Stock exchange announcements have long been ripe for what we used to call computer-assisted journalism and what we can now process through generative AI. They’re clearly formatted and relatively easily parsed. They also matter.

Several decades ago Reuters trialed an automated extraction of journalistic value from stock exchange filings and, of course, it has come a long way since then.

A small-but-innovative New Zealand business news site, BusinessDesk, has led the way in the contemporary use of generative AI in that country, and, sensibly, it has focused initially on historic and current announcements to the stock exchange.

By running nearly 3,000 announcements through generative AI, the site has created a historic database that underpins the value of stock data and news its readers rely on to trade. It also cuts to 30 seconds from half an hour the time it takes to create an article based on a statement.

I suspect it is only the start, but it seems to me a sensible place to start: Stock exchange filings have a consistent format with little room for error, the corpus can be tested in chunks to confirm accuracy, it can be dreary for humans (though stories are often hidden in them).

I talked to Matt Martel, the publisher and general manager of BusinessDesk, about his decision to launch this experiment in Artificial Intelligence, which rapidly became a publishable product. BusinessDesk was acquired two years ago by INMA member NZME. 

This interview has been edited for clarity and brevity.

INMA: How are you collecting and extracting the value from this and converting it into publishable stories?

Matt Martel: We launched NZX [New Zealand Stock Exchange] data on the Web site two years ago, and we’ve stored every announcement on a server since then. In December, we wrote code that allowed us to push anything we wanted to ChatGPT and for it to return as a JSON format that we could automatically ingest into our CMS.

There were 2,700 announcements sitting in our database on day one, and we went through and summarised those as news stories. So that was 2,700 articles in an hour. We weren’t happy with some of the results so we tweaked the prompt and did it again. We then attach each article to our stock market graphs and show which announcements have moved the market.

INMA: Presumably because you’re putting specific content into ChatGPT, the risk of hallucination is much less and so you can rely on it being accurate?

Matt Martel: That’s right, to an extent. If we send an announcement that’s missing a full stop at the end, it will send something back that fills in the rest of that sentence with hallucinations … . It’s still trying to hallucinate and build things in if we don’t have a full stop at the end of the last paragraph.

When we start asking it to do things such as write an article as a business reporter, or to put it into Guardian style, or to do things which are more than summarise the article, then it starts to just do too much. All we wanted it to do was a very basic job to summarise this announcement as a news article. And it does a really nice job of that.

INMA: What has this saved you in terms of editorial resources?

Matt Martel: At this stage, this is just for the NZX data so our journalists don’t need to be involved in editing and moderating the content that comes back. But we absolutely keep an eye on it. We read everything but after publication. We were running the system on our test servers and tweaking it for months before launch, so we were pretty confident about it. We are asking it to take a single piece of content from an NZX announcement and then create a news article. AI is really good at that.

But we’ve had to learn how to work ChatGPT and now Bard. This has our data editor, Andy Fyers, learning how to create the correct prompts for the articles. The next stage is to get ChatGPT to start writing headlines for us, which will be live shortly.

INMA: And you’ve associated those historic items with movement in the stock that you can now associate with the charts on each stock, right?

Matt Martel: We’re hoping to show which announcements have moved the market, and our goal is to also append our own stories to the charts. One day we want to be big enough to show that our journalism also moves markets.

INMA: What other uses are you imagining?

Matt Martel: Where it will be interesting is for fact-checking, where we can say this fact looks like it’s not correct because it differs from what we’ve previously published. That also then allows us to create category pages or topic pages so any proper noun can then become a topic page, which will help with search discovery.

INMA: I see you label whatever you have used Artificial Intelligence to create.

Matt Martel: We’re very clear when we have used AI and it’s almost as simple as that. We need to declare whatever we do.

INMA: Matt, I think you have a clear idea why you need to leap on AI train now?

Matt Martel: We’ve seen this from the demise of newspapers and the rise of the Internet. It’s going to happen anyway. So you’re probably going to be involved. I’d rather be at the forefront of it and understand it than in three or four years realise that our competitors have done the same thing or someone new has come along and destroyed our business model.

If you’d like to subscribe to my bi-weekly newsletter, INMA members can do so here.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .