All right. We now have another Anthropic model.

It's literally set the world on fire, so let's see how well it does.

It actually knows how investment bankers think, and I think that's really impressive.

This is quite a lot more usable than the result we got last time, and that was six weeks. Just taking a look at this, I would say we're firmly in the A-minus territory here.

A-minus or B-plus. It's done much more complex processes than last time. It's done much more sophisticated DCF calculations.

Welcome to this week's episode of "What's the Big Deal?" Right, Graham, what are we going to be talking about this week? We're talking AI again. We've got a lot of news in the last couple of weeks about Anthropic's recent model.

I know everyone loves Fable 5 for coding, but we want to put Fable 5 to the test, at least while it's still out and available to the public, and comp it against the last Claude Excel test we did to just see how well it builds a DCF.

We're going to give it a really simple prompt, basically keep everything else the same, and compare A versus B and see how much the world has moved on in the last-- I watched our episode the other day or last night, and I think at one point we said, "Let's compare and contrast in, say, a year when the technology moves on." And here we are, literally a month, maybe six weeks later, talking about, all right, we now have another Anthropic model. It's literally set the world on fire, so let's see how well it does.

Absolutely. Yeah. So about six weeks ago, we asked Claude to build a DCF model of Lululemon.

With mixed results, I seem to remember.

There were a few challenges in how it was calculating free cash flows and unusual assumptions. And yeah, I do remember you saying, Graham, "Let's revisit this.

Let's circle back." And I think, as you said, we've now got this new version, Fable, which we are going to put to the test and just we'll have a side-by-side comparison, basically, won't we? We'll be able to use the same prompt and see how the world has moved on.

Yeah, exactly. So last time, I watched our last episode yesterday, and the main, let's call it screw up, I think we used Opus last time, that Opus kind of made in the model was it didn't even calculate NOPAT.

Mm-hmm.

And I think we kind of agreed that in this case for Lululemon, because Lululemon has zero debt, just cutting out a couple of steps on the NOPAT and the free cash flow calculation was kind of okay, but also not really. It's not best practice, and at no point did Claude even say, "This is why I'm using this calculation." So- Mm-hmm ... we want to see if it's grown up just a little bit.

Fantastic. Exciting times. Look forward to doing that.

As a bit of context, I do remember last time Lululemon had been having some challenges. Its share price had been really languishing because they had some governance issues. I know they had a proxy battle going on at the board level, and there were concerns about the strategic direction. And actually, it seems like the world hasn't really moved on in a sense that the proxy fight has been resolved, but they're still having major issues with their strategic direction. They have a new CEO, but still the share price is languishing. So let's see what the DCF tells us about the fundamental valuation of Lululemon.

All right. So I'm always interested to see how well these models think with really simple prompting.

Mm-hmm.

Now, there's a whole science and a whole thought process behind prompt engineering.

I think the better these systems get, the less we're going to have to pay attention to prompting because one of Fable 5's advantages as it relates to coding and software development is actually you don't need an insanely detailed prompt. It's actually pretty good at figuring out exactly what it is you want and what it should be doing. So I'm going to start simple with this and see how we go.

I've got the Claude Excel add-in open over here.

I've got Fable 5 selected. Hopefully, I don't run out of credits because they're being pretty stingy with Fable 5 credits these days, and I burned through quite a few of them just on other stuff.

So let's see if we can get through this without Anthropic turning us off. So all I'm going to say is build a DCF for Lululemon.

And should we just give it the usual parameters, just literally to say follow financial modeling best practices, literally make this look like it came from a bulge bracket investment bank? Ooh, I like that.

That seems- Yep ... fair enough for me. Okay. Follow financial modeling best practices and make this look like it came from a bulge bracket investment bank. And I got a typo in there, but I'm sure it'll figure it out.

Yeah.

It's early in the morning here. I'm in Baltimore right now. I got up at 6:00 this morning after not a whole lot of sleep, so forgive a couple of typos here. All right. Off to the races.

So while Claude is thinking away, Debs, what have you been using AI for recently in this kind of area, and what have been your recent successes and failures? Ooh, that's a really great question. Yeah.

So we are, I say For a lot of our summer training, AI is a big theme. It's a big topic.

Yeah.

We are sprinkling AI use throughout the workflows, is how I'd describe it.

So every key topic that we're delivering training on, we are adding an AI element to it. It's really important- Yeah ... that if you're a new analyst, you still need to know how to build a DCF.

You still need to know how to spread the comps for a company. But you can definitely now automate it, once you're comfortable with the fundamentals, that you can then automate it with AI.

And so that's what we are doing at the moment, is at the end of every key topic, is showing the workflow application of AI.

So the use case.

Yeah.

But then also we're spending a lot of time making sure that there's sense checking going on, checking the outputs like we're going to do today, hopefully, when we've got the Lululemon DCF. And then also making sure that there's a really strong analyst overlay. Because fundamentally, if you're an analyst and AI is doing all this work for you, you can't just then hand that over, even if you're happy that it's been sense checked and it's correct. You can't just hand it over for review.

There's usually some element of judgment that's needed.

You need to be able to be comfortable having a conversation with a client about it.

So I would argue that the professional skills become even more important. Your ability to talk to the numbers, your ability to use your judgment- Yep ... on top of the calculations becomes even more important. So we're really trying to integrate that into our training.

In terms of the success- I very much agree with that. Actually, why don't we-- let's come back to this- Oh, yes ... in just a sec, because we've got our first response back and first question here. Not question, rather, here's my game plan.

So here's my plan for a banker grade Lulu DCF.

Pull the data, it's saying it's going to pull data from the latest 10-K. It's going to, let's see, use the 10-year treasury from official sources.

It's going to put together a WACC sheet, CAPM cost of equity, a DCF sheet with a bull care base case, a-- sorry, bear case, base case, bull case.

Let's see. EBIT margin, DNA, CapEx, and working capital, terminal growth, WACC premium, and an index-driven selected case.

I'm not sure what it means by index-driven selected case.

Let's see what it comes up with there.

Mm.

Protection evaluation. Oh, okay.

Five-year revenue to EBIT to NOPAT to unlevered free cash flow build.

Oh.

Mid-year convention discounting, perpetuity growth, terminal value, EV to equity bridge, implied price per share versus current.

Sensitivity tables, three five-by-five grids at the bottom with a few different sensitivities. Formatting, that bold bracket style, dark blue section headers, blue fonts, inputs, black formulas, green sheet links.

Let's see. Banker number formats, blah, blah, blah.

Actually, for a pretty simple prompt, those questions aren't bad. Note, Lulu has a January fiscal year end, so fiscal year 25, end of Feb 25 will actually be the last full fiscal year.

I'll check the more recent 10-K ending Feb 26 is out.

Wow.

Okay, it's asking now for questions.

Graham, this is really impressive. Does this look right? Anything we want changed, projection horizon scenarios, or specific assumptions? No, but a couple of things I would just pick out of there.

So this is already quite advanced because it's identified it's going to use mid-year discounting.

Which when you start doing basic DCF, you assume that all the cash flows are going to occur in one year's time.

And the reality is that most- Yeah ... bulge bracket investment banks, you're mid-year discounting, because we all know that cash flows don't suddenly magically appear on the 31st of December in one year's time. They are generated over the course of a year.

And so the use of mid-year discounting shows it actually knows how investment bankers think. I think that's really impressive.

Yep.

It's also identified the approach that's going to be used for calculating terminal value. There are different approaches that can be used.

We can use a multiple in the terminal phase, or we can use a growth perpetuity formula, which has a growth assumption in it.

So I think the fact it's identifying these features shows, it gives me confidence that it knows what it's doing.

You make a point on the terminal value.

There's two things I want to do.

One, instruct to present the terminal value two ways.

Let's show us a perpetuity growth and an EBITDA multiple approach. Now, I remember the other point you were talking about last time, which Claude didn't even think about this at all, was the date from which we are performing this DCF. And you were saying in your equity research days, you used to use daily discounting, in essence, where I guess you've got a date that you are, I assume a cell input that you're changing basically for today, saying, "Give me the value of Lululemon as of this date." So do we want to instruct Claude to say, "Give us an input for, in essence, today's date, and discount back to that period"? Absolutely. If we specify the valuation date is today, and that's the 14th of July, it will have to make a few adjustments.

It'll have to do a time apportionment for the first year's forecast because you only got future cash flows that need to be discounted, and then we're discounting to exactly today rather than assuming a year or half a year.

Sounds good. All right. I'm just getting a little prompt together here.

Now, it sounds nice and easy. We're asking it to do all these things, but we're doing this because we're really comfortable that we know the process behind it. So as we said earlier, when we're training on DCF, it's really important that you do work through the mechanics, you understand all of the different elements before you start asking AI to do all the heavy lifting for you.

Because then you can sense check the outputs.

Agreed. Okay, so I just said, "This looks good. Two additions.

One, provide two different terminal value calculation methods, perpetuity growth and EBITDA multiple. And two, provide a valuation date as of today, July 14th, 2026, and allow that date input to be changed if we want to see the valuation as of a different date." Excellent.

Happy with this? Yep.

Okay. Enter.

All right. Now, while Fable is building, I'll keep my eyes on it.

If it comes back and has another question or two, then we can pause and revisit. But to your point a few minutes ago around professional skills being a lot more important, I really fundamentally agree with this.

Now, I think we've probably covered this in some way, shape, or form before, but one of the things I talk to a lot of the people I'm training about is, one, I think you guys are starting this work at a really interesting time in the sense that you have these tools that will do a lot of the initial heavy lifting, say data analysis, some of the grunt work, on a much more automated basis, and your life is in some ways a bit easier. But what it means is you're not going to have as much experience going from start to finish, building from scratch, and knowing where, one, AI can make mistakes, and two, where you're going to find mistakes in models just more generally. So you've got to be a lot better at looking at something and spotting trends and spotting errors and knowing what kind of questions to ask.

So a lot of what I do in the classroom now is looking at model output for a model that we've built and just saying, "Okay, what doesn't look right here? Let's talk about this trend.

Can we explain it based on something else that we know about, say, the inputs to our model, or is there an error we need to go back and check?" And that's the skill set that analysts need to develop just almost immediately these days.

Yeah, I completely agree. And it is really difficult to make that leap without having done all the building yourself.

As you said, building things from start to finish can be a good education in knowing where things can go wrong- Yeah ... knowing where things can get fudged.

And having to leapfrog all that and go straight to being able to read the outputs is going to be really challenging.

My top tip on that, I'd be interested to hear yours, Graham, is to look at as many examples as possible.

The more that you've looked at example DCFs, if you hit the desk and you've got access to previous versions for other deals, for example, or if you're in research for other ones built by your team, you start to get a sense of the key ratios that you expect. And that, for me, is really powerful.

When I'm sense-checking, I often look for what I know is a normal level of CapEx to revenue. What's a normal- Yeah ... level of working capital build? The conversion of NOPAT to free cash flow.

You start to get a sense of those kind of numbers that you're used to, and that's where I tend to spot errors. But Graham, what do you think? I think it's a combination of that.

I also think, one, so I started a new-- They're a mix of summer analysts, summer interns, or, sorry, full-time analysts and summer interns. I started a new training yesterday, and we were going through some basic Excel and financial statement modeling.

And certainly, as it relates to something like financial statement modeling, I think it's really easy as an analyst to just go through the motions, follow the instructions, and get to the right answer.

I spend a lot of time talking to people about the assumptions that go into the model and say, "Okay, we have this financial statement model. We have this set of inputs and assumption drivers. I want us to figure out which of these is the most important." And when we look at some of the model outputs, really try to explain trends based on this set of inputs that we have.

I'm sure you probably use the same example.

We use an old, I say old, it's probably five years out of date, because dates from this perspective don't really matter, right? It's more about the actual mechanics of putting something together.

But it's an Apple financial statement model as of, I think 2018 is the last actual forecast year, or sorry, the last actual historic year. And then 2019 forecast a revenue decline. 2020 has revenue going back up.

So because in an FSM model, as an example, you have so many things tied to revenue growth, you get some funky-looking inversions. You have some working capital unwind in 2019 that then reverses in 2020. So I'm always making sure as we're going through to say, "Okay, what about this doesn't look right?" Or, "What about this looks funny?" And then based on our assumptions, does this make sense? That's a really good point, and I wonder if we might spot some of that with Lululemon because, as we said, it's been going through a bit of a turnaround, and I think that does make it quite challenging.

What we tend to like, well, not like, but what we expect with DCF and forecasting in general is nice, smooth, gradual growth.

And that's certainly- Yeah ... not what we would expect for Lululemon.

Exactly.

How's Fable getting on? Okay.

So we've got another stop here now where it says, "Okay, here's the data I've pulled. What do we think?" So it pulled a few years of actual results from 10-Ks. It's got the share price. It's got diluted shares outstanding.

It looks like it's done, okay, diluted shares outstanding, 114 million.

Let's see, 111 million common plus 5.1 million exchangeable as of, let's see, at fiscal year-end per the 10-K.

Less 2.2 million quarter one buybacks per the Q1 release.

Okay, this has actually done a decent amount of digging in terms of just news to say, "Okay, what is the right diluted share count here?" Okay, net cash, zero debt, undrawn $600 million revolver.

I think we knew that from last time, so that looks like it's remained constant.

10-year Treasury, 4.58%. Beta assumption, 1.2. Says no official source publishes beta.

Blue input that you can flex. So it's being honest about the fact that it's just made an assumption here.

Mm-hmm.

Tax rate, 30%. Management guidance. Okay.

Then we say, okay, context, 2026 guidance was cut in June.

See, we've got the-- I don't need to run through all the numbers here, but we've got the revenue growth, and margins for the bear case, the base case, bull case.

All cases, DNA 4.5% of revenue, CapEx going from 6% to 5.25%.

Let's see. It hasn't specified why it's making that assumption, but I'm sure we can ask the question. And the WACC at 10.6%. Valuation date input July 14th, 2026, drives all mid-year discount periods dynamically.

Terminal value, perpetuity growth, and EBITDA multiple approach side by side with a method toggle feeding the headline bridge.

This looks a lot more robust. I seem to remember last time the WACC was a bit odd, wasn't it? For the- Yeah ... for the original DCF that was built.

So 10%, that's kind of my go-to starting point.

Right.

So this is looking sensible. And even the CapEx, I think last time it gave a really punchy CapEx assumption.

So yeah. Let's see what the outputs are then.

I just said go ahead and build it. So let's let it crank over here.

Okay.

If it stops and asks another question, we'll take a pause.

Okay.

Otherwise, we'll see what it comes up with when it finishes building.

So Graeme, whilst it's doing that, I've got a question for you because, when you said you wanted to put Fable to the test, I will confess, I wasn't really familiar with Fable. I haven't used it at all.

And then you mentioned just earlier today about the fact that it's kind of allowing some free access. Do you know the situation there? Is this kind of like a bit of a teaser to kind of get people locked in? That's a good question. I don't know if you know the backstory behind this model, but I'm sure people have probably heard of the Anthropic Mythos model, the thing that was basically so good they had to withhold from public release.

Yeah.

Fable is the more general release version of it, and it's basically, the way I understand it is it's the same model but without some of the crazy cybersecurity kind of hacking logic that people use Mythos to bolster their security with. That's the one they really don't want to get in the wrong hands. So they released Fable 5, I want to say about a month ago, there or thereabouts.

And then pretty quickly thereafter, the US Department of Defense said, "Hey, this is so good. You can't put this in the hands of anyone.

Only US citizens should be allowed to have access to this model." And really the subtext there, I think, is it's basically the US government saying, at this point, we're picking OpenAI and Sam Altman as the winner in this space. I'm not sure it was really motivated by anything other than just political goals. Long story short, they, as in Anthropic, came to an agreement with the US DoD to get Fable released again. But to the sort of disappointment of a lot of normal users, and I say normal users, I have literally the $200 a month Claude subscription, so it's not like I pay nothing for this. But Anthropic has said, "We're going to give you Fable 5. It uses a lot more usage credits.

You only have so much access to it, and after a certain date, it goes to usage only." So it's not included in your plan credit anymore.

You have to basically put the quarters in the machine every time you want to literally get some Fable 5 tokens.

Now, I don't know where this ultimately goes because it's kind of funny, when you log-- Let's see if it even shows up in this interface here. No, it doesn't. I think it's probably in the normal Claude app.

But whenever you go to choose the model, it says, "Fable 5 included in your plan until July 7th," then July 12th, then July 14th.

Ooh. Then it's July 19th. So I don't know if they're just pushing it back to get as many people addicted to it as they possibly can, and then they make you pay per use.

I have to think at some point, though, ultimately the world is competitive.

OpenAI will come out with something good, and then Fable will get included in your plan usage.

But then it's going to get superseded by the next latest and greatest model.

So it's always this little chicken and egg situation.

But I have to assume the reason they're doing this is to try to maximize their run rate revenue in advance of the IPO.

Yeah. Okay. Yeah, that's a good point.

But it is interesting, isn't it, that actually we started off, all of us, I think, started using AI with no real concern about limits and even really the cost of it. It seemed something that you could use in a very affordable way. And already- Yeah ... so many of us at work are finding we have subscriptions, but on a daily basis, you reach your limits.

And then you have- Yeah ... to request more tokens. And you can see at some point, we do reach that point where you're now having to make an economic decision around the cost of using AI versus someone actually having to go through the thought process themselves.

And so far, it's been so much cheaper.

Yeah.

But there will come a point- Well, and it- ... where you're thinking, hang on, there's an analyst that you can pay 100 grand a year versus a number of tokens which are creeping up in price.

100%. And the other thing is, to the point of having to know how to do this stuff yourself, what happens when your tokens run out? You're like- Yeah.

... "Oh, wait, I've never done this before. I don't know what I'm doing.

I just have to stop working." Yeah.

No, you still need to know what you're doing.

Yeah, absolutely. Right. How's it- All right. It's working ... getting on? It's just churning. So the one, not the one, but one of the main distinctions between Fable and some of the other models is it will just crack on and get on with it. It doesn't necessarily stop and ask as many questions. We had the two kind of question rounds. Is this approach right? Okay, here's the data. Am I okay to go? Now I'm very much expecting ...

the next step here to be pretty fully thought out, unless there's a major question or something happens along the way and it says, "Hey, what do you think about this?" So let's give it a few minutes and see what Altimeter pulls together.

But it looks like if I look through the steps it's following, so it's setting up sheets, build the WACC, controls and market data, build scenario assumption blocks, build financial projections in DCF, build discounting, terminal valuation, and valuation bridge, build three sensitivity tables, format, and then verify. Doesn't sound like a crazy list of steps, so let's see how it goes.

So Graeme, we're demonstrating a really clear, obvious use of AI in terms of building a DCF, but we've been reading in the news, haven't we, about other tasks which AI is being used to kind of replace some of the workflows that investment bankers traditionally do. Can you tell us a bit about- Yeah ... one of the deals that you were highlighting to me earlier? Yeah, we just saw, I think it was in the last week or so, there was a story in the journal about CVC using an AI agent.

This is an attention grabby headline, right? It says, "CVC uses AI to replace investment bankers to run a sale process." And it's that, but it's also not that. Basically, CVC was selling a portfolio company, it's called Sprouts, and the article was touting the fact that in the data room was an AI agent that was there to answer prospective buyers' questions. Usually, you'd have a team of investment bankers and analysts really prepped to go through the data room and come back with any kind of questions.

Here, we've got an article basically saying, "This is the first time that AI has been used to do this." Actually, I would be 100% shocked to find out that this is the first time this has actually been the case, because if you're going through a data room these days with access to AI, are you using these tools to pore through the data room and answer the questions you have? 100% you are. I think this is the first time where the seller or the banker, at least publicly in the news, has said, "Hey, as part of our data room, we're in essence including an AI chatbot that will help you go through all the data and answer any questions you might have." And by the way, I'm sure we've talked about it before, but I think that that skill in particular is one of the best uses of AI today. Just going through these large- Sure ... datasets, answering questions, pulling out trends.

In the FSM modeling exercise I was running through with these first year analysts and summer associates yesterday. Part of the exercise historically was going to public company filings, going through 10-Ks, pulling out data, and I said, "I'm not going to make you guys do this." Not least because you guys are in mid-market private equity.

You're not going to have to go through public company filings, do diluted share count calculations, all this kind of stuff when you're building a model.

But one of the things I talk about is just how good AI is. If I take that model template and say, "Okay, Claude, GPT, go pull out Apple's last three years of historic filings literally back to 2018." Fill out the inputs for the sheet, pay attention to the notes in some of the columns saying, "Okay, grab this piece of data from this page," and it will literally fill out the entire input sheet.

We'll have a comment in every cell saying, "This is where I got this number from, this page of the 10-K or the 10-Q." For that kind of work, it's really impressive and saves a ton of time.

So it doesn't surprise me that we're seeing this being the first public use of AI that we're talking about in the news.

But no doubt, privately, this has been the case for quite some time already.

Okay. But it is kind of a step on the journey to some of the work that's previously been done by investment bankers being done by AI.

And that- Yeah ... is potentially a risk for some of the fees that are being generated by the investment banks for their sell side process, isn't it? Because sell side process has a number of steps. You've got to think about the marketing that you're doing, trying to generate interest from the buyers, managing the data room.

Yeah.

And as you say, part of that is now answering questions from prospective buyers. There's still other stuff that's probably still being done by the actual bankers themselves, things like advising on the valuation. We've talked about things like valuation opinions and things.

So those are still- Yeah ... there might be small cogs in the process which are gradually being transferred to AI. But it is a risk, isn't it? That the bankers' fees start to go down if more and more of it can be done by AI, and specifically, that more of it can be done by the firms themselves rather than the advisors.

Ooh, when do you think we're going to have the first AI signed fairness opinion where- Ooh ... Claude says, "I think this deal's good." That is a really good question.

Or fair enough.

Yeah.

It's about risk, isn't it, really? Because the fairness opinion is the actual work involved, as we saw when we looked at the, it was Paramount, Skydance, Warner Brothers deal.

The actual work- Yeah ... involved looked quite limited, but it's the risk that you're taking on because you can get sued if you provide opinion that doesn't hold water.

Yeah. No, 100%.

Interesting. Right. How's the model getting on? We- Let's see. It is. It keeps building.

It said, "Isolating which formula failed." So it's made...

I don't know exactly what it failed at, but it highlighted that it had made a mistake, so now it's going through and correcting itself.

It's still crunching. The failure was the LET formulas. What's LET, Debs? Do you have any idea? I have absolutely no idea. Well, it says it's not- Is it- ... supported in this Excel build, so maybe- Is it- ... I'm only kidding with Excel.

Is it hallucinating some finance acronyms now? It's quite good that it self-corrects.

For a long time- Yeah ... as you say, it was really needed a little bit of coaxing, didn't it, to come back to you and say, "Oh, I found a problem." Yeah.

"What should I do?" No, 100%. It's just getting on with it.

I'm very much expecting by the time it finishes getting on with it, we'll have something that's at least semi-usable.

Am I expecting this to be perfect and as good as a really good analyst if they've gone through and done the job properly? No, probably not. But on the basis of some really quick prompting, I'm expecting something at least semi-usable.

So Graham, I can actually see some- So we give it a few more minutes here ... I can see some assumptions, though, on the DCF.

Shall we have a quick look at the assumptions on the DCF tab, and see if there's anything- Yeah ... that jumps out at us? Let's see. Okay. I didn't want to interrupt it too much.

I was going to resize some columns just to let us see everything, but it's already gone through and done some of that.

Okay, so we've got a case selector.

Let's see. We've got a choose function, not necessarily my favorite- Uh-huh ... function for this, but you know what? Not horrible.

Okay, terminal value method. We've got, okay, one perpetuity growth, two EBITDA multiple approach, current share price, dilute shares outstanding, market cap, cash and cash equivalents.

We already know we have no debt, so we have negative net debt, enterprise value, and WACC of 10.6%, which I think is about what it said it was going to do- Yeah ... in its setup assumptions. And we'll go through, and we can take a look at the WACC calculation in a second. Okay, so we've got three cases here. Bear case, base case, bull case.

Bear case looks like we've got revenue down 2% this year, flat next year, and then 1% growth 2028.

Base case down half a percent, 3% growth and 4.5% growth.

And then bull case, 0.5% growth, 5% growth in '27, and 7% growth in '28.

I don't know the company that well, but directionally, just in terms of thinking about what we know about historic results, what management guidance is for the next year, just looking at, say, the variance in these cases based on the number of years, there's not a huge range in the bear case to the bull case this year. Right? And even in the bull case, revenue growth in '26 is only 0.5%. Right? And then as we get in the outer years, it's making some slightly divergent projections about what those three cases are going to deliver.

I think that displays at least a bit of thought that is above and beyond just simply saying, "My bear case is -5% every year, my base case is flat, and my bull case is 5% growth every year." Absolutely.

The thought process, I think, is actually halfway interesting here.

And I've got some numbers in front of me for consensus, and it looks like those base case numbers are pretty much in line with consensus.

It's found some numbers somewhere which are publicly available, showing consensus information. And that, yeah, the starting point- Yeah ... usually for bull and bear cases, you're going for 100 to 200 basis points. That's 1% or 2% above and below the base case in terms of revenue growth.

Yeah. And then an adjustment to margins based on what is reasonable within the industry.

But yeah, it looks like a pretty sensible profile across all of the cases. And also, as is normal, we only sensitize revenue growth and margins.

We don't sensitize usually the other assumptions unless there is something really specific that could happen in terms of store growth, store expansion that affects only one of the cases.

So- Yeah ... all the focus- Yeah ... of the bear case and bull case is around growth and margins, which is, yeah, sensible.

And by the way, I just hid the window because it said the model is complete as described. So now we've got something, at least at this stage, to go and take a look at. So I actually just want to see also just how it's built the model. We've got a selected case assumptions.

This drives the model. We're using some choose functions to pick between those three cases above. Again, for a model with this number of inputs, i.e. not that many, it's not a horrible way to go about it. I'm not super mad about this. Okay. For each, we've got an input for terminal growth rate and exit multiple assumptions. So on our bull case, we've got a 9.0x exit multiple, 8.0x, and 7.0x. And then for perpetuity growth on our bear case, we've got 2.5%, then three, 3.5% growth on perpetuity.

Okay. Doesn't sound crazy. All right.

Now let's look at our unlevered free cash flow calculation- Mm ... because this is what it really did not think about, or think about properly last time.

We got three years of historic results.

Just want to check these comments here. It's saying exactly where it pulled each one of these figures from. Then we've got a calculation for revenue growth, which I'm assuming is just going up and pulling from our case assumptions up above.

We've got EBIT driven by an EBIT margin.

We've got depreciation, amortization.

Depreciation, amortization, let's see how we're calculating this.

We have got an assumption for percentage of revenue.

Again, for a high level model like this- Mm-hmm I don't think that's crazy. Right. Gets us our EBITDA.

Let's see. Then we've got, it's a slightly funny presentation, but EBITDA, and then less taxes on EBIT. It looks like this calculation is actually correct.

And then NOPAT. So EBITDA is just a presentation line here. It does look like it's actually calculating NOPAT correctly. It's taking EBIT, less taxes on EBIT.

Then add back DNA, subtract CapEx, changes in net working capital. So it's got an assumption for net working capital, which is not a huge net working capital impact.

And actually, how have we thought about that assumption there? Again, okay, just net working capital balance as percentage of revenue.

Mm-hmm.

Again, by the way, how do you forecast net working capital balances when you're doing a high-level model? Do you pick a percentage of revenue and forecast that way? 100%.

I think a couple of things, a couple of sense checks that I always run, CapEx ahead of DNA, usually, because you need a company to spend more on its growth than the depreciation on its existing business.

So you've got a nice- Yep ... ratio of CapEx to DNA there.

The net working- And one thing, by the way, on that, what I don't know is based on if we'd gone through and really said, "Okay, what is Lululemon's expansion plan for the next couple of years?" Mm.

Would you make some specific assumptions here? Yes, this looks a bit more high level, but for this stage, I think that's probably okay.

Absolutely.

Sorry, go ahead.

And then the other thing is something that you mentioned that is to do with the net working capital, that sometimes when you've got a contraction in revenues, the fact that you can end up with some slightly wacky working capital flows. Because usually, a contraction in revenues, if a company's struggling, it's not usually going to result in cash generation from working capital.

If you link working capital to revenue, you can end up with this odd working capital inflow. They've got a very small $3 million, whatever. So, it's a small inflow, but it's actually not too crazy- Yeah ... which, again, is quite reassuring.

Yeah, agreed. Okay, let's look at the actual, let's see, discounted cash flows, fiscal year ends date. We've got a, let's see, year ends... This is, in essence, just running a YEARFRAC formula here.

Let's see.

Unlevered free cash flow included.

Okay, what's it doing here? Fiscal year-end date.

So that's a bit weird because I would usually expect it to be the year fraction to be one for every year except the first forecast year.

Because what you're doing, if you're discounting to July and you've got a January year end, you've got a certain number of months until the first year end, and thereafter, we're going to be discounting full years. So I'm not quite sure why you end up with more than one year.

Okay. In essence it is that, but it's looking at the fiscal year-end date, and it looks like some year's fiscal year end is January 28th.

Then we're going- Oh, it's a 52-week- ... February ... fiscal year. Yeah.

Yeah. Okay.

Okay.

So that, I think, is the issue here.

Again, this isn't going to change the math- No ... materially at all.

Okay. Then we've got our mid-year discount period, discount factor.

Let's see. Where's our discount right here? Okay. Weighted average cost of capital.

Okay. Present value of unlevered free cash flows.

Then let's look at our terminal value calculations here.

Two different approaches. Terminal year free cash flow, one plus the growth rate.

We're using the normal perpetuity discount formula here. We've got an implied EV to EBITDA multiple based on- Good sense checking. Yeah ... that perpetuity growth rate calculation. Method two, EBITDA.

Okay.

You know, it's not crazy.

It's not crazy, is it? I mean, even just seeing the- Really ... implied multiple, high single digits, low double digits is kind of the rule of thumb, isn't it, for a terminal multiple.

Terminal value as a percentage of enterprise value, 73.5%.

That's a nice little sense check in there.

Make sure you've not got all of the value baked into that terminal value.

And by the way, we didn't even prompt to say- Yeah ... "Give us that." The prompt we said was, "Follow Bulge Bracket investment banking best practices," and it has just figured that out.

Yeah.

It's just made a decision. Hey, I assume they might want to see that.

Right. Actual valuation summary. Okay.

Enterprise value less net debt equals equity value.

Net debt is negative, so our equity value is higher than our enterprise value.

Okay. Diluted shares outstanding, implied share price, current share price, implied upside, 45%.

Oh, yeah.

Ooh, okay.

So this is the crux of it really- Okay ... isn't it? So basically, Lululemon- And it's- ... as a company, is going through some really challenging times.

We mentioned the share price is really low at the moment compared to even a year ago. And the DCF is still showing a fundamental value for the company, which is well above the share price.

That's really interesting, isn't it? Now, from memory, last time we did this, Claude basically solved for a current share price- It did ... for valuation- It reverse engineered it ... that was basically in line with the current share price.

Here, it's actually taking a view.

It's saying, "Hey, I actually think there's upside to the share price here." Now, the one thing that it's still, from my perspective anyway, this formula I think is technically correct.

Anytime we're doing a sensitivity analysis like this, I would always just use a data table because it's a lot easier to figure out what's going on. I'm sure we could go back and prompt Claude and just say, "Hey, these sensitivity tables use an Excel data table instead of this crazy sum products," whatever's going on here.

I can't audit that formula. I don't know about you.

I don't know if you can look at that, Debs, and just say, "Yeah, that's right." Absolutely no way. But I think what's really interesting is that clearly AI The default approach now is to make everything as auditable as possible, which in a sense is good.

Yeah.

It's just you do end up with these ridiculously convoluted formulas, which in theory are auditable, but to the average analyst, because you never build these formulas yourself, as you say, you just run the data table, which is basically a little macro, and all the outputs are hard-coded.

We would never know whether those formulas are correct or not.

Exactly.

Yeah.

But what is important is at least can you identify trends and decide whether these are directionally correct? Absolutely.

Because here we've got our share price calculation sensitized by the WACC and terminal value growth rate.

Share price going up as we increase the growth rate, going down as we increase the cost of capital. Same thing with revenue growth and margin, beta, risk-free rate. Can you take the inputs and figure out, do you already know what it's supposed to do in terms of affecting the outputs? And at least look at something and know if the trend is right.

Mm-hmm.

That's your first step. Would I go back and prompt to replace these with data tables just to make it really easy? Yes.

But as a first stab, it's actually pretty good. And if you have Excel set to partial calculations, this is still going to calculate every time without having to hit F9. So I guess a little bonus there.

We are running close on time for today.

I know we've both got to get out and teach and get into the classroom, but we talked last time and I said within a couple of years, we're going to get to the point where AI is going to be, at least for this kind of work, as good as that first-year analyst.

Based on this prompt, build a DCF for Lululemon, follow investment banking best practices.

And fine, we gave it some specific instructions in terms of, hey, provide the terminal value on two different methods and give us a date that we can use for the discounting. That was it.

Yeah.

That was literally all we gave it. And this is quite a lot more usable than the result we got last time. And that was six weeks.

So Graham, am I right in thinking that last time we graded it and we basically were pretty down on the results? I think I gave it a C.

Yeah, I think I gave it a B minus, and I think that was being generous.

I think you gave it a C something.

What would you give it this time, Graham? Ooh.

Okay, based on, in terms of output per prompts, to get the full grade, I'd have to go through it.

I'd have to go through all the source filings.

Yeah.

I'd have to read the press releases and see has it actually parsed all that stuff correctly. Has it come up with the right view on bear case, base case, bull case? That's the real detail work you'd need to properly grade it.

But at a high level, just taking a look at this, I would say we're firmly in the A-minus territory here, I think.

Yeah. I would say- For time spent ... A-minus or B-plus. It's done much more complex processes than last time. It's done much more sophisticated DCF calculations.

There's one little niggle I have, and it's something I didn't pick out last time, which is on the debt side, it treats debt as zero.

The reality is they've got a huge store portfolio, which is leased.

Technically, I would treat that as debt-like.

It hasn't flagged that.

Yeah.

So I would want to roll up my sleeves and dig through and just check that that isn't creating an actual error in the outputs.

Yeah.

But I think I'm warming to this now.

I reckon B-plus from my perspective, which, as you say, in six weeks is a massive improvement.

Yeah, in that short amount of time.

So should we do this in another six weeks and see where we are? Hey, we haven't even tried GPT for this.

Yeah.

I think generally speaking, I've had better luck with the Claude Excel plugin.

Likewise, yeah.

So we can pitch them against each other.

Maybe we'll do a side by side, give them both the same prompt and see how they do.

But in not that much time, some pretty impressive progress. So Fable 5 does seem to be living up to its reputation.

Yeah.

And people continue paying for it when you've got to buy the credits and it costs $50 to put this model together. Would I just go through and just do it myself? Maybe.

That's the question to answer the next few weeks or so.

Absolutely.

We'll see what Anthropic decides to do about it.

Great. Well, thanks to those of you that listened to our little challenge for Fable, building a DCF for Lululemon.

I hope you found this episode interesting.

That's all we've got time for this week, but that's thanks from me and see you soon.

All right. Thanks, Debs. And by the way, if you want to see us do any more of these, or if you've got a modeling challenge for either Fable or the latest OpenAI model, let us know in the comments down below, and we'll try to get to it in an upcoming episode. But until then, take care everyone, and we'll see you same time next week.