Stuff. This is the sound of a busy trading floor in Wells Fargo skyscraper in New York City. It's filled with stockbrokers who are researching companies, writing reports, chatting with investors, and processing their trades. These are classic and highly paid jobs that have been around for more than a century. Imagine if one day those jokes went away. Yeah, this is a scenario that Ken Center, a veteran Internet analyst, one of those classic brokers, is forcing himself to think about.
Canada developer sidekick named Brian Healy have spent months building an artificially intelligent robo analysts that performs a lot of what Ken currently does for a living. Ken calls the system ERA, which stands for Artificially Intelligent Equity Research Analyst. It's an elaborate warning to investments, firms and banks that they should learn and embrace AI to catch up with Internet giants like Google, even if it means automation destroys
some lucrative all street jobs. And it's a broader wake up call for all of us. Artificial intelligence is here and it's changing complex, high paying jobs for good. We better be ready. M Hi, I'm brad Stone, I'm Julie Verge and I'm Alice de Ba, and this week Undercrypted, we're taking you inside Ken's quest to build an ever evolving and always learning software program. This program could eventually put Ken out of a job, or at least do several parts of his job a lot better than him.
Stay with us, So, Alian, Julie, you guys recently went to go meet the brains behind this project. We did. It was a cool October morning. We picked up an audio recorder and headed down a few blocks to Wells Fargo's office on Park Avenue. And Ali, You've known Ken the stock analysts for a while, Yeah, all the way back to the days when I was a reporter voters in the Wall Street Journal. I'd often call him for
advice on tech companies and news. He's tall, with wavy, sweep back hair, and when we met him, he's wearing a crisp suit and tied like he was born in the outfit. Brian, on the other hand, comes from a tech background. He helped build Alexa, the digital at home assistant whose voice you hear on Amazon's Echo speaker. Brian has a goatee and no nonsense, close cropped hairstyle. And I have to admit he didn't look so comfortable in that suit that he was wearing. It sounds like an
odd couple. How did they meet? It actually kind of started out by a chance. Ken was organizing an AI conference in two and had reached out to Brian's US to be a speaker. He was busy, so he recommended Brian go in his place. They hit it off, and soon after the conference they started working together on other projects. You know, I would come up on the weekends and stuff and um talk about the technology and just sort
of be a resources available when he's talking to clients. UM. And then it just kind of kept evolving from you know, we kept working together. While spending all this time researching how Internet giants were using AI to create better products, they came upon a scary but important question. My goal was to try and get Bryan to help maybe just sort of deepen my understanding so that I could help our clients to understand this. And I guess we sort of put it out there is almost a question kind
of like what what can't be automated? At this point, that's when they agreed to create a software version of Ken but better era. This is to Brian built reads news stories on companies and distills that into a sentiment score from one to zero, with one being wonderful and zero for absolutely terrible. Then she'd monitor the stock market to see if these positive or negative articles would move
share prices. If she spotted a correlation, she'd remember that and use it to make predictions in the future, including buy, sell, and hold ratings. Finally, or would sure not short rand summaries explaining the predictions she was making. The project started out as an experiment not for public consumption. Could Brian use machine learning, a hot type of AI to make a robot analyst that was actually useful? And Brian said, well,
you have to explain to me what you do. So I said, okay, well let's let's see if we can we can do this. Um, I'll take you through what I do as an equity research journealist, and you know to the best you know the best I can, and you take me through what would be sort of the tools that would could be used to replace or enhance what I do. He used techniques like natural language processing to build a system this summer in his spare time. Because we have to remember, Brian has a day job
as head of AI at a company called Lola. You know, this would probably be a good time to define our terms for the uninitiated. What is artificial intelligence? He's Brian with an explanation. The most basic answer, Machine learning is any engineering technique that means the software is not discriminately programmed. So I didn't write code that had specific discrete branders and comportent things. It was a system of statistically learning.
In general, the industry term for machine learning means learning specifically from large volumes of data. It gets very complicated. Beyond this, we won't take you down the rabbit hole, but if you only take away one thing from this podcast is this these computer programs update themselves without humans
having to do very much work. You know a model that's it takes data and passes it to a data analytics team, and then they pass it to a product development team and their product development teams who works with the engineers to bring it back into the model. Um, you're you're allowing your essentially seeing where data can drive
the changes within a model's performance. What you're talking about is removing certain bottlenecks in terms of how these companies innovative ERA makes stock calls, and it also tracks how these recommendations end up panning out. If their prediction turns out to be wrong, the system remembers it and it's less likely to make that decision again in the future. Okay, so let's get back to the story. So here Ken
and Brian trying to build this robo analyst. Initially, eras summaries were tough to eat, they were disjointed, and the grammar was off. You could really tell it was written by a computer. But after a couple of months her reports started making sense. I think I sort of realized it could be a thing when it started actually producing language. So when that got generated, it was kind of like, Okay, this is actually kind of meat like in saying what
it's doing. It sort of explaining itself, and it just sort of highlighted that as we keep consuming more data over time, these are only going to get better, and so we should just keep doing it. So this is when the project starts feeling real. Yeah, and Ken decided it was time to share ERA with the outside world, but first he needed to get permission from his company, which also happens to be one of the biggest banks in America Wells Fargo. This must have been a pretty
big risk for Ken. Wells Fargo is a pretty traditional bank and it's come under a lot of regulatory scrutiny over the past few years for creating accounts that weren't opened by actual customers, and the bureaucracy that comes with doing anything new and a large company is significant, and bread on top of that, Ken had just joined the company. When he told the compliance department about what he was
up to, they weren't very pleased. There was sort of a disbelief at first that we we actually wanted to put an analyst out there that we weekly was artificially intelligent, and that we wanted to provide predictions around the stocks, and that actually, you know, could write its own research, and not only that, but it could bold the sections of text that it felt really it wanted to underscores
being important to its specific stock thesis. Right, So that was sort of I think people just sort of kind of through their head back and we're in a bit of disbelief. Why were they so shocked? So equity research analysts spend all day thinking about the future of the companies they cover and how that will impact stock prices.
They send these reports to clients, making different recommendations on whether to buy, sell, or hold, along with a price target, and that's usually an estimate of where they think the stock is going to be in about twelve months. And
if they're wrong. If they're wrong, investors can get mad and go somewhere else for advice and trading services, and these are things clients pay a lot lot of money for the analyst whole reputation, and by extension, the bank's reputation rests on the accuracy of their stock ratings, so there was a lot that could go wrong. Okay, so so far, the analyst Ken has formed an unlikely friendship with the developer Brian. They built this bot to recommend
when to buy and sell certain companies. Their ambition shocks Wells Fargo's compliance department. What happens next well Can eventually succeeds in convincing the compliance department to let him proceed, and it comes time to unveil Era to the public. This happens on September two thousand seventeen. She started making stock recommendations using this complex new software that only a few people on Wall Street even understood. What was the
response like, is your AI technological work with Era? Is it gonna put securities research a business? No, and we did it more as a study. I think for clients who tend to be a little longer term in duration, they're interested in how do you build it right? What is you know? How do you think about the application of this technology? For clients were a little bit more short term and focus, you know what are Era's predictions this week? Right? Can EARRA help me manage news flow
and help me synthesize? What can means here is that the summaries error rights can help clients pick through the daily avalanche of online news to find real developments that will actually move the stock. We recently asked Brian to hook up ERA to an Amazon Echo speaker. Alexa. Ask my AI analyst what she thinks of Google as of yesterday. I think Google looks like a hold and that forecast is good until October. Here's why from Nashable dot com.
Alphabet just took an important step towards becoming a major force in the online payments world. As tech Crunch reports, Let's say they mark AI is an important topic to cover for a number of names. Well, they're you know inbox? Could you know swell with all these AI articles, but they may or may not actually be relevant to the stock's performance. So this program launches in September and it starts making all kinds of calls. The first week went fine.
The second week the Era through a curve bool downgrading shares of Facebook, recommending the investor to sell the shares of the company. And now we should add here that as of this taping, Facebook stock is of about fifty this year alone, and forty six analysts that we have listed on the Bloomberg terminal, only two of them have cell ratings on the stock. So Era made a contrarian call. And the funny thing is Era's downgrade contradicted Ken's own opinion because he has a by rating on the stock.
Turns out Era had read thousands of stories about Russian ads on Facebook that were designed to divide US voters out of last year's US presidential election. Russia probe is focused on people, including three former Trump aids, charged with crimes. Another phase focuses on corporations like Twitter, Google, and especially Facebook. Today, politicians who are up in arms. Congress had called for hearings all these negative articles drowned out other positive pieces.
