in this chain, the more sensitive its stock market returns are to a deterioration in the TSI. For countries with exposure to the value chain above the median level (Chart 2, red), changes in US-China trade sentiment can explain 10% of stock price variability between January 2018 and June 2019.
Source: Paper, page 33
Which Sectors Are Most Sensitive to US-China Trade Relations?
For a sectoral analysis, the authors home in on the US and China. Trade tensions tend to negatively affect sectors across the board. In China, IT related sectors are most responsive to a deterioration of the trade sentiment. These sectors are also the most affected by tariffs. Smaller and younger firms are at more risk, too, but utilities and financials are the most insulated. The key determinants of sensitivity are the degree to which the sector is directly impacted by tariffs and the proportion of revenue it receives from the US.
In the US, financial, IT, consumption discretionary and industrial sectors are most at risk of worsening trade relations (Chart 3). It is no surprise IT also emerges as one of the most affected US sectors, since some big tech firms located in the US were at the centre of the trade tensions. Similar to China, US firms with larger revenue from China are also more exposed to a deterioration of Chinese trade sentiment.
Source: Paper, page 36
The Importance of Social Media
Lastly, the authors deconstruct the TSI measure into one constructed purely of social media articles and another from traditional media. Interestingly, TSI-social-media is always significant in all specifications, while the TSI-traditional-media is not. Specifically, the tone extracted from social media explains 90% of the 10% stock price variability seen in highly exposed countries (countries with exposure to the value chain above the median level). This highlights the significance of social media content in investment decisions.
The paper overcomes the disadvantages of traditional media-based sentiment studies by using deep learning techniques to analyse big unstructured datasets. Examining the tone of Chinese media outlets, they find that a worsening of US-China trade relations can have significant and wide-reaching implications for equity prices.
These findings are in line with other estimates of the US-China trade war. Literature shows that a 10-percentage point increase in tariffs has reduced global GDP by around 1% after two years. This will have had negative implications for businesses and consumers. Post Trump, expect the US trade stance to remain tough, but it is unlikely equity prices will be as responsive under a clearer and less chaotic strategy.
In a neural network, neurons are fed inputs which have weights assigned to them. These inputs can be a sample of external data, such as images or documents, or they can be the outputs of other neurons. After receiving inputs, the neuron produces a single output, which can be sent to multiple other neurons via connections. Each connection is assigned a weight that represents its relative importance.
Information is fed through neural networks until it reaches the output neurons, whose output accomplishes the task they are assigned to do (such as recognising an object in an image, or a keyword in a text). To find the output of these neurons, one needs to take the weighted sum of all the inputs, weighted by the weights of the connections from the inputs to the output neuron. This weighted sum is then passed through an activation function to produce the output (Chart 4).
Convolutional Neural Networks
A convolutional neural network is different from a neural network because it operates over a volume of inputs. Each convolutional layer convolves an input, trying to identify a pattern or extract useful information via filtering, and passes its result to the next layer. Each convolutional layer gets smaller because neurons in the subsequent layers receive inputs from only a restricted area. This constant shrinking of layers means more filters can be applied to each layer, allowing for greater depth and better extraction (Chart 5).
For textual analysis, the deep learning method can be seen as imitating what the human brain does with this example: ‘Yuor ability to exaimne hgiher-lveel fteaures is waht aollws yuo to unedrtsand waht is hpapening in tihs snetecne wthiout too mcuh trouble’ (example courtesy of BIS).
Amstad M., Gambacorta L., Chao H., Xia D., (2021), Trade sentiment and the stock market: new evidence based on big data textual analysis of Chinese media, BIS Working Papers, (917) https://www.bis.org/publ/work917.htm
Note, the TCI measure is simply a count of articles with keywords. Given that sentiment was mainly negative 2018-2019, the authors multiplied the daily sums by -1 for the sake of comparison. The TSI measure is generally less negative because it is a series containing -1, 1, and 0s. ↑
Sam van de Schootbrugge is a macro research economist taking a one year industrial break from his Ph.D. in Economics. He has 2 years of experience working in government and has an MPhil degree in Economic Research from the University of Cambridge. His research expertise are in international finance, macroeconomics and fiscal policy.
(The commentary contained in the above article does not constitute an offer or a solicitation, or a recommendation to implement or liquidate an investment or to carry out any other transaction. It should not be used as a basis for any investment decision or other decision. Any investment decision should be based on appropriate professional advice specific to your needs.)