Using AI for Financial Modeling: Systems, Skills Files, and Validation

I think is also an important consideration as we'll try and display in this in this session. You you have to understand the jagged edges, right? AI is certainly AI and finance is not a solved problem, right? It's not a solved problem. When people say AI XL is a solved problem, I'm like you are wrong, right? You are very wrong. We have made insane improvements in the last 6 months, but there are still many jagged edges that are very problematic for the investing process. So what we we're trying to do here is shine a light on those Jacob processes, think about workarounds, but al be be very objective about where these tools will solve.

Thank you everyone for for joining today in this third installment of this random webinar series on augmenting investment process with AI. Um today we will talk about Microsoft Excel and what can and and my experimentation can't be done with AI which I think is in a way it's as an important piece to understand the true limitations of these tools. Um so I'll show you what I've learned in three weeks. So hopefully you don't have to spend the same amount of time that I did going down all of these random rabbit holes uh to preempt a couple commonly asked questions. Yes, the recording will be sent if you're registered on Zoom also be posted on YouTube which we'll release on our our Twitter fundamented edge. Uh no, we will not share any of the skills.mmd files. uh we we're beginning to do that selectively with clients and uh sort of firm clients and students of our programs but we won't be doing that in the open webinars. Um so uh with f without further ado I will go ahead and share my screen and we will jump we'll jump right into it.

Uh disclaimer nothing we discuss will be investment advice. If it was you probably shouldn't listen to me even in a good year. I was right 52% of the time in the bad year. I was catastrophically wrong. So, don't listen to some random person on the internet. Always do your own due diligence. Uh reminder of my background. I I was an investor. I was a fundamental uh analyst and portfolio manager for 13 years. I randomly backed my way into seats where I sort of had a front row seat to the uh combination of fundamental and quantitative investing. really this the outside of my career in 2008 at a Tiger Cub. Um but I found it deeply intellectually fascinating um to think about the combination of fundamental and quantitative uh investing. Uh and so I I bring many of those priors uh to this work. Sort of the deep engineering mindset of the fundament fundamental investors uh workflow and sort of a quantitative lens of taking what is uh is often a very intuitive or tacit approach to investing and systematizing that. We do that in our analyst academy where for 60 hours we break down fundamental investment processes. So that's a good starting point to think about a augmentation uh sort of a a combination of uh the priors from quantimental with just an experimentation journey for the last two or three years. We started a web webinar series on on AI vendors in fall fall of 2024. Um, I just I just looked actually I ignored a direct message from Gabe from Rogo for like 18 months before I finally got back to him and then they just raised at a $2 billion valuation yesterday. So, congratulations uh team rogo. Really my learning approach is taking the priors of uh priors of the lived experience as an investor. Um and a lot of a lot of uh a lot of conversations uh with people who are in the trenches building uh finding the right people to follow on YouTube and Twitter has uh has great leverage. Uh most of what you'll see is just simple parallel experimentation. I think the best way to learn AI really is just start using it and AB testing. Do it the old way, do it the new way. And you'll be surprised. You'll be surprised and often frustrated. Sometimes things that work one day won't work the other uh due to the non-deterministic stochastic uh uh nature of these tools which also are often throttled etc. throttled and nerfed. Um but I'd say that's probably the most helpful process. Um so use me as a guinea pig. I know you all are very very busy. And then we ran an AI for investment research cohort starting in September where we had 150 analysts and PMs really learning a lot of this together. Um, we're starting to do more client engagements to help funds actually adopt agentic investment processes. And part of my experimentation has been building an AI native mock portfolio, pretending I'm back in the seat managing capital again. I'm at a little bit of a sticking point on that. So, I'll I'll update you on that. Excel is actually a key sticking point on the ability to do that as a side hustle. Um, and then of course, I always sort of half jokingly point to Cunningham's laws. The best way to get the right answer is to post the wrong answer on the internet. So, I get a lot of people telling me I'm idi idiot on X and sometimes I am. But, you know, if they point out things that uh I can learn from, uh, thank you Cunning Cunningham's law.

So, I don't know everything, but I know one thing. I know that we all have the same 3,000 hours in a year. Um, you know, roughly roughly 60 hours a week is sort of, you know, by sideish standard. You know, the 7 to 7 Monday through Thursday, cut out a little bit early Friday. um than a Sunday afternoon, you know, give or take five or 10 hours um sort of standard for for for for an investor. And so part of the approach in thinking about AI augmentation is starting with that known constraint of ours and started to think about you know the exercise of will AI generate alpha starts with that known constraint. So to me I sort of think about uh coverage capacity. This is more of a multi-manager mindset for sure. Um but a typical multi-manager mindset is there's sort of a cap on the number of companies an analyst or a PM can cover generally that you don't see that much above 50 simply because the the general consensus on good coverage good institutional grade coverage is we have our own models we update our own models we meet with management regularly we there's a lot of motion on earning season there's a lot of just desktop research reading sellside research we're reading KS reading 8Ks and primary research talking to the management team surveys etc. So at a very broad very broad strokes you know 60 hours per year per name that's just the the the staying in the flow I guess is part of building a model as well too. So it's sort of very rough explanation of the coverage capacity the workflow of an institutional investor. So before we start about start thinking about scaling this we have to think about you know how do we scale these efficiencies in a rigorous way what part of these areas can we actually scale and so if you sort of take it on faith that these tools are ready or near ready for scaling fundamental investment process let's dig in and sort of experiment and go through the individual pieces. There are certain things like meeting management where you know we're not there's sort of no clear uh uh line of sight into scaling but one of the big areas of scaling is is modeling right if I can click a button and get a financial model that's that's that's that is no different from my existing financial models if I can click a button and upgrade a model in a a comprehensive uh accurate way that starts to do one of two things. one that starts to really scale the names I can have under coverage in that same 3,000 hours so I could cover more names as a healthcare investor. I always went back and forth on covering European names. You know, just the scaling of what I had to do the blocking and tackling in the US generally meant that I ignored European names. I could pick up Europe and add that sort of orthogonalized alpha stream to my book or I could go deeper on existing names. I could take this I could take these these uh hours saved and I could take you know I could all of a sudden be doing 30 hours per name on primary research or 12 hours per name meeting meeting management right so I can redeploy those hours in in workflows that deepen my comprehension deepen my comprehension so I'd say you know as in sort of the evolution of of AI AI moving into Excel is a big deal is a big deal. Um the question is what is what is it what does it mean? Right? And so let's first talk about from the investor lens what a model does, right? Because the the the the the design structure purpose of a hedge fund or loan only model is going to be very different than a sellside model and certainly different from a um uh from an investment banking banking model. The biceet equities model has a few characteristics. It's it's built to be maintained for years, right? You're updated every quarter under time pressure. It's built as a communication tool, right? Analyst sends to PM, analyst sends out to the team. We're all looking at the model as a quantification of a thesis. You It has to be flexible. You have to be able to run scenarios. The model is a thinking tool. It's not an answer unto itself. And so you have to be able to understand, hey, comps up, comps accelerating, decelerating by 100, 200 bips. What what does that mean to EPS? What are the incrementals of the business? How does that flow flow through? So we're really optimizing for fast updates, easy auditing, analytical flexibility. I tell the story, the first model I built, my first hedge fund job, I had manual searchs and VLOOKUPs and all this crazy stuff, and I got yelled at just like, don't do that. Right? The models are meant to be, you know, simple. And some of the models you'll see on the buy side are like you'd be shocked at how simple they are. And that's fine, right? Simple is great. They need to be clean. They need to be functional. They need to be flexible. They need to be in intuitive. And the standard we recommend is a quarterly update shouldn't take more than 15 minutes. If if you're building a model to that takes more than an hour to update, you've missbuilt that model. There are shortcuts and data updates, you know, uh data sections, etc. um that I would argue you can get virtually any model updatable within uh 15 months uh 15 minutes. So they're built to enhance a deep understanding of the business to track fundamentals over time, a baseline for riskreward and to help you navigate earnings and news. Hey, an 8K drops, something happens. Let me pull my model and see what that means to earnings. Earnings cut 10%, stocks off 3%. Hey, guess what? I'm going to sell that. earnings got 3%, stock off 10%, it bounces back next year. Hey, guess what? That's a viable opportunity, right? So, it helps to frame uh helps to frame and quantify what's going on in the in the market. So, the buyside model is a it's a thinking instrument. You can't you can't oneshot a correct buyside model, right? We're not here to press a button, get a DCF, get an answer, and the game is done, right? It's sort of a sort of a silly silly uh silly silly uh use case. Very few investors use DCFS or point in point in time uh price target derivation anyways. They will use them to understand you know expectations as as one tool to understand expectations but what we're really focused here on on on earnings models and a lot of the value the ultimate sort of make or break of the thesis is in the in the forecast right the model is a place to encode the analyst best judgment about a business. Uh so that's that's important when we think about engaging with LLMs where open AI's hired all these investment bankers. There's a lot of in the train in the LLM training corpus. Investment banking uh highly influences the thinking about financial modeling as Buffett highly influences uh the thinking about investment research. That's great if you're an investment banker. That's great if you're a Buffett lookalike. That's a little bit different if you're working at a high velocity multi-manager hedge fund, right? Right? And so we have to take account for that when we think about creating uh the the the system. So we're not building banking transactional models here that support a deal. We're not building cellside equity models here that that uh support coverage and distribution and tend to hug consensus. We're trying to get to the right right answer. Right. And the use case of the model really is five-fold. It's um it's uh you know helping you go deep, right? The mandate in the institutional buy side is know your business is cold, right? Know all the nuances. You know, my first PM would ask me, you know, I talked to him about a company. He's like, walk me through the P&L. Like, what do you mean by that? He's like, start at revenue and from memory, walk me from revenue down to earnings per share down to free cash flow. I'm like, I completely failed on that the first three times. the trauma of that failure encoded that in my brain to understand that when we cover businesses our job is to really understand the nuances the quantitative and the qualitative nuances of the business. So the model is helpful uh to to to do that. Uh in that sense building the model the act of building a model is an educational process right why for you know really five years at least I built every model from scratch. It helps to enforce mathematical rigor in the investment process by translating qualitative insights into precise financial models. Hey the co the CFO said things are a little squishy. What does that mean? Does that mean 100 bips or 300 bips? Hey a competitor said they're taking share. I validated with three other data points. What does that mean to organic revenue growth, right? Quantifying those narratives down to numbers is a really important process of the analyst and the and the financial model is the critical sub substrate to do that ultimately to drive differentiation and v variant perception. Right? The market thinks key drivers going up. I think it's going down. Right? out of covering 150 names at any time maybe I have 10 or 15 opportunities where I clearly see a varying perception the market's right more than it's wrong in my experience but using the model systematically to identify those opportunities to understand what what catalysts are really investable catalysts where does the beat and raise happen where does the miss and lower happen where does something happen in the business that meaningful meaningfully changes the net present value of the free cash flow stream to that business and so so really pinpointing the specific events whether it's a regulatory event, a Supreme Court event, a a you know phase 2 biotech readout etc. And then valuation is really a secondary principle at most hedge funds. They don't start with hey this looks cheap on consensus. They start with let me understand the business build my forecasts and then often you know three-year forecast is no more complex than what's my four year out of number and slap a multiple on that. And so the outputs of your of your modeling process build the baseline of your valuation work. Right? So this is like all really important stuff. Um right and when people say well why do you build a financial model at a hedge fund Brett? Why don't you just use cap IQ numbers or fact set numbers? I'm like you're kind of missing the whole point of what the job is. This is the whole point of the job. And having the financial model as the workbench of the analysts like I couldn't have done my job without a financial model. like you'd be you'd be investing on vibes. Um um and so that's how critical the financial model is and that's why you know I sort of take a different lens on this session as well too versus many of the vendor trials vendor demos you see is like hey press a button you know has a DCF like okay that looks impressive maybe to a neoight but from a practical perspective uh there's real no usability of that model that's that's generated uh not to mention AI slop and hallucinations and a few other things that we'll we'll get into um we'll get into today. Um so really you know the model is the workbench of the entire institutional equity research process. It's it's you know I have the model up through the entire thesis development arc from idea origination to triage to due diligence to you actually building the model going deep on the key drivers because ultimately your model will hinge your thesis will hinge on you know two or three inputs into the financial model. Um so the actual the mechanics of building the model are more commoditized in nature. The real value the real differentiation comes from simple inputs into just a couple cells in a financial model. How does that form you know help insight formation you know how does it help expectations and valuation work understanding positioning what people are playing for etc. Thesis construction. So how does the model sort of your your your um your workbench for building the thesis? Um but also you know at at large hedge funds most your models are built right you have you have existing models a lot of the action happens in the active position management space right this is why I think early stage chat bots haven't been that helpful for hedge funds like help me summarize this 10k it's like well bro I've covered hca for 10 years I've read every 10k um I don't need a summarization I don't need to get up to speed and have an LLM tell me what the how the business makes money I know that Right? I need a tool to help me drive differentiation on three key drivers to help me through the earnings process and the catalyst path and maintenance process and management touch points and understanding news and catalyst and events and understanding how the riskreward adapts over adapts over time. And so just one of many reasons why I think chat bots haven't really hit you haven't reached escape velocity at um uh in the hedge fun world certainly by any means. Um, and part of the reason why agents are so damn interesting because you can start to wrap a lot of this around with with agents. So, we'll we'll talk a little bit about that, but sort of an important point is that investing is a basian exercise. You you you sort of develop a prior and then you you're updating that prior really daily, daily, weekly, monthly. Financial modeling is not a oneanddone exercise. Um and so a big part of the financial modeling uh exercise is building something that is that basian tool to update news flow competitive announcements. It's a tool to know the business cold and then it's a tool to monitor the business uh clo closely. Um and so you know in that sense like an AI slot model or even templates I think um templates haven't really always hit the mark on monitoring the business closely uh element. So what you see if you walk into any hedge fund desk is you'll see folders of you models like this right where you'll see you know every every you know company has a financial model with you know segments and consensus and riskreward and key drivers and the full three-statement model when I was a PM with a nineperson team we had 300 financial models I could click open any model at any time it was updated I could get a sense of what our view was where we sat relative to the street and that's a really helpful uh foundational um uh footprint to start trying to build the portfolio that beats the market because of 300 names are trying to have maybe 25 longs and 40 50 60 shorts and um you're not starting from scratch on every on every name. Now that's again that's a very multi- multi-name. Uh when I worked at a Tiger Cub I'd cover a hundred stocks and I didn't have all hundred of my models updated. If a name wasn't interesting, it would go out of date. It would go out of, you know, updating for two or three, four, four quarters. So, uh, one thing we always tried to highlight in our curriculum is, you know, there's no right way to be an investor. This is a very heterogeneous process. You know, I I'll hear, you know, very impressive investors with $500 million positions being like, Brett, you know, I don't build out quarterly models. I'm like, okay, well, if that works for you, please don't. Please continue doing what works uh what works for you. So you'll hear anywhere from yeah I build all my models from scratch to you modeling is a waste of time. I just use canalyst and I say great right the modeling approach that supports your investment process is the right approach for you. Now a often there's a mentality there's a philosophy at these firms and the tiger cub community it's it was often certainly when I was there was a no no template rule. the multis I've been at no template rules build everything from scratch um you know, if you're a small underresourced generalist if you're building every model from scratch versus using catalyst that's probably not a great idea and so just framing that heterogeneity as we get into AI I expect and would actually suggest that everyone sort of thinks about using AI in their own specific uh uh specific uh way with a few with a few highle considerations and high level highle principles Right? AI can there are many perils to using AI, right? Defect defect and slop and probably the one of the most um uh concerning ones is comprehension degradation, right? If if I have an analyst who is AI pill and 18 months in they've not really understood anything about the businesses. It's very hard to take that leap uh take that leap to becoming an alpha generating stock stock picker. Right? So AI can do your thinking for you, but it can't do the understanding. It can't it can't really facilitate the comprehension, right? So the promise is acceleration, the promise is rigor, but the promise is really insight, right? So one of the core principles is if you're not using AI tools to develop and deepen insight, you have to sort of reset and rethink how you're using uh the tools.

I think is also an important consideration as we'll try and display in this in this session. You have to understand the jagged edges, right? AI is certainly AI and finance is not a solved problem, right? It's not a solved problem. When people say AI XL is a solved problem, I'm like, you are wrong, right? You are very wrong. Um, we have made insane improvements in the last 6 months, but there are still many jagged edges that are very problematic for the investing process. So what we we're trying to do here is shine a light on those Jacob processes, think about workarounds, but als be be very objective about where these tools uh where where where these tools are with a venture capital funded vendor who needs to sort of sell uh sell that things are great. Sometimes you don't get the balanced uh expression of those jagged of those jagged edges. Um, you know, we'll talk a little about tool tool selection retrieval capabilities is still I don't think multi-document retrieval has gotten a lot better still not a solved issue there's still an ongoing debate on MCP and CLI etc uh and then validation um and this is where I think there's been some interesting breakthroughs certainly in my experimentation so we'll walk through that as well too so four governing principles of all AI deployment in my opinion um are you know above all we have to optimize for rigor over speed we play the most competitive game in the world. Markets punish mediocre research. There's all sorts of checkpoints on that. Uh so it all says we have to demand rigor. We have to build a a broader net for signal identification. The idea sort of the mental idea is like you can have an alt data team in your pocket or investigative journalist in your pocket or a quant NLP forensic accountant in your pocket. Broader net more signals. Um that that that you know with a validation system that can't be a bad thing. Just conceptually that can't be a bad thing. We want to compress the mechanical parts of the parts of the job. As long as with the important coffee, as long as that comprehension doesn't compression doesn't impact comprehension. Um, if we can free up from the 38,000 hours, 600 hours uh to redeploy on high value tasks, that's a that's a nice win case. Um, and then a two-tiered validation system is critical. Sort of systematic AI verification, but then the artisal human verification. So, these are four principles that I that I believe. One prior in that that I point people to is is that sort of the the prior of visible alpha and and um the spreading of consensus. Yes, believe it or not, when I first started on the buy side, I had to go email 12 sellside analysts, collect all the models, spread those spread those hand by hand, talk about mind-numbing work. Each took me one or two hours. Maybe I learned something the first one or two after the 15th spread. Um, it was really annoying. Uh so when visible alpha came out and a software system does that does that and eliminates many hours of that grunt work. I would argue there's there's very little if any alpha signal in that but that frees up maybe that freed up 50 hours in in a year for me as a junior analyst. So those are the sort of like lowhanging fruit uh acceleration and mechanical opportunities that I'm chasing down there. I think there's a long a long list of that. So when it comes comes to modeling right you know it depends a lot on where you're at what firm you're at but you know my lived experience I think anywhere from 10 to 25% of my time was spent on the purely mechanical parts of the modeling I'm not talking about doing due diligence to input input data points I'm talking about you know just entering the data and organizing building the analytics and connecting and validate the numbers and updating for quarters and doing a resegmentation rebuild and hey this model's four quarters out of date or hey there's a large transaction will you model that into the core core core operating operating operating model. So even as experienced analysts you know it's not crazy to think that I'm spending maybe three four 500 a year hours a year out of my time on the purely mechanical. Now I didn't put a slide on this but if any of you have been at a a multi-manager at one firm and go to another firm you will intimately know red pen theory where you you don't get to often most often you don't don't get to take models with you. So rather than spending time on a garden leave with your family or partner or traveling, often you're you're uh starting to think through the process of rebuilding your models from scratch, which if you've had the same models for 10 years, all of a sudden they are locked in a system and you have to rebuild those. Uh that is an incredibly painful experience having gone through that hoping having gone through that myself. Uh so if I could help the next garden leave PM not have to go through that I will have uh I will have paid it uh p paid it forward. So if you've been uh tuning into this seminar series you know that one of my touring tests for investing a AI is if I can cover and ramp on the 157 names that I used to cover five years five years ago. Now I have some historical context but the idea is can I refresh? Can I do full ramps? Um, my sort of test is like, do I feel like I could go and pitch a portfolio of these names to a to a director of research or a bisdev at a at a hedge fund. Um, I'm and I'm not there yet. Um, but a big part of that is can I update the models that are 157 names, right? There's I don't know how many models I have, 80 or 90. They're five years out of date. Um, and there's probably a few dozen fresh rebuilds. So just that modeling exercise alone um you know, to if I were to you know spend 20 hours on each model 157 names I've already burned through 12 months uh 12 months of of uh a pure modeling. This is you know, sort of many reasons why PMs hire junior analysts. This is one of them. Uh the just the simple bottlenecks that exist in this process. Now in reality what I've done at various times is I've sent much of this work to an India uh team to do the the basic inputting of numbers but still that doesn't take it from 20 hours down to two that maybe takes it from 20 down to uh down to eight once I have those models up creating tracking systems for those 157 157 names. So I was a healthcare investor, healthcare X therapeutics. Uh so this is sort of a rough uh rough coverage area. If I can't, I'm not going to go just start building these models from scratch. I want to find a way where I can press a button uh show an example of my model and have this model architecture built for me. I was I was hopeful that maybe I'd get a little bit closer uh in this exercise than I did. So, I'd I'd say um unfortunately uh we're not there yet, at least from my uh experience. And I'm going to walk you through in the remaining 80 slides, 80 slides uh why. So, I think it's possible soon hopefully to, you know, take these 10, you know, 10 new new analyst builds, new model builds, you know, cut the time down by 90%, you do the updates. This is probably where we are at today, closer to today. The resegmentations, we're not at today. the large transactions I don't think we're at at today. Um so I think with the emergence of new tools it seems possible to reduce the intensity of the mechanical parts of the investment process and I'm going to show you a few few things to do. Uh and then again the access time can be redeployed to go deeper expand expand coverage. There are some positive signals right um you rogo um um you know has I think made made good strides in the investment banking process. Uh one of the things that you can do in the rogo system now is you can basically put a ticker press a button and you get a very nice like to me institutional grade accretion dilution analysis. This gives me the structure right and then I just need to go do due diligence on what I think the synergy phase in will be or what the core you know masimo ebida number will be but it sort of takes a lot of the mechanical annoyance of building an accretion dilution analysis. When I had an India team, one of their mandates was any name under coverage that does an overnight acquisition or early morning acquisition, build out an accretion dilution analysis. So before the market opens, I can see what the deal looks like and we can make a decision to buy, sell or hold that hold that hold that stock. This was in the 2015s when healthcare M&A was going crazy. But that was a nice advantage of having an overseas modeling team. Well, I don't need to do that anymore. I could press I could press a button and in in in rogo an architecture like this saves times, reduces errors, and a time crunch frees me up to actually research those key inputs. You know, everything's everything's triage in uh in investing. And when I get the press release at 7 a.m. that there's a large acquisition, the market opens at 9:30 a.m. whether I'm done with my analysis or not, I have a 2 and a half hour window to get it to get as much insight on that transaction as possible. So if I can accelerate those those moments, it's really that's that's really really helpful.

One thing I want to point out in the session too is that AIXL is not just for uh for building uh for building models. I've found 5.4 and 5.5 in particular to be quite Excel fluent. Even simple things like uploading a model into a GPT. Now don't do this before you get a appliance uh uh approval compliance approval. Um, but I could say, you know, help me think through the right way to approach revenue build. So as an ideation partner, Excel is there today. AIXL is there today, right? It gave me some interesting ideas to go to Amazon and break down net service sales, you know, the disclosure between online source, physical stores, the product level disclosure, and then map that back into the the net product and net service sales, which fit on the fit on the source of the P&L. So solving these little puzzles that aren't really that value addited but can give me a a way to think about revenue build. So much of the value in a good buy side model is the revenue decomposition, the revenue build. And I've always been a component of a multi- multi-angle revenue build. If you look at a company, build out the revenue two or three different ways. Build it by product, build it by geography, build it by GDP, you know, GDP growth plus a market market share, right? look at revenue three or four different ways because sometimes you can look at seasonality. You can pick up simple alpha and revenue modeling just by applying three or four different lens lenses to that. So we have a prompt to basically give me all the different ways based on what's reported in the K that I can help design this re this revenue uh build and I've had revenue builds on names like IMAX that have been hundreds of rows uh before and I think actually give you you know helpful insight into how the business uh the business operates.

So, I'd say the general update, uh, if I had to sort of start with the conclusion is that, um, I don't think we're there yet. You know, I I I can't walk into a client, like, hey, all of your Excel can be done with AI now. I don't think we're anywhere close to that. I think any anyone um, who says that is wrong, uh, in my opinion. Um, but we've seen improvements, right? 2025 a era AIXL was was was not good. uh the first version some of the AIXL tools were bad uh certainly in the pre-coding agent era the fact that you a that LLM are now agentic and can tool call coding is a massive innovation that uh has shown huge improvements why because LLMs are natively bad at math prone to hallucination and didn't have great retrieval attention degradation lot you know uh sort of lost in the needle in a hay stack issue multi-document retrieval was very had, you know, finance benchmarks from in the '60s was like is it sort of seemed crazy. U the tools were just like so bad. Uh lately they've shown a big improvement with coding agents which are more deterministic and quantitatively fluent and then the ability to pipe in and do connectors to MCPS is a big is big unlock. I'm still seeing I think uh some issues when I get into deeper use cases. I don't know. My hypothesis is that there's still a little bit of a token bottleneck. Um that the effective context window even even the agentic system spawning sub aents is is still not is still not big enough. Multi-document retrieval can be hundreds of thousands up to a million of tokens consumed. And I asked at Perplexia Computer how big one of my models was and just my Danaher model is 1.18 million tokens, right? And so what I'm seeing as a as a conclusion is when I cut my model down to one tab, one simple road update, it's doing a very good job. When I have my six tabs in there, you know, over a million tokens, it's really getting it's really getting get getting lost. I sort of hypothesize that this is a simply a token context window issue. Uh I had some disagreement on that twi on Twitter that is actually sort of a you know uh uh coding in Excel bypasses that context window problem. Um but that was that's been my uh sort of across use cases. The more complex uh the use cases, the more my skills are ignored uh the more simple parts of the task are ignored. So I chalk that up to the fact that the effective context windows of the foundation models are still in the 500k to 2 million range and we some of this may just be as AI has been a waiting game for the underlying models to models to improve. So I think the word on tool I don't think there's a complete tool yet um but there's some interesting pieces emerging. I've been a fan of the agentic workspaces. Uh codeex which came out sort of updated this week. I find very interesting. Perplexity computer I've been a fan. Claude co-work I think is promising but uh still a little buggy. Uh Claude and Chat GPTXL I didn't see that much difference in they're promising but still fresh and buggy. Uh DOPA I'll walk you through a bit. Rogo, I think, has shown a vision what's possible investment banking and a few of their AIXL vendors I think are that I picked back up um still aren't there yet, but they've improved quite a bit um since uh since fall fall of uh fall of 25 um for sure. I do wonder on some of these tools um how you how you you know they sort of uh overcome the compute sort of subsidization uh dynamic uh but that's a debate for a debate for another day.

One of the things I I like to point out to people is um is um you know the the importance of accuracy and underlying data. One of the things that makes me you know go mad with Claude is if the connectors break Claude will go try to find a number on a on a blog somewhere which is just a really bad idea. Um so I've been become a fan of DUPA. DUPA has these data sheets which are sort of AI generated AI cleaned. They've sort of long before LLMs. uh they're not really in a usable structure, but it's a really valuable source of comprehensive uh comprehensive data. So, it's been one of my favorite MCPS and has been a key unlock in getting getting accuracy. And so, what you can do in these Excel uh cloud XL now is you can upload skills, you can upload connectors and that has been an important important unlock unlock for me. Why is that important? Because if you oneshot this sort of principle cross AI, like maybe we get there where you can oneshot and AI knows your intent and it knows by your history that you're a hedge fund person and you care about this part in models versus not. We're not there yet, right? We're certain we're certainly not yet there yet. We're we're in an era of AI where systems still matter, right? data matters, context matters via workflow orchestration uh agent.mmd skills.mmd uh files. Uh so you see this if you try to oneshot if I try to oneshot I go into a AIXL and say build me a financial model for for for Amazon give me a three-statement model. Um, you know, it uh it's sort of funny like one of the first thing that looks is like does a balance sheet balance? Um, you know, almost in none of my models in the forecast period do I even focus on modeling out every line of the balance sheet to balance balance the bal balance the balance. What what what you what you get out is sort of like I still chalk it up to AI slop like yes, does this does this looks like like something a a four-year-old will build? No. like we're sort of past that formatting structure has gotten a little bit uh better, but it's still not there yet. It's it's still something closer to that you get out of a faxet, you know, a fax set down download. There's no decomposition of the revenue. There's no revenue bill, which is I would sort of argue in most companies the number one most important part of a financial model. um you really no analytics around the the um the the uh the the income statement um you know summarized sort of compressed balance sheet um etc. Uh if I asked chat GBT what it thought, what you sort of see is you get um you know very sort of simple it says top tier undergrad early investment banking analyst level. I wouldn't even give it give it uh give it that unfortunately. Um so systems approach uh as we sit here April 2026 really important you know one shot. Most of the my friends were like hey Brad I tried to use AI on these things like it just wasn't good. I'm like well what do you do? like, "Oh, I went into Claude and asked it to build me an Amazon model." Like, well, yeah, no surprise. Um, a oneshot approach with no validation system, with no steering of the model, no connecting of data is going to be bad. It's just going to be bad. And pretty much every workflow, like I don't think there's that many workflows where that that still can be one shot. uh in my in my experience you have to apply a systems approach and intentional design an engineering mindset the institutional rigor sort of enforced v via institutional data institutional workflow and then validated accuracy right I I'm not going to oneshot a model send it up to my PM and have you know one-third of the numbers be be uh be be uh be incorrect right and so part of that institutional rigor is shifting the thinking of the financial model I remember 18 months ago I was talking about fine-tuning and you know you finance tuned models and we're going to go into the parameters and we're going to make this model think like a hedge fund analyst via the parameters and the parametric memory and I don't hear much about that anymore because skills file skills atm files have become so powerful that I can change the way the model acts and thinks simply by putting workflow context next to model model intelligence. Why is that important? Well, one of the core issues in financial modeling almost many use cases is the underlying quality of the training corpus for finance is awful, right? You know, investing theory on the open web is highly influenced by Buffett and blogs which are often more often than not Buffett derivatives. Um the the financial modeling corpus is highly influenced by invest invest investment banking, right? So if I asked Chad GPT oneshot, I say if you had to pick, you know, one and only one test of whether I build a financial model for investment research accurately, what would it be? You know, sort of clowning the model here. It's like, does the balance sheet balance in every historical and forecast appear without a plug? Again, I have no plug. I've I don't know if I've ever built after being yelled at the first time I did it. I don't know if I've ever built a plug into a three-statement model to support uh an an investment idea. So this is an investment banker. [clears throat] It's not a stock picker thing. And that matters, right? If I'm trying to build 157 models. I don't want 157 investment banking models, right? Okay. So that's that's one shot. That's fresh out of the box. How do we how do we um how do we apply a systems uh a systems thinking? You've seen the slides in the other seminars, right? We want to understand, you know, we want to understand the superpowers and jagged edges of AI because if we see that we like we sort of know, okay, because of the training process, this is why it does it. I can go back in with a skills at MD file and change that. And so that intuition is important. Um, you want to think about intentional tool selection. You know, your data strategy is super important. your workflows, prompts, and skills that are customized to you is an incredibly important part of the frontier right now. As so much of rag and retrieval and model engineering has been abstracted into agentic harnesses, what's left amongst sort of the commonality of all of the sort of current leading agentic workspaces is is sort of two things. It's the data connectors and it's the skills architecture that basically encodes your workflow into the system. And as I sort of think about my hypothesis of where this uh where this game ride is going is those are the two irreducible elements um that um that need to be uh addressed. Um and then a two-tier validation system. How do you systematically but then an analog way check the numbers um and a system that enhances comprehension and and and rigor.

All right. So what what is a skills file? Uh if you haven't understood skills, it's not that complicated to understand. Go to this you YouTube video from Shaw to Lebby in

23 minutes. You'll have a basic understanding. It's simply a way to encode. They can get more complicated when you talk about what's a what's sort of a you know deterministic Python call versus you know how do we chunk skills etc. There's sort of you know an emerging uh developer community around skills engineering that matters certainly when you start to scale it across a system. But it was a very very civic, very specific uh perspective. Uh skills are just a way to encode a process, sort of like a stack of prompts or a menu card of prompts. Which is nice because I had 306 prompts that encoded my entire investment process, which was great and impressive. And people like, "Wow." I'm like, "How often do you use those?" I'm like, "Well, crap, if I have to put in 12 prompts to analyze a new CEO, that's a pretty cumbersome process." That the the prompt chaining and all these complicated exotic prompt engineering techniques that I wasted all this time in 2025 learning, I think are completely abstracted now into skills. I can put those same 12 prompts into one skills file to evaluate a management team. I can just call that with a forward slash in my agentic workspace, put in a ticker, and the user, it's like incredibly, incredibly user-friendly. I can, you know, the sort of new investor dashboard. I'm just going to press a button and I get a 10-page report on a management team. It's sort of fantastic how the abstraction of all the engineering has has happened.

Um, so, you know, skills are really sort of created by pioneer by Anthropic, but really becoming the standard uh that work across uh work across platforms. Prompts are still a thing for sort of simple use cases, but for actually wrapping your process in an agentic approach, skills are the important, important context. The wonderful thing about skills as happened in prompts where where LLMs were great um at building uh prompts. So I never had like I haven't written a prompt by hand for the last six months because LLM capability can write a prompt better. You can get to a very high standard of skills creation simply by by documenting your process in a deep, comprehensive way and putting that document next to it next to uh a well-structured agent and ask them to build a workflow orchestration. And so this is a lot of what I've been going down. Now, I don't think that's an institutional grade standard, but getting 85% of the way there um is um it's a really good uh starting point. And so what I've been doing is, you know, to build this is like, okay, I'm going to go in and effectively, you know, create, I think this is an 88-page, you know, buy-side financial modeling and guidebook that talks about historicals and cleaning numbers and basically everything I could think about about the financial modeling process that I've collected in my brain and in my educational program for 20 years. Let me put that into a textbook and let me, you know, present that to model intelligence and ask it to build skills. And it sort of shot. Sometimes I try things like, "There's no way this is going to work." And this is one of the things I'm, "Oh wow, this is pretty um uh pretty good." So I've shown this people people this workflow and, you know, I sort of say like, "Start thinking about, you know, articulating your workflow, workflow context, create these standalone documents for your own process." For me, it was, you know, the notes and process learnings I've gathered over the year, some voice brain dumps, and iterate, you know, a number of sort of steps of of um of iteration.

What you see when you do at, for example, when I upload my model foundation skill, right? And I ask ChatGPT that same question, "If you had to pick one and only one test of whether I built the financial model or for investment research accurately?" All of a sudden, it gets the right answer, right? Does the model replicate how the company actually makes money? Uses the real operating drivers management uses internally? Check. Nicely done. Why? Because if revenue is wrong, almost everything downstream is fake. Precision. Nice job. Nice job. ChatGPT. A good single test is not, "Does it balance or do the formulas work?" Those matter, certainly they matter, but their hygiene. Highest signal test is whether the model captures a business engine correctly, right? Can every major forecast line be traced back to the three most deterministic key drivers? And the CFO test: If the CFO read the revenue bill, would they say, "Yes, that's how the business actually works," right? And so this is just a demonstration of the power of skills, right? To be able to go in and articulate all of these concepts are skill are concepts that I've articulated in my workflow document that I've presented to model intelligence in the meta-skills creation process, and it accurately captured those back in the FE model foundations uh skill. So I've rectified the pro I rectified the native challenge of of uh financial, you know, the financial model corpus uh that's investment banking, investment banking heavy. So this is, you know, this is sort of what I've been building out. It's the, you know, it's the model foundations. It's things that are always loaded: formatting philosophy, tab architecture, the exact way I want it, raw build from scratch. It's, you know, an model updater. It's an audit model. "Hey, before I show this model to my PM, go and check every number, every formula." Uh, that's working quite well. Uh, the qualitative to quantitative: "Hey, management said this, what does that mean for EPS?" "This company lost this contract. What does that mean for EPS?" I think there's a huge hurdle to uh jump over in terms of automating uh sort of automating ourselves. Um, you know, model earnings prep, scenarios, a key drivers, what are the three things that matter? And so these are the eight master skills for financial modeling. And I'd say again, when I keep the models simple, uh don't overload the context window, they work quite well. When I try and have it do too much, certainly my raw build skill is just not working. Um, and I don't know if it's just too complicated or what what's uh uh what's uh what what what what what's happening, right?

And so in, I'd say I'm sort of seeing whether it's Cloud Co-work, Cursor, or Perplexity Computer, this sort of convergence of this super app strategy where there's one agentic workspace. All I have to do is go connect and go go get my connectors and upload my skills and um, you know, I'm not sort of steering people to use Perplexity Computer. You could take a skill, embed Perplexity Computer, take it over to Cursor. Like, "Here, here are my eight modeling skills. Can you revise these that are in line with how Cursor would like to see the skills?" And that takes 15 minutes to to uh to to uh to do. And so ultimately, what I've been doing is building these model model workspaces uh where uh this is in Perplexity Computer. I have a, you know, workspace where it's a new model creator or model validation or a model updates. I can upload the model. Uh Perplexity Computer doesn't have an Excel app yet, and they only just recently got connectors to Duopa, etc. I can go in and upload a model into model validation, "Go check this model," which is a workflow I quite like. I think is a big unlock. Um, so generally, um, the the great news about the trajectory of AI and finances, you know, prompt library with 300 skills is really hard to use. I think UI and UX will matter a lot. So building these spaces is is going to be really important, certainly to move beyond the hacker era into the broad adoption era of AI and finance. So, I'm starting to think about all these different UIs in a case where um I could have just these different buttons to push, right? So, I could have I could have my D, you know, my Uber model here uh at the top. I could upload my Uber model and I could have a UI where there's an update refresh. I just press the button, it updates it. I could have a research model, you know, drill down into each line item. "Let's go do some, you know, let's go research um uh you know, revenue per ride," etc. Validation and sensitivity. "Let's go validate, you know, the formulas and codes and audit and do a bull-base bear and understand what's baked in the stock with the DCF." You know, challenge and sharpen. "Let's go push back on these estimates." "Hey, you're modeling an acceleration in volumes, but Whimo is expanding into all these markets. Will that lead to a market share?" So, having these debates to help you develop a variant view embedded in the model. And then ultimately, how do you how do you identify differentiation out of the model? So I think this sort of UI game now is interesting with each of these uh each of these steps in a model workspace being an underlying skills architecture. I'm starting to think about for firms like, if you have a firm of a hundred investors, how do you build these in collaboration with the engineering team and then all of us to roll this out to investors, just like, "Here's your modeling UI, right? It just works. Press this button and be surprised and delighted by what you can do with your financial modeling use case." Again, not there, not there now. Not all of these buttons will work now, certainly not reliably. Um, but I'm sort of taking on, you know, making a wager on that we continue this trajectory of improvement. And could this be a 26 uh thing where you have something like this? Could this be a 27 thing? I'm starting to think that maybe that's maybe that's maybe maybe that's true.

All right. A few other um a few other uh considerations, and we're getting into level the the three-level test. Um, you know, one thing that you I'm still seeing this is really annoying is the numbers still aren't correct in a few of these model updates, right? Claude was the worst performer on these, unfortunately. Um, and I asked, you know, "Why the numbers are wrong?" It says, "The reason is simple. I pulled the historical numbers from web search results and uh and secondary sources, news articles, uh data aggregators rather than from the actual 10K filing." If I had an analyst that updated a model pulling from a news report, a news article from Benzinga versus going into the 10K, that would definitely get a bad analyst review. Um, so these are still jagged edges that need to be need to be ironed out, right? And so part of the reason we're not there yet is these tool these things still happen. The connectors break. Claude forgets it's connected to Duopa. Um, so you need to remind it, "Don't, don't, don't do that." Even though in my skill, right, I explicitly call out the sourcing architecture. So data MCP is sort of critical for these in-the-flow quantitative quantitative use cases. It's part of the reason why I think, you know, through much of last year, uh, I would hear a lot of uh firms trying to use AI for earnings season, for morning news monitoring, and for earnings previews. And I would point out the risk embedded in those workflows simply because there's no natural validation loop. If I'm waking up at 6:30 a.m. and I get an AI alert that my company cut guidance by 7%, right? And there's some volume in the pre-market, and I want to pump my position in the pre-market, but that's a hallucination. That's a real problem. That's a real problem. And so I've sort of had those I've had those workflows as red lights in my um in my schema. Um, maybe they're moving into yellow lights uh now where I've had green lights as idea generation, sniff test, where there's sort of a natural validation loop. I make that point to say, when there's no natural validation loop for low-latency, high-impact use cases, you had better be sure that your data data lake and API connectors are really, really reliable, that you have a validation. And so it's particularly critical for earnings news, anything without a natural validation path, uh uh, you know, don't use ChatGPT for that. A much more more comprehensive system is important. Use it for a topical brief web scrape of, you know, uh, you know, learning about the Medicare Advantage world. That's great, but be careful. Just, just, just be careful.

The great news is that when I get the Duopa MCP connector uh and I start to do my level one test of update a model for update a model for the quarter for Amazon, I do the, you know, I I updated it by hand, and this is my sort of checking function. Guess what? All of a sudden, it does it. It does it well. I tried to move out of Excel into Cloud Co-work and Cloud Code. I had so many issues, and I, it must be a skill issue. Um, I read all these, you know, articles and tweets and YouTubes about Claude Code and VS Code and Claude Co-work being great. My experience is from time to time good, plagued by approval fatigue, failed extraction, uh doing hard yaka, having to allow once every time there's a web search. Um, I just don't find it very user-friendly, telling it over and over to pull from the filings, and then we're going to we're going to the website again. Um, and so, you know, rounding rounding numbers issues, um, updating a model and just pulling in numbers that are rounded. Um, you know, no, I don't want that. I want accurate numbers. Like this this matters. We don't want to round numbers. Um, so moving beyond, you know, Cloud XL with the MCP connector, I still had some issues. So I have a I have a stack of modeling evolutions evaluations. Some I'll share, some some I won't. I I'm not so hubristic to think that anyone from any of the foundation labs cares or is even watching this to go and retrain the models based on these evaluations, but just in case I have have and I'm building a broader set of model evals, thinking about some URL sandbox, etc., to continue to do that, which is interesting. Um, but uh for this for this use case, we'll sort of talk about three levels of three levels of of of valuations. Number one is something I can hand to a a liberal arts intern. Um, two would be something maybe an investment banker can handle. Three would be like a banker who knows, you know, knows something about buyside modeling. So, a super simple task, the sort of the prototypical simple task is, you know, and Amazon, if you cover like, you know, the model's pretty simple, not that many adjustments. You get a you get a below-the-line ad back from time to time, but they don't go crazy with the non-GAAP. It's a pretty simple, it's a pretty simple model, about as simple as it can get. Um, just update this one quarter P&L, right? And I and I moved over all the formulas. All I needed for this use case is find where the numbers need to be input. This links down to revenue, you this down to operating income. Just update the numbers, right? And Claude with Duopa, 33 out of 33. Claude without Duopa, it was one out of 33, and I don't know where the numbers were from were from. They're trying to pull from the web, but when I used the connector, uh, great. Perplexity Computer, which this is before the Duopa connector, I ran this analysis, it had the Perplexity numbers also great, 33 out of 33. To have an extra set of check checks, I'll move this into my validation space and Perplexity Computer test this model, test whether this model was updated for the Q2 to 25 model update, uh, uh, correctly. Um, then in this case, this was uh this was one of the Claude examples. It actually caught a few of the differences. This is, I don't know exactly why, but Claude without MCP had small had small small differences. So it goes and does a line-by-line test, and I I didn't do this on a hundred use cases, but on the, you know, 10 use cases where I had the Perplexity validation space check the numbers, it was right. It was sort of accurate in the in in in the checks 10 out of 10 times. So I say in general on level one use cases, um, we're kind of there, right? We're there. Uh, these, you know, simple model updates, simple analyses, um, it's really good. Some really good evolution. I'd say that's an improvement from where we were in the fall of 2020, 2025.