Why is it so important that we all see these ads right now? Well, when you look at them, they are I have seen them, and this means eras AI algorithms picked up on this in advice selling to avoid a drop in the share price, which she had predicted over the next week, and what ended up happening, Erra was wrong. Facebook start ended up rising a bit over the next week. Some clients weren't happy. A few saw it as proof that what Ken and Brian were doing
didn't matter. For others, it confirmed their view that the way Ken was using AI was fundamentally flawed. That was a disappointment because I think that you know, people are are their zero in on things that they're kind of missing the bigger picture here. To Richard Johnson, a vice president at research firm Greenwich Associates, it was a sign of how far AI and finance still has to go. The robot downgrading Facebook, I think that's probably where this
type of analysis is going to struggle right now. I think we're very much in the early days of it. But you know, in that example, perhaps you know the album Nettle bit of fine tuning and to kind of you know, give less weight to the Russian fake news type stories that are in the media because I think, you know, we all kind of thing that your Facebook will survive, will will survive this. When Ken first explained error to Wells Fargo's compliance team, he said the only
relevant rule they could find was recent guidance on robo advisors. Now, these programs automatically decide how much of your investment portfolio should be in stocks, bonds, and other assets. One recommendation that they had was to tell clients about quote changes to algorithmic code that may materially affect their portfolios. But
this doesn't apply to era's self learning approach. What I think it's very interesting about this and why there's such a great learning opportunity as that robo advisors tend to tend to be programmed. What can means to say here is this, robo advisors are mostly programmed in the old fashioned way humans right, software code that gives step by step explicit constructions on what to do in certain situations.
With a product like ERA, you don't program her, right, the data programs her so her, so her algorithms are constantly changing as a result of that, and so we to some extent we moved out, I think a little further than what you know, finn Ra and kind of these agencies that regulate, you know, our industry are are used to. The solution to this was to include a bold faced disclaimer and every report stressing the era's stock ratings were not investment advice and should only be read
to gain a greater understanding of artificial intelligence. And when we met Ken in New York, we mentioned to him that he'd been putting the disclaimer near the top of each note he publishes about ever, maybe I did just to kind of, you know, continue to get it as soon as I break from in terms of what I'm writing about to what when Aaron starts writing, I'm just hitting with the disclaimer. I don't want to control what
people trade on one way or the other. If they take her advice and they feel that, you know, they agree with her sun reads and they like her points great, you know, I don't want to tell them, don't you know, don't take an information that she has. So where's Eric today? She's now covering more than five hundred stocks and reading
about half a million news stories a day. That's half a million a day, a few weeks ago, the system also analyzed tech company earnings for the first time and issued nine recommendations on stocks like Amazon, Google, and Netflix. And this time did I get it right? This time? They all proved to be accurate after a week. Well, recently these stocks have performed so well my dog can probably recommend it. But seriously, it does beg the question.
Has Ken succeeded in creating his own autonomous replacement? Here's what he had to say about that. But I think in the end it's sort of evolved over time. Is one more of an enhancement. It certainly took took out maybe or it showed how a certain amount of the work that we do in terms of handling news volume and and trying to sort of be specific in terms of waitings of that news volume in terms of stock prediction,
how it could be I think improved. Other analysts agree on the point that cutting the workload is a great idea, but they're way less keen to say that an AI analyst can help them make stock recommendations alongside their own calls. As far as having an air bra our official intelligence Bank research analyst, I don't think that's happening in the the next decade um, but I'll often watch my back. That's Mike Mayo, a veteran bank analyst at Wells Fargo,
who came in to chat with us about Kennon Bryan's creation. Now. I was super excited to see Mike because I remember my first day as a reporter on the financial market. Mike was one of my first calls. He actually picked up the phone, and I feel like I'm some super cool twenty two year old talking to this big guy analyst. He's a sort of a finance nerd, but you can tell instantly that he's so passionate about what he does. For him. ARA is all about enriching his role and
making banks more efficient. I don't want to spend all this time looking at data feeds and checking articles and you know, checking barons over the weekend and did I catch the Wall Street Journal story? And a lot of time for looking at all this information just to see if we're missing anything. Now, Mike said that is associate analyst probably spends about seventy of his time checking news articles, gathering other relevant information, and manipulating data, which are all
tasks that AI can automate. Mike himself reckons that probably about a third of his time is spent on those tasks. The search. Yeah, you can automate looking for the needle in the haystack, or in this case, a few needs. You can automate finding a few needles in the haystack. That would be fantastic. That give us more time to go kick the tires and talk to management and have
much more creative research. Richard, the researcher from Greenwich Associates, estimates that of finance jobs at risk from AI automation and research jobs among the most exposed. I think there will be an impact on jobs. I think for sure, we know that there's significant productivity and efficiency gains that can come from it. You still need, uh, you know, some people, you know, some human analysts to kind of
interpret a lot of the data and so forth. Of course, in this example about the world's fargo on the programmer, he obviously had to put a lot of inputs into that, telling the machine learning algorithm what to look for and so forth. So that type of skill is still going to be needed. But maybe you don't need a team of fifty junior analysts trying to so will that iconic trading floor Hubbub and chaos ever be silenced by machines.
