Wednesday, April 13, 2022

Self-organizing fraud (or plain old fraud)?

Russ Mitchell writing for the LA Times:
Whether Twitter bots are being deliberately programmed to manipulate stock trading is among the questions that Kirsch and his research assistant, Moshen Chowdhury, are trying to answer.
A Twitter bot is a fake account, programmed to scour the social media site for specific posts or news content — Musk’s posts, for example — and respond with relevant, preprogrammed tweets: “Tremendous long term growth prospects” or “Why Tesla stock is rallying today” or “Tesla’s Delivery Miss Was ‘Meaningless.’” The bots can also be programmed to send nasty or threatening messages to company critics.

Kirsch and Chowdhury collected and reviewed Tesla-related tweets from 2010, when the company went public, to the end of 2020.

Over that period, Tesla lost an accumulated $5.7 billion, even as its stock soared and Musk became one of the richest humans on the planet; his net worth is estimated at $275 billion. Operational results can’t justify anything close to the company’s $1-trillion market value, based on any kind of traditional stock-pricing metric.


Using a software program called Botometer that social media researchers use to distinguish bot accounts from human accounts, the pair found that a fifth of the volume of tweets about Tesla were bot-generated. That’s not out of line with giants like Amazon and Apple, but their bots tended to push the stock market and tech stocks in general, with those companies as leaders, but not focus on any particular narrative about the companies.

While any direct link between bot tweets and stock prices has yet to be determined, the researchers found enough “smoke” to keep their project going.

Over the 10-year study period, of about 1.4 million tweets from the top 400 accounts posting to the “cashtag” $TSLA, 10% were produced by bots. Of 157,000 tweets posted to the hashtag #TSLA, 23% were from bots, the research showed.


The researchers are looking at the timing of the tweets and options activity in the overnight stock market, among other factors. One big unknown: whether the bots are the work of entities with a direct financial interest in Tesla.

I'm out of my area of expertise here but here are a few points that worry me about all this. 

1. Though we don't know the exact magnitude, we can be reasonably sure that bots are a relatively small part of the concerns with Tesla's valuation and Tesla is a relatively small part of the concern with bots. That does not mean that this is a trivial issue. When you have problems this big, even a small portion can still be pretty damned large in absolute terms. 

2. Bots and the larger problem of fake accounts and falsified metrics are not by any means limited to Twitter. The company formerly known as Facebook has a huge problem with fake accounts and fraudulent ad metrics. I'm sure there are similar issues with Snapchat. Even Tik-Tok where it's hard to imagine a big bot presence certainly has issues with cooking data such as viewership numbers. 

3. Social media metrics have gotten to be a huge part of data-driven systems and I assume social science research. I know from the business side they come up all the time. Our metrics, models, and algorithms are all vulnerable to these kinds of fraud. The big question at this point is how bad the problem has gotten. 

4. I wonder if our mental models for thinking about the next generation of fraud are sufficiently complex and open to unknown unknowns. In this case, my guess is that these bots come from multiple sources. Possibly Tesla or people working covertly but directly with the company. Investors looking to pump and better time the spikes. Fan boys wanting to amplify their displays of love and loyalty for Elon and their hatred of his enemies (particularly when those enemies are female journalists). We could easily see something that looks like coordination but without any direct communication, and as bots grow more sophisticated, their behavior will grow more complex and presumably more difficult to predict and detect. Which brings us to... 

5. Bots are a moving target.

No comments:

Post a Comment