I hope you're doing well and good afternoon or good morning depending on which time zone you're in.

So today what we have is we've got half an hour session, quite a short session on just the use of AI in banking and the tools, the AI tools have got significantly better over the past six to 12 months and there's a lot more things that you can do with them and it's just worth understanding some of the use cases.

And I'm going to focus mostly on like an investment banking context.

So I'm gonna firstly just talk a little bit about the kind of key issues and mechanics before I go into the actual use examples, but it's gonna be a pretty practical session.

If you have any questions during the session, just in the question and answer section just below me, if you can click that and ask any questions, that would be great.

And I'm just showing my screen so you should be able to see my slides right now.

So in terms of what we're gonna cover, there's actually four main areas that I want to focus on in the next half hour.

One is trading comps and benchmarking.

Second is deal comparables and finding targets.

Third is forecast building forecast models and making sure that your assumptions in the first year of any forecast matches what the guidance is from the management team for that particular company.

Then also news and share price graphs where you can link major events or major moves in the share price to key events and being able to do that quite quickly and easily.

And then lastly, you've got a few bit of treats or Easter Eggs, which are graphic inputs and and company profiles.

So that's kind of what we're going to cover in this session and we'll, we'll spend about half an hour doing that.

Just before I actually jump into these examples, there's just, I want to reiterate a a few things.

When you're using these AI models, most of our clients are moving to in-house AI models or in some cases third party AI models that they're actually hosting internally.

And that's because there's massive sensitivity about non-public information being leaked into the public domain.

So that's something that's really, really important to caveat.

First I'm going to use chat gpt, I use Microsoft co-pilot as well and that's pretty well integrated into the productivity suite within Microsoft.

But actually still nevertheless the underlying technology is the same.

I still find chat gpt actually a more effective tool just gets better results from it.

But I want to kind of give you the, the some of the mechanics of what you can do there.

So firstly, few reminders. You want to specify roles.

So I'm gonna be specifying your a high ranking investment banking analyst.

Specify an audience if you need to.

The tone is actually kind of connected to what I call the temperature.

Actually it's known technically as a temperature and a low temperature.

And temperature is normally measured on scale of about 1 to 10.

So like 1 to 3 temperature gives you much more factual results.

In other words, a much greater degree of accuracy.

A temperature of seven to 10 will give you a much more creative and varied responsible kind of risky from a analysis point of view.

So I would always take a low temperature.

Now in a lot of models you can't actually set the temperature. You can in some, but if you can't set the temperature then you'll need to prompt it. So you need to say make sure you are very accurate in your answer the length as well and an emphasis on particular areas.

And then the last thing to say, just when you're prompting is prompting this just from one prompt.

You want to continually re-prompt and you can always ask the model to interrogate its own answer. In other words, is this a good answer that you have given me? So just within that caveat and one thing I also want to just reinforce non-public information is really problematic in AI models if they are hosted externally.

And so you need to make sure that you are very careful what you put into these models.

Okay, that said, the first use case is finding comparables. And just bear in mind that the models, the AI models that have been used, large language models, the cutoff for chat gpt is October, 2023.

It can still find more recent information on the web, but the cutoff for when it was learning things was this time last year.

So the first prompt I'm gonna do is your investment banking analyst.

There's five companies in the specialized animal health industry in Europe, provided one sentence description for east and format result in tables so I can copy in it into Excel.

So let me just kind of do that and I'll just show you the results.

Should be pretty quick to do. There we go.

Just got my chat gpt.

I'm actually gonna go down and copy a prompt that I did earlier.

So let me just copy that 'cause that's pretty, it's gonna just save time, although you're just going to be watching me type stuff in, which is never very useful.

So put a you there.

The other thing to note, you can actually make typos in the prompt.

That's no problem. So this is, I want to find five companies in specialize animal the health industry. So this is actually probably all like target analysis or sometimes comps analysis.

So I'm gonna hit enter and you can see it's searching the web and it's coming up with five companies. It's given us a size indicator from sales.

It's decided whether it's private or public.

And I, and I can do kind of redo with just public companies.

Okay, so we'll just redo that and it will just get public companies.

So we should now just have a series of public companies there and that kind of gives you a pretty nice use case.

You could also put in analyze or the set by profitability, growth and margins and let's see if that kind of gives us a little bit of more metric.

Now of course you can't rely on some of these numbers, but this can give you a really nice first cut.

And then you could also prompt it, say which of these would make the best m and a target? And let's put in for animal, let's choose one of the large ones, the back.

So if you're just thinking about acquisitions within the sector, you can kind of prompt it and hopefully it will.

So it kind of gives you a nice little grid here. And I'm actually gonna, this is not great.

I'm so put the, put the results into a grid with a temperature indicator or the suitability.

So this is a really great way if you're looking at ideas within the sector.

And you can just, you know, if you've seen all seen these grids in a deck, right? And it's just really easy to do here.

Of course you have to check the results, but it's really nice and you can actually do this for private companies as well. It saves lots and lots of time.

So I've just done it for public companies and then I can say, so redo the analysis with the top 20 private companies in the EU.

So we want EU come and let's see what it comes up with that. So you can kind of key keep on prompting this.

And this is really fantastic at finding good targets.

So I mean this is, this has given, yeah, it's gone through there and it's probably, you need to be a bit careful that's a little bit of tube and wider search there.

But I do find when you're trying to find comparables, whether it's target, identify or trading comparables, it's really good.

So let me just do another one of these quickly.

So it could be, I'm gonna do something a bit more general.

So let's say we're looking at Kellogg.

I want to do a trading comparables, well I pushable to, I'm an IB analyst still remember that I want to do the a trading comps analysis analysis for Kellogg in the us for the key competitors, the key publicly traded competitors.

Now when I do this, one of the things I find in Factset, CapIQ and Bloomberg, their screening tools don't tend to be that good often because the databases they're using for screening and things like SIC codes and those are so out of date, particularly if you're looking at the tech sector.

And the good thing about MLM models, they will analyze things like company description.

So I find that you get a much better indicator of key competitors using AI models than you would do from screening tools in some of the databases.

Now it's put in Nestle, it's an, they have ADR I think in the US market, but that's technically a public company.

So it's just kinda worth understanding.

But I find this for finding comparables really, really fantastic.

Now this is, I'm gonna put in a prompt now, which is where you really want to be careful putting it into an open a mo I model if you are working on transaction.

But the other thing I've done, so I am an IB banker advising Danone on who, who are looking, looking for targets, the USA rank companies in the table in, in a table based on three criteria, strategic fit, profitability and growth.

Okay? So then it will, it will go away and find good targets.

Now as I said, you want to be really careful with putting this because bordering on non-public information, obviously there's no, I'm not advising this deal, but I just wanna show you as an example.

But this is a really, really good way of getting targets when you're doing analysis and I think it's because of the MLM aspect of being able to search lots of languages.

It's really, really fantastic.

Now let's do some analysis where you're being a bit more specific.

So what I'm gonna do here is if I'm going to do deal sourcing, so I'm just use comparables and target analysis.

We've done that. Now I'm gonna do deal sourcing.

So I'm gonna ask it here using Coke's investor relations site, summarize their recent M&A transactions and and how much they paid.

Okay? So this is actually pretty nice because this is again saves you a lot of work and you can control.

So you can go to FactSet and see the m and a activities, but it won't necessarily give you a little bit of color about the deals that they did here.

So it's a really, really nice thing to do that sometimes though when you're kind of giving it these really broad things mean I've told it where to go to find the information.

And you can see it does actually give us some specific prompts to where this information was from, which is great, but let's say I want it to be quite specific.

So I've got a report here that I'm going to upload from my computer and I've got a, the capstone part is beverage coverage m and a port.

Okay, so I'm gonna ask it, I've got a report here.

So summarize all the transactions listed in the attached report in a table and I like to put in a table 'cause it's just easy to copy and paste. So I'm just gonna do that. So here we go.

It's just taking that other report and you can imagine that's just so much more straightforward and it's going through and the trouble is it won't necessarily get all the items in this report, but it'll get quite a lot of them and it will just save you time extracting it from a a document.

So I find it just really super helpful if you have a document and you want to extract content from it, again, it has to be checked.

Okay, so this is one that says this is, yeah, August, 2023.

Yeah, I've done that for one particular year.

The other thing you can do is that if you have an earnings transcript, I'm just gonna go to our platform, Felix platform and let's I go to Coca-Cola and I'm going to go into the transcript of the earnings call.

Let's do a year end earnings call.

So I'm just gonna select all this, copy it and I'm gonna dump it in and I'm gonna ask it to interrogate the M&A opportunities in the following earnings transcript.

And I find this really effective.

So there's something just pasted it in.

You sometimes find that these models, they're not terribly good or they kind of limit how much you can put into them the amount of text.

Actually chat gpt is pretty good, but some just gonna use it.

So here it's going through the earnings transcript and I find when you're doing this, earnings transcripts are better than 10 K or 10 Q filings.

So I know it's giving here components of the analysis and it's giving some suggestions of areas, areas and let's say can you suggest some ideal companies to acquire and hopefully using the above analysis.

So here it will go out and find potential targets as well.

So from a deal sourcing point of view, it's really, really fantastic and it kind of gives you, it shortcuts that analysis really, really carefully. It's not foolproof but it's pretty good.

Now next thing I want to do is just show you if you've got a specific data set.

So I, in this case I've downloaded a set of M&A deals from FactSet and I love FactSet but I often find FactSet problematic because it just gives you a lot of garbage.

If I just show you the actual Excel files, if I go to my desktop and I'm gonna find the analysis here. If I go to mergers FactSet, so this is the dataset that I use and you can see that these are all the deals in the non-alcoholic beverages sector downloaded from FactSet and they're 175 deals.

This is quite a big data set and I could spend an hour going through each of these in a granular way.

It's potentially waste a lot of time doing that.

So I'm just gonna show you how, if I can just say review the attached M&A activity to grid, and find deals related to the energy drink sector.

Summarize deals, deals and provide and provide a background explanation for each deal.

You are a high ranking IB analyst and want a high degree of accuracy go without saying.

So I'm just gonna attach the file as well.

So lemme just upload it from the computer and that was the that there upload it.

Oh lemme just close you down.

Go and uploading that.

So this is a very specific data set that I'm using and then I'm asking it just to go through this.

It's opening it and it's gonna go through 175 deals.

There's quite a bit of information there already.

It's analyzing it and I just want deals in the energy drink sector.

So that gives me all the deals in the non-alcoholic drink sector and I just want to prompt, literally fine tune it.

So here we go. It's giving us the deal here.

It's giving a bit of background analysis and this I just, this is for me when I'm doing particularly M&A deal comps, it's really, really helpful.

Now this is, do this, I'm gonna do this for all the deals and put the results into a table.

So I'm just gonna do that.

And it should be able to put the results in the table that, so actually I prompted this earlier and it was a much better outcome.

So this is one of the frustrations with AI. You do get differing outcomes depending on when you prompted.

So let's just see, yeah it's just only got one, there's only one deal in the energy drink sector.

I don't think that's, so let me just show you this is one.

So that is the same prompt, exact same prompt and I get different answers.

So you can see that that's quite a dramatically.

So let me just change the prompt, let me redo that and see if I can go into the prompt.

And this is one of the downsides if I just go and upload it and I go to, there we go.

Put that all prompt in. Let's see if it will change it because that's kind of strange to see how different those results were.

One just gave me one deal and the other gave me a really nice grid.

I think we did use the right, yeah we did use version two, not version three, but you can see how I prompted that.

It's still giving me one output.

I'm just wondering whether I used a different file.

No, I use the same file.

So that's really interesting giving me completely different results.

That's kind of frustrating but that is one of the downsides.

Okay, next up is taking a look rather than deal sourcing is forecasting.

When I'm building a forecast for a company, I often want to be able to update my forecast with the guidance the company has given for the next 12 months.

And I'm gonna do this using a 10 K filing, which is a big filing and then earnings call slides.

Okay, so I'm gonna just go to the Coca-Cola's 10 K.

So lemme just open a new prompt here and I'm gonna ask it and summarize the growth forecasts in the document by segment.

And I'm gonna ask it can you provide estimated, let's just do that.

Okay, so I'm gonna upload the whole 10 K file and the file is really big.

I generally would suggest to you you don't want to copy and paste 10 K in there.

It will blow up, it will say you've got too many characters but in an attachment they find that much easier to deal with.

So in this case it's gonna read the 10 K filing and now often you don't get actually specific items in the 10 K filing.

So it kind of gives you some information but it's too generic so that's not terribly helpful.

So I'm gonna do um, another prompt any forward guidance in the information below and I'm gonna attach the slides.

So these are the track, these are the presentation slides.

This is gonna be for Kellogg in this case.

So this is Kellogg's earnings call slides that have been that I've uploaded there.

And let's see if it gives us any forward guidance in the document because I'll generally find that earnings call's gonna give you much better information.

Any specific, any specific I did get a specific guidance or Kellogg just see if there's, because I've got a better result previously.

Yeah, there we go. So do and what about, what about about growth rates? And I think I kind of confused it because I had both Coca-Cola and Kellogg in there.

Yeah it kind of gives them us whole numbers.

I actually got a better result. Let me go back.

Let me take a look to see if I have the example from previously.

Just see guidance Go activity forecast summary.

So yeah, this is I think I actually put the, the transcript in there. So lemme just do that. Let me just show you if I put the transcript, I generally find that transcripts are really good.

Let me do it for Coca-Cola.

Again, I'll go for the ending 2004 and I'm just gonna ask it the same thing.

Give me, give me any guidance on the next year's revenue and margins from the below transcript.

So I found earnings transcripts.

Best thing is because it's a language model, they're generally much better than even like structured documents. So it's going into the, it's quite a long document here.

I'm just gonna pull in hopefully stuff which will be useful for our model.

Just conscious of time, hopefully it'll coming up.

So there's a lot of data there. There we go.

So this is actually pretty good.

So giving us revenue for growth of 6, 7% and this helps you if you're building a model, it just gives you a really quick glimpse into what's going on.

So the earnings call transcript is actually really useful source of information.

Okay, another prompt I'm going to do, which I find just painful as hell is where you have those share price graphs and your MD says, I want you to identify where the share price has moved significantly and annotate why it's moved, whether there's been a earnings release or whether there's been a major macroeconomic event.

So, I'm gonna put in a new prompt here.

I'm gonna upload, I've got a file.

So if I come here I've given this is a FactSet download of dates and the share price for Coca-Cola over a five year period.

And so I'm gonna upload this close that, I'm going to upload that file into the prompt.

So I'm gonna come here to this price summary and then I'm gonna prompt it with using the attached file for Coca-Cola's share price.

Identify any reasons for price movements up or down in a table, just see what it comes up with and hopefully it will just summarize the key moments.

Now often sometimes when it's using numbers it doesn't tend to be as good as if it's using language. That's why things like the transcript is actually a better result to do that.

So, it can actually do reasonably good job at calculating differences, seeing the big drops and falls.

This is, so it's going through, yeah there's quite a big, it's done the whole, here we go. Let's just see if it's summarizing.

Yeah that seems to have do that, do the top 10 price changes.

So hopefully that will just summarize it because it put way too many there so I can just finesse it down a little bit.

And then just see if it's, it hasn't actually put them in order weirdly.

Okay, so the top 10 such given us more than 10, which is a bit frustrating.

So this is the problem where it's using numbers generally these AI models, they don't like numbers because they're language models.

So that's kind of the problem you have.

And then what you could say is you could say, analyze the attached document from FactSet and link to price changes.

So this is a news feed that I've taken from FactSet, so just pull it in here and hopefully I'm gonna there in that news download, open that and then hopefully I can link the share price changes to that.

It's actually got less reliable just over the day.

You can see that taking a while to upload it maybe.

Because the Americas are open right now.

So that could be that sometimes it does degrade a bit.

So you can see here that actually when it's dealing with numbers it's not very good because I asked for the top 10 and it's given me 20 in that case and it's still trying to upload. Let me just prompt this, see what it comes up with.

Generally speaking, if you give it something and if it's text, it's much better than using numbers.

Now This is, yeah, it's given me a lot in October, 2024.

It's not great though.

And let me just see if we've got another example of where I've done that.

So just see, okay, price movement analysis.

Yeah, it's not great.

And that's partly because we're using numbers there.

Now I'm conscious of time so I'm going to give you one little nugget, which I really do think is fantastic and that is extracting graphics.

So in this case, if I go to a dataset, let's go back to my prompt, I'll go to Coca-Cola, I'll go to Kellogg actually, because I think Kellogg's slides are better.

So I'll go to Kellogg and I'm gonna go to the earnings call in 2001. These are the slides and in one of these slides it covers us a graph and I want to extract that graph and put it into a table.

You could either type it out, but this is where actually this is really, really great.

So I'm gonna just do a screenshot, select it.

Once I've done the screenshot, I'm gonna go back to the AI model, a new prompt.

And I'm gonna ask it, can you summarize the below graphic into a table? And this is just really fantastic if I just paste that in.

So it's gonna paste it in and hopefully it will just produce a table with those numbers. You've gotta check it, but you can see here, um, oh it's actually Kellanova.

Where don't you actually put numbers in check Kellanova? It's 13, 8, 5, 7, 5.

Yeah and it's got the correct dates there.

Q1.

Yeah, so you can see that that's a really fantastic thing. You can do screenshots and you can get the data extracted without having to kind of go in and retype it yourself. So I use that all the time.

You go to the FT or Wall Street Journal, you can just select it and you can dump it in.

Anything that you can screenshot, you can extract the information super easy.

That's as much as we've got time for the half an hour is up.

I hope there was a few tidbits of use there.

I think one of the big things to remember is AI models are language models and therefore they're not so good at dealing with numbers, but they're very good at screening, much better than FactSet, CapIQ or Bloomberg.

Really, really good at screening deals or comparables.

Very, very helpful. Okay, thank you so much.

I hope you, you have a really great weekend.