For now, I'd say that the human analysts are safe, but I'm I'm definitely not ready to put my money on them and their associates all being there in five to ten years. Ken himself is holding out hope for human stock research. He used his final moments of the interview to re emphasize that ERA is an enhancement rather
than a replacement to everything that Mike said. As we look out over the next so many years, it is really an opportunity for us to see kind of an enhanced performance and an ability for us to move into sort of a creative research sphere that would be difficult to do otherwise. So, Julie, are other banks starting to do this too? Were showing signs of following the path forged by Brian and Ken. We've seen a little bit
like this. Um nothing exactly what Brian and Kenn are doing, but Morgan Stanley has a sort of bought that will help research analysts dive through earning season, which is obviously one of the busiest times of years for these guys. Tons of news articles coming out each day, so we'll help automate what those earnings are and whatnot. But otherwise there's not anything exactly like ERA that's skimming through news articles,
social media and whatnot on a daily basis. I heard you refer to Era as a she a couple of times. Ken did as well, why are we giving this AI analyst agender. It's a little bit of a canned response, and that it's just because there aren't enough female analysts on Wall Street. So I guess if AARA takes over, then there won't be enough male analysts on Wall Street. Okay, Alice, it's a little surprising that it took Wall Street this long to do this when algorithmic trading it already transformed
finance years ago. What took the industry so long? Well, some of it is evident in the performance of error actually, So each time it makes a call on a stock, that's a data point that can be fed back into the system. Now you compare that to you know, automated trading systems that hedge funds they trade, you know, maybe you know, ten times a minute or something like that. So you just have a lot more data points to
train the systems on. Oh, so that leads to the last question here of course, as journalist, we are not allowed to invest or buy individual shares in the companies we cover. But let's put aside those rules for a second. With what we learned today about Era, Julie, would you follow the bot's recommendation? Would you take advice from an Ai? Not at this point, but I with how quickly she's gotten smarter and learned from her recommendations, I would say in a year or two, I want to be surprised
if I was al Stair. I'm a I'm a human all the way. Um, the only thing I probably would use it for is to scan the all the news articles to see what's most important that that I would make my hand decisions. Yeah, I think I probably agree with you. I'm gonna I'm gonna sound not as future leaning as as Julie. I guess I just think that the inputs that Air is looking at, primarily it seems like it is the news is you know, one important
to mention, but it's not everything. And it just does seem like there's so much intuition and knowledge of a management team and it's past mistakes and future opportunities that goes into making these stock picks. I don't think air is there yet. But Julie, maybe I do agree with you that one day, probably within our lifetimes, perhaps we
we will be taking training advice from an artificial intelligence. Yeah, and I guess maybe the most likely scenario is that it's just Ken and Mike and not the three or four people they have working for them, right, Like, they just have an era and that's all they need. The problem is if they if they give you bad advice, you know, can you really blame the robot? Tbd And that's it for this week's episode of Decrypted. Thanks for listening.
Is artificial intelligence changing your workplace? Send us an email at decrypted at Bloomberg dot net, or you can reach out to us on Twitter. I'm at at Julie Verhe, I'm at Alista M. Barr, and I'm at Bradstone. If you haven't already, please subscribe to our show wherever you get your podcasts. And while you're there, I hope you take a minute to leave us a rating and a review. This does so much to get us in front of
more listeners. This episode was produced by Pie Good, Cary Akita, Liz Smith, and Magnus and Mixon