Content area
Full Text
Over the last decade, financial markets have witnessed the rise of new systemic risk factors such as the quant meltdown of August 2007 and the flash crashes of May 2010 and August 2015. These events were driven in part by the sudden disappearance of liquidity providers, leading to large price movements. As prices shift, market makers who rely on mean reversion will pull out of the market, triggering a self-reinforcing feedback loop in which prices move even more sharply because of declining liquidity, causing more investors to liquidate their holdings in a panic because of increasing price volatility.
This same period has also been characterized by the rise of alternative sources of data, especially those that can be used to measure quantitatively news content and social-media sentiment. With the use of machine learning tools, it is now possible to process large amounts of text information about an asset and assign real-time investor-sentiment scores based on natural-language processing of these texts.
In this article, we study whether social media and news data can provide us with insights on market panics and manias that are not already captured by existing data. Although these new data sources have the advantage of being available in real time and represent the beliefs of a wide variety of investors, inferring information from them requires separating useful signals about investor sentiment from everyday noise. In particular, we focus on the following questions:
• Given that social-media users represent a small fraction of market participants (and many non–market participants), do these sources contain relevant information about liquidity?
• To what extent can social media give us insights that cannot be inferred from more fundamental sources, such as traditional news feeds?
• Do positive and negative sentiment have asymmetric effects on markets?
• Can news and social media be used to predict future levels of liquidity?
• Can social-media information be used to improve trading strategies?
We answer these questions using three different approaches:
1. We regress several measures of volume and liquidity—log number of trades, log number of quotes, log number of trades outside the quoted bid-ask spread, log turnover, and average spreads—on news and social-media sentiment indicators.
2. We perform a series of intraday event studies on abnormal social-media sentiment.
3. We...