Research Shows Altered Retweets Could Fool Stock Traders’ Algorithms | Coordinated scientific laboratory
The social media realm can be an unpredictable place. Dark and anonymous characters can have considerable influence, especially if they have gradually gained followers and have apparent credibility. Their messages can influence the decisions of various groups, including stock traders.
Social media posts and tweets help influence stock traders’ decisions to buy or sell parts of their portfolios. An example played online and on Wall Street last year. Stock prices soared for GameStop Corp. and AMC Entertainment Holdings Inc., two companies that have not performed well in recent years. Members of the WallStreetBets subreddit pushed the pressure in an attempt to disrupt the status quo on Wall Street and put power in the hands of ordinary people. Business Insider reported that rising stock prices cost hedge fund short sellers more than $1 billion in a single day. Short selling is about making money when the value of an asset drops.
Traders use algorithms that comb posts by keyword to find mentions of stocks and companies and gauge whether stock prices are rising or falling with the economic tide. But not all algorithms are infallible. In fact, there could be an inadvertently embedded weakness in this system which, if exploited, could cost traders dearly.
Dakuo Wang, the head of human-centric AI research at IBM Research, said an adversarial attack via a retweet can expose this flaw in traders’ deep learning algorithms. He explained that using a mathematical calculation, the team can identify which tweet about a particular stock or company is most important when it comes to stock forecasting.
“With this unique tweet, our method will also calculate which word is more important, then we will try to replace this word with a semantically similar word. We want our fake tweet to fool the deep learning model so that the model will predict the opposite result,” Wang said.
For their experience, researchers used a dataset consisting of 10,824 instances, including relevant tweets and numerical characteristics of 88 stocks from 2014 to 2016. Traders used the buy-hold-sell strategy where stocks are bought if they are expected to rise, hold them for one day and then sell the shares the next day. For the attack, in the tweets for example listed in their searches, words such as “filled,‘ and ‘alert’ have been replaced in retweets with negative terms ‘exercised,and “unstable.” Keeping the rest of the text the same, the retweets not only slightly altered the model’s perception of the information, but were also subtle enough to go unnoticed by the human eye.
“Our results show that if marketers truly take social media posting as an input to their deployment models, we can hypothetically attack their model and make a significant loss to their portfolio,” said Yong-XieUIUC PhD student in mathematics.
Whatever ? If an investor started with $10,000, they would lose $3,200 after two years using the long-term buy-hold-sell strategy in a simulated investment.
As revealing as this research is, the researchers wanted to clarify one aspect of it.
“We are attacking [traders’] predictions, we’re attacking their profitability, but it’s not about manipulating the stock price,” Wang said, “and the goal is to expose that risk to the ML and Finance community so people can create more robust algorithms to defend against such attacks”.
The researchers also explained that while this loophole exists and is concerning, it is very unlikely that a random person will cause serious damage to Wall Street.
“Attackers cannot see the results of their attack,” Xie said. “It would be like someone trying to disrupt the market, but not wanting to achieve anything for themselves.”
Instead, the real danger would come from an investment firm trying to attack a competitor by exploiting the vulnerability in the algorithm.
“They can, for example, disable their own model for ten to thirty minutes and then launch attacks in the retweets. Suddenly, all their competitors’ models will be influenced and they can benefit from it,” Wang said.
The team is looking to do more research, but said it was fortuitous that they worked together in the first place. None of the four authors normally works together. Both Wang and Xie attribute IBM-Illinois Discovery Accelerator Institute for cooperation.
“I hope we will have more and more of this kind of work with other students between IBM and UIUC. We want to help the world become a better place, right? But you can’t not do it without a team,” Wang said.