Algorithms at the Associated Press churn out 5,000 earnings articles a quarter
In early 2014, journalists at the Associated Press (AP) wrote articles on roughly 400 US corporate earnings announcements each quarter. Today, using algorithmic ‘robo-journalism’ technology, the media co-operative generates almost 5,000. The effect on newly covered companies’ trading volume and liquidity around the earnings period has been dramatic, according to a new study.
‘In the three days after an article’s publication, we find an increase in trading volume of about 38 percent compared with the sample average,’ says study co-author Ed deHaan, assistant professor of accounting at the University of Washington. ‘We don’t see evidence it affects stock returns. But liquidity is a useful end in itself.’
The research paper also finds that ‘increased trading volume potentially creates a deeper market, and [that] lower information processing costs can result in less information asymmetry, less price protection, and greater willingness to trade.’
For financial market academics, the novel study further erodes the efficient market hypothesis – which is good news for IROs. ‘If you’re a strict believer in market efficiency, there really isn’t a big role for IR,’ comments deHaan. Yet from a practical perspective, its implications are more nuanced. ‘Maybe not all firms want the extra visibility,’ speculates deHaan. ‘Maybe some want a quieter life.’
Managers with bad news to report, or those eschewing a short-term focus, for example, may find this new spotlight on earnings vexing. Either way, more eyeballs on earnings means IR departments will need to gear up for the increased attention.
At the same time, while AP claims its journobots are less prone to mistakes than their human counterparts, the system’s error rate isn’t zero. In 2015, for example, Netflix recorded better-than-expected second quarter numbers.
But the AP report ran the headline: ‘Netflix earnings miss Wall Street forecasts’. The reason for the discrepancy likely involved a concurrent 7-to-1 stock split – data not then incorporated into AP’s algorithmic calculations. AP soon corrected its article, et the nature of the internet means the false story still lingered in corners of the digital firmament.
‘A human journalist might pick up [anomalies],’ says deHaan. ‘But for a robo-journalist the input data is critical. Any time an article comes out, firms want to monitor it to ensure the information is correct.’
Firms may also want to weigh in with AP and other news organizations on the content of their automated earnings reports. ‘Robo-journalism is in its early years,’ notes Elizabeth Blankespoor, assistant professor of accounting at Stanford University.
‘Organizations like AP are exploring possibilities and looking for feedback. There’s an opportunity for companies to say, for example, Perhaps this isn’t actually the best way to structure and communicate the information. You’ve chosen four pieces of [earnings release] information, but there’s a fifth that’s true for all companies in our industry. It’s in this particular location, and would be valuable for investors.’
For now, writing algorithms works best when synthesizing quantitative, structured information, just like that found in earnings releases. But deHaan envisions a not too distant future when artificial intelligence will effortlessly incorporate qualitative information into communications.
If so, it may be time to broaden those skill sets: an Oxford University analysis of occupations at risk of computerization points to a solid future for recreational therapists.
This article appeared in the spring 2017 issue of IR Magazine