Level two, a little bit more, a little bit more mixed. I took an Uber model that was four quarters out of date, and I said, um, uh, Perplexity couldn't do this. So, I gave this to Claude using the Duopa MCP FD FE model update. And this was like one of those moments where it's like, I was delightfully surprised. I'm like, "I don't think that it'll be able to do this and roll forward the estimates and keep the formatting." And it basically oneshotted this with the with the model update, right? This is helpful, certainly if I'm a generalist or uh a broad specialist. I'm not updating a model every quarter. This was four quarters out of date. Add some rows, you know, roll forward the formatting. This is pretty nice. Like that might be a that might be a 45-minute hour-long exercise to get an Uber model up to date. To be able to press a button and have that done in Excel. I was pretty excited. I was pretty, pretty, pretty, pretty, pretty excited by now. I ran that through the, you know, the validation space, and it did make a few mistakes, right? But, you know, I was encouraged because my model validation space caught it. There was a couple of wrong signs and the net income, I think an NCL NCI input, it missed 200 million interest expense. So a couple more complex below-the-line things, but out of 140, 145 inputs, it went through and said that 140 were right, right? And notably, it gave me the five that were wrong, so I could go back in and, you know, fix those five. So I've sort of like, I I use this slide as like really great news of the progress we've made, simply because agents can now agents directed by skills can go in and validate and check models, right? I had struggled so much intellectually with AIXL, this concept of 70% accuracy on finance benchmarks. It's like, "Well, how do I identify the 30% that's wrong?" And so the answer to that now is like, "Create a validation workflow, create a validation skill." Well, before you take any model that is sort of accurate, go through and actually run it through a systematic validation. One step I didn't show in this webinar, but I've done elsewhere, is bring it back to the thesis. My thesis on Uber hinges on these three points, right? To to help me have further understanding of the accuracy of the model. "Give me a checklist of data points that I can go in and check in the in the in the model to ensure accuracy." So that's another way, sort of that, "Give me a checklist of things to go in the in the model." I could hand that score, I could hand that checklist to an India team to go do that by hand checklist. Even if I continue to use India and train India on this, I could take, you know, a 20-hour process down to two hours with this with this system, so I can 10x productivity from an Indian Indian modeling team. Um, so there's all sorts of different workflow workflow uh dynamics. I think the other thing you're starting to see evolve too is click-through visibility. Part part what I really like about the Duopa structure, um, is be able to click into a tab and see actually where that's uh see where that's from. So I think that will be the institutional standard is click-through vis click-through visibility.

All right, so I'm riding pretty high on the Uber on the Uber accuracy. So I'm like, "All right, let me give it something that's harder, right?" And a harder one would be Danaher. I hadn't touched Danaher since Q1 of 2021. Uh, this was really built around at the time there was an acquisition of the Cytiva business from uh GE, and you had this sort of really big COVID tailwind in the business that I was trying to sort of figure out and decompose so I could see what the core was versus temporary, etc. Do I short this thing on a temporary tailwind thesis, etc.? Um, so, you know, I had sort of the the model architecture built around those two pieces. So part of updating a model is like, "Well, I don't want to build the model around that architecture anymore. I need actually a resegmentation. I need a restructuring of the business." And if you have covered Danaher, you know that there's kind of a lot going on. So I had a model that's five years out of date. That's hard, even if the the model structure hasn't changed, but there's a different modeling emphasis. It was Cytiva COVID testing and Cytiva sort of, you know, M&A coming in, the 2026 emphasis more about a possible bioprocessing acceleration and Masimo integration. So I need to just do some rejiggering of my model, restructuring of the model. I want to sort of model in the Masimo integration and see how that impacts revenue and EPS and actually put that one-page accretion dilution into my core operating operating >> [snorts] >> uh model. So that's a that's a big exercise, right? Right. So when people say AI, you know, "Yeah, uh, financial modeling for for uh Wall Street is solved," I was like, "They're really even close like in the neighborhood of um?" I guarantee you I have the, you know, one of the deepest skills architectures and context architectures of financial modeling flows. Um, and we're not even in the neighborhood of getting close to that um right right now. The other challenge, um, the other challenge is Danaher changes is it's restructuring, right? This sort of a a maddening thing for those who are building deep financial models. You go from three segments to four, four segments to three segments. I basically need to recast my entire financial model, and I don't want to lose the 10 years of history. So I add rows below below it. Sometimes I'll try and reconcile if you know old to new. I want to see continuity of operating results over time because his, you know, history lessons are helpful in better forecasting the uh the the future. So there's been some re real restructuring. I got a little excited for a minute here when I updated you that my into my FDA model skill, Perplexity Computer's, you know, identified accurately, "This model needs a complete structural expansion." I'm like, "Well, check, yeah, it does." Um, it pulled up in 2023 that there was a, you know, they had spun off the Veralto uh business. I'm like, "Hey, listen, you're doing pretty you're doing a pretty good job. Like, keep keep going." It had pulled in, you know, where the data uh was to come from, which was part of my FD modeling skill. Like, "Yeah, going to the Ks and Qs, going to the press releases, like uh going to an EPS reconciliation table, like, yeah, let's go." So, promising, you know, I waited five or 10 minutes, and this is what I get out, right? There was sort of no formatting, roll forward. Um, and sort of the core heart of the business uh that I wanted to see, you know, the revenues, profits of the segments, it just completely skipped. And so I put in the P&L um poorly, and that's about that's that's that that's about it. I didn't even consider the restructuring, even though it sort of, you know, it caught that the model needs a complete restructuring. I wasn't able to add rows and rebuild the model around the way I the way I wanted to to to do it. I didn't even consider that as a judgment call, even though I uploaded my Danaher thesis into the context window. And my conclusion is just like, there was just too much, like 1.8 million tokens and all the other documents, and model intelligence isn't intