Okay, welcome everybody to this webinar on the latest ChatGPT features.

And this has got a real kind of focus on finance.

So what I'm gonna do first is introduce myself.

My name is Alastair Matchett.

I started my career at JP Morgan in the financial advisory or M&A unit and I covered financial institutions, banks primarily and then oil and gas. Then I went to the private equity world as um, a fund in the UK called 3i and then went to the education business and teaching on Wall Street for more years than I would care to admit.

But there's been a really huge change in the ability of ChatGPT to add productivity.

However, one of the big issues is how it relates to finance because it's very cool at producing documentation and it's a large language model.

And what that means is that it's a probability based model.

So it's going to guess broadly what you want based on probabilities.

They essentially looking at relationships between two vectors and a set of data and then they're applying that to the task at hand.

Now that is great because it's a very creative tool for language.

Where it's less good is for funding very specific things you would need in say an M&A context.

So what I'm gonna do is show you a few of the new features because I did a session this time last year on ChatGPT and how amazing it was, but actually there's been some really important features that are actually significantly going to improve your productivity.

If you have questions as we go through, there is a question and answer um, button at the bottoms, please don't hesitate to ask questions there.

And there's also a feedback form in the resources.

We're gonna focus on this session on ChatGPT because that is we widely available.

A lot of banks haven't in introduced that into the firm, but what is happening is that co-pilot, which is the Microsoft product, uses ChatGPT as its engine.

And I suspect the next years there will be a rollout of copilot and we'll do a separate session on that.

But actually fundamentally in terms of the things that copilot can do versus ChatGPT because they're using the same underlying technology, it's pretty similar.

Okay, let's get started.

So I'm gonna share my screen now 'cause I've got some slides and what I'm going to do is I share my go through this session.

I'm not going to generally prompt ChatGPT because it just takes too long.

Um, the prompting is really quite slow.

So what I've done is I have already done the prompts and got the answers.

I may do a little bit of prompting so I'm gonna show you the prompting and then the answers I've got.

So before we actually even look at ChatGPT I want to kind of give you a heads up on some of the kind of key things that have changed in the last 12 months.

Okay? So let's take a look at the really big changes that have happened and I kind of put these into two blocks.

Um, firstly there's internet connectivity and that is good because it means you can actually search the internet, you can tell it to use the formatting from our website for example.

You can ask it for news stories, you can do things like trading comps, um, research, which is really, really great.

So the internet connectivity is a profound change to the system.

However, I think that's less useful from a finance point of view.

But I tell you where I think the really significant benefits are gonna come from or are is number one the ability to upload files.

And that was, you could kind of do that beforehand, but now we can upload really big files.

So they have really dramatically increased the file size.

You can upload into ChatGPT and there's a, the model is actually called ChatGPT for turbo and that allows many, many, many more tokens and that's how they measure the input into chapter BT than previously you could put into the system.

So that's been a really significant change is the increase in tokens. And that's really great 'cause now you can put a 10 K or an annual report into ChatGPT and you can query then report and that's fantastic.

The second big thing that I think is very helpful for people in finance is you've got optical character recognition.

So you can do a screenshot and you can tell ChatGPT to convert that into a table.

So this is me, this is that gonna be really significant.

So you can info memo blocks of data, you can literally do screenshot and copy it and you can get it to extract the data.

So those are really significant.

They also have recently launched an app store kind of app store within ChatGPT I, I've really played around with this a lot and I'll be honest with you, I'm being pretty unimpressed.

It feels like there's just a kind of, it's a very simplistic kind of uh, cover over the ChatGPT engine.

I haven't really seen a massive benefit.

There's one tool that I think could be useful but broadly speaking I think it's not such a great, um, there's the apps are not so great at the moment.

I'm sure that will change.

So, um, that's the content of what we aim to cover in this session.

So the first thing I'm gonna do is to show you the internet connectivity features in ChatGPT.

Um, I'm gonna start with what I think is good and that is if I come to an example here.

So I asked ChatGPT to summarize BBC news.

And the first prompt when I did this first it said I'm unable to directly access BBC News website. So that's the first thing that there's quite a lot of restrictions out there about which websites ChatGPT can access.

But weirdly I prompted it again and it did give me an answer but it doesn't get the answer from the BBC news news website.

It's getting it from ground news, which I think is scraping the headlines that are coming out of BBC.

Now when I did this, I looked at the b BBC website to do comparison and actually it was pretty accurate.

It did include pretty much all the major news stories from the BBC site, but it's not going directly to the b BBC site. It's kinda using third party site in terms of the breaking news stories.

So just be careful about this.

But nevertheless it is a very, very useful overview of the news stories and you can kind of re-prompt it. You can say give me the major news stories in five bullet points, um, and you can re-prompt it, re interrogate the answer.

So I think this is helpful because it's better than sometimes Googling it because that's gonna give you it in the format that they design.

But in this case you can kind of re-prompt it and get it in the format that you want.

Exactly. So that is an example of where the internet connectivity is a really good, um, helpful tool for you.

But be super careful because actually I found a lot of the features, um, the heuristics have been um, really terrible.

So if I go down to, here we go, the Kellogg stock price graph, I've asked, it said I said generate the stock price graph for Kellogg and it's done it over the course of the prior year and it is garbage, complete garbage.

And actually if you read what they've said here is his generated graph showing the mock stock price of Kellogg. So this is a complete heuristic, it's complete garbage, it's not correct.

So I would say getting any kind of financial information out of the the web is pretty dangerous.

So I think this needs to be used very, very carefully because a lot of websites have come restricting access to ChatGPT.

So be really, this is a classic heuristic, um, uh, old mirage, um, whatever you wanna call it, um, where it's just creating data out of thin air that isn't real at all. So just be quite careful with the internet connectivity.

Um, one thing I have found it incredibly helpful for though is where you are doing trading comps or credit comps and you want to identify which firms are in a sector.

I've for a long time when I've used FactSet CapIQ or Bloomberg and I've done screening tools in there, it they're just clunky.

They use things like SIC codes which are really out of date.

And so it's very, very hard sometimes to triangulate down a particular industry.

And actually that's I think where ChatGPT is very good. Just bear in mind though, it goes, the latest model goes up to December, 2023 so it's not going to include the most recent months.

So just be careful about that.

But let me give you an example of where I did this and I purposely chose a really specific industry, a little kind of niche industry.

And I remember working on this when I was a banker.

So I've asked it to list the main comparable companies operating in the European animal health sector.

So this is kind of real small sub-sector, estimate their size and include both public and private companies.

Assume you're an investment banking analyst, it's generally always important when you're prompting to make sure that you tell it who you want it to be.

And it's gone through and it's pulled out the main animal health companies in the European market and it's kind of given some stats, market capitalization here, sales just be a bit careful these stats 'cause they can sometimes be wrong but usually is in the ballpark.

Um, and this is really, really great.

What you can do is you can say, well actually I I would like these um, in a kind of table.

So lemme give you another example where I've done something similar in this case if I go, I've asked it for the top 10 investment banks and their website addresses.

'cause this is really helpful if you not just don't just want to get their names, you want to get their website addresses, you can just click, click through.

So here it's given the top 10 investment banks according to ChatGPT, it's given the market cap total assets, the location and it's given the website but it's in a kind of granular list here.

I could have asked it to put it into a table in ChatGPT, but what I've um, also done here is ask it to put the list in an Excel format.

So then it takes that list and puts it into an Excel format.

The problem is that this is going, yeah it's, I did this a few days ago and so it's kind of outdated so let me just show you what it's pulled out.

I had to re-prompt it this morning.

Oh gosh, I couldn't do that.

Let me just um, redo the spreadsheet with tickers.

Um, I may actually just pull up the spreadsheet 'cause I did this earlier today.

Yeah, here, pull analyzing it takes a long time when it's doing this kind of irritating.

So just bear in mind that sometimes the speed is pretty slow. But let me show you the result of what it produced when I asked it to convert that output into an Excel table.

So let me just pull it, um, around here just there we go.

Use that and just move it to this table we go.

And you can see here, um, this is the grid that it's produced. So it's got the um, ranking here, the name, the ticker market cap, total assets, location and website.

And you could then obviously use your FactSet or Capital IQ or Bloomberg coding to generate some multiples off this.

So I would say this is probably the best use in terms of finance, of ChatGPT, internet connectivity is that you can, and previously it was quite out of date but you can do lots of comps, mining and get tickers and everything like that saves a lot of time and I think it's fundamentally much better than Capital IQ or FactSet or Bloomberg at screening.

Um, I found for many years the screening tools in those platforms pretty um, tricky frankly particularly FactSet.

Um, and I use FactSet as my main tool.

So that just gives you an example of how you can get it to analyze, um, get, get you comps and then put it into a table now.

So the next thing, and this is something I I really literally just blew me away because previously it was always a problem getting data into ChatGPT and now you can add files, large ones into ChatGPT and this is because the model is now called ChatGPT 4 Turbo.

You'll need a paid subscription for this.

But it's really fantastic at pulling data out, doing a really interesting complex search from the document because it will do a kind of more intelligent search and control f and it will give you the page references as well.

So it's really fantastic. I found extracting tables really problematic.

Um, and it's still very text-based. So let me just give you an example of this.

I'm gonna go back chapter bt um, and I'm going to show you this is um, a prompt.

Lemme show the original prompt at the top.

So firstly I uploaded the PDF of Kellogg's 10 K, it's about 150 page document pretty large and you can just do that by clicking on the um, little uh, paperclip down there.

Then I said summarize forms Kellogg in the attached document and just go through and gives you a little summary which I guess is kind of useful.

Um, gives you a nice little summary of all the main things that are going on.

And then I said, what's the financial performance of the latest year compared to prior years? Again, this is pretty text-based.

I didn't think that this was particularly good just doing the analysis and I think ChatGPT is probably less good at doing financial analysis than extracting data or producing text.

But where I've started to find it really useful is, well actually before I take you on that, I also asked it to download financial statements into Excel.

Unfortunately this is, this is um, gone now but this was not good.

What it did, it took the three income statements in one tab that was okay, but the balance sheet and cashflow statement, which put in two separate tabs were completely empty.

So it only pulled in one set of financial statements. It did find the statement, but the balance sheet cashflow statement, it didn't pull it out.

So I think getting extracting data from the document is pretty clunky still.

So where, what is useful? Well, asking it specific things about the document and now he said which pages um, are the references to non GAAP earnings.

So when you're looking at a 10 K and you want to get the cleaned earnings, you can um, query this and it's saying that the um, measures are found on pages 39 to 40 and this means you can find the exact place in the document where that's located.

So you could ask it for change of control provisions if you've got a um, a loan documentation or you could ask it for say all the lease references in the document or operating leases.

You can really kind of triangulate it and it gives you a much, much better performance than the control F tool where you'll get a hundred references and you've gotta go through each of them in turn.

So I think this is where you can do really clever searching in the document.

I also try to get it to help me with modeling and frankly this was just garbage.

I said I want to build a forecast model, identify the key value drivers and it came up with just a whole load of text-based output, which is just useful for useless, sorry for financial modeling.

Then I thought, okay, well let's um, summarize the revenues.

'cause in the MDNA section of Kellogg, it breaks down the revenues, um, based on price changes and volume changes.

And that's how I would tend normally build a model for Kellogg.

And again, it didn't really do a very good job at that at all.

So that was a bit disappointing.

So the thing that you can is for example, say find all the um, data related to operating leases in the document with page numbers and then assume you are an investment banking, sorry banking.

So I dunno if that, that may take a a while to do. But it will go through the document and in the document it'll find all the references to it. So it's really quite a good tool to search and it will give you clickable links. So if you look at the top, um, in my prior reference I'd go up and where I asked it for the non gap earnings, here we go.

Um, you can see you can click on that and it will take you to the place in the document where that item is.

So that's pretty um, useful. Let's see.

Okay, that's not so great is it? Um, yeah, so this not a great answer but I found that actually this is the best um, use of it is querying the document.

You can upload really quite large documents now.

Okay, um, what else is good? Well this is bad view low.

So this is opt opt optical, sorry, character recognition.

So what this allows you to do is you can pull into ChatGPT screenshots and ask it to convert the table into an Excel file.

And this I think is probably really supremely helpful in terms of productivity.

So if you want to do a screenshot, the shortcut to do that is control shift s so not control shift S windows, key shift S.

So if you do the windows key and shift s what it will do it, it'll open the screenshot tool, you can select the screen.

And what I did in that 10 k is I selected the breakdown in sales of volume and prices.

'cause I put them to my model to be able to build a forecast of sales based on price changes and volume changes.

And this is really significant because this is always a pain, particularly you've got an information memorandum, et cetera that you want to extract data from.

Bear in mind be careful because if you upload it into ChatGPT, there's a massive confidential issue.

So don't use any non-public information in ChatGPT I think as Microsoft co-pilot gets rolled out because there you can segment internal data and external data, there will be some opportunity to do that, but that's gonna take time 'cause the bank's gonna have to be very careful at segmenting the search data that co-pilot uses.

So watch that space, I'm sure that would come in.

So let me just show you an example of where this has done.

So when this is done, so if I go into, um, I think it's this prompt here, let me just, here we go.

No, it's um, here we go.

Yeah, um, now that's visualization.

I've just been using Excel data.

I think this is, yeah, so here we've got um, convert the image below to an Excel spreadsheet and um, I just open it up.

Lemme just quickly um, redo the spreadsheet, um, take a while. So what it will do is it will extract the information and produce an Excel table of that graphical information and it's really pretty good um, because it's not kind of looking at the large language model, it's just looking at the data that you've given it and it does optic optical character recognition.

So I would say this is a very, very useful tool about getting um, graphical information into tables in Excel.

I find it really, really fantastic.

Saves huge amounts of time.

Um, let me just wait until this error analyzing.

Let me just, I'll tell you what, I'm gonna do a new prompt and what I'll do is I'll go to Kellogg, lemme just quickly go to Kellogg, um, and extract the table.

Let me go into, let's go to the press release and just get an extract, um, pull this out.

So I'm just gonna do a screenshot, um, and then I'm gonna go to ChatGPT and I'm gonna paste that in and say produce um, Excel table of this information.

And so we'll go, we'll take that screenshot and it will create an Excel table.

So hugely time saving doing this rather than having to go and do it yourself.

And it's really pretty accurate at doing this.

I I'm going to prompt this just 'cause I wanted to show you this 'cause I think this is one of the really, really useful tools in ChatGPT this the ability to screenshot and you can do websites as well and you can also select websites, dump it in and it will produce an output.

But the graphical stuff is amazing.

That's the one one danger of course is that in examinations we used to kind of use graphic image, it's to sort people copying, pasting and chapter BT getting the answer now they can just do a screenshot.

It's getting the world is um, changing really rapidly.

Um, but I think that this is a really big improvement for you as a kind of investment bank analyst analyst.

I'm just gonna wait until this is finishing.

Let me just hop over to my slides while we're wait for that.

So that's a huge time saver.

The ChatGPT apps, as I said, I have found these less useful.

There's one example I do want to show you which is the diagrams app, which I thought was kind of useful but the other ones I found pretty clunky.

I'm sure that will change but at the moment I spent a lot of time, the ones I thought would kind of be useful.

Co co-pilot credit writing coach, um, web browser, the negotiator that wasn't too bad producing uh, little how to guide Excel GPT.

Again that's pretty useful.

So lemme just go back to see if that prompt is finished and it's still analyzing.

So this is, yeah this is one of the problems.

It's pretty slow. Um, but I need see if I can pull out my Excel so if I can find that output.

Um, let me quickly see if I can find Uh, the output Kellogg National statements. Lemme just see. Yeah, so this is something that ChatGPT pulled out of the Kellogg document.

And you can see here this was a screenshot and it's pulled out the key numbers there.

So you can see it's really pretty good at extracting um, data from, from this.

So let's go and have a look at some of the other things that I've pulled out.

So I'm just gonna show you the visualization.

So let me come down to um, I think this is my visualization.

Maybe you just gave a little Notes, it's not there. There We go. Yeah, here we go.

So this is why I used the app, the um, diagram app and I asked it to produce a diagram based on Kellogg's results and actually the result was pretty good because one of the things it did, I mean the actual diagram's pretty clunky but it's taken the timeline of 2023 and it's kind of said fiscal year end spinoff completed independent company trip began trading with independent company record date for spinoff shared distribution, waterproof spinoff.

So that's quite useful at creating a diagram from the information that's graphical and I felt that was the one thing that was pretty cool In there.

Um, the other tool that I used is the Excel tool.

And what I wanted to do is help me figure out a function that I don't use very much, which is the Lambda function.

And I think this is really helpful.

Asking ChatGPT questions about how to do something is good because what it allows you to do, it kind of gives you a step by step process, can give you examples, you can re-prompt it so I don't understand this, gimme more detail.

So I think if you've got a new analyst or you are using a function you don't use very much.

If you go into Microsoft help area, it's just not very good because it's very, very generic.

Whereas you can tell GPT you are financial analyst, you are using this for a finance solution and it will give you an example relevant to that.

So I think the Excel function help is very, very, very useful.

Non-public information, uh, it's it's public information so that's fine.

Um, the other thing I did is just again the negotiator said, just give me how you would negotiate sell the large company like Kellogg.

Actually it was pretty good. And so for a new analyst, this is very, very helpful for them to kind of get a step by step guide about how to do things or how things work in the industry.

Also, if you have got a document, let's say you want to understand the 3 38 H 10 election in a merger in the us which allows you to convert a stock merger into an asset merger from a tax point of view.

Normally if you Google that, you'll get a whole load of articles from Ernst and Young, pwc, Deloitte, et cetera.

That'll be a hundred pages long go into huge amounts of detail and you know, you need to add two hours just to read through it.

You upload that into GPT and ask it questions about document, summarize it, gimme examples, what are the key pages? And I found that very, very useful way of taking a really technical document that you don't really, um, you spend, you could spend hours trying to figure out and get it to help you understand it quickly.

I think that's a very, very um, useful tool.

Just before we finish, one of the other things I try to do and think about kind of all that use examples.

Um, in this case I asked it for London to Rio Flights on a particular day, pretty useless and it gave me one not great.

Um, taxi services to Pineville, which is a small town in the us Again, it gave me one, I didn't think that was very good.

But what was quite good is if you are planning travel in this case I said what's the best place of getting to falling water from New York? Falling water is near Pittsburgh.

And this was very helpful because tell me actually the best way is by car.

You could also take a train to Pittsburgh and then do a car rental or you could fly to Pittsburgh and do a car rental as well.

So it gave me all the options.

So if you are planning a trip to a client that's kind of in like a weird location, actually this is very good at kind of giving you the root mechanism, it's not so good and actually scheduling times for them.

So that's most of what I wanted to cover in this session.

Um, and just kind of recap, the really big changes are internet connectivity, number one.

Number two, the ability to upload really large files.

And number three, optical character recognition.

That changes tables into Excel.

I know that was a bit clunky, um, in my prompting, but sometimes it is a little bit erratic. Depends on the time of day.

Um, one thing you'll find at like 2:00 PM London time, which is 9:00 AM New York ChatGPT gets like really busy because you've got two main financial centers or two main global cities, um, using it a lot and you'll sometimes find that it really slows down at certain times of the day.

So you just need to kind of be careful about what times of day use it.

I use it all the time, obviously being massively careful about confidentiality, but as copilot comes in which uses the same underlying technology, I think that will be solved.

Thank you so much for listening.

We have a feedback form which if you click on resources, we'd love you to fill out.

Um, and that will also allow you to ask us questions on an ongoing basis.

Next Friday. We are doing enterprise value complexities.

We'd love to see you there. I hope you have a great weekend.

Thanks very much for watching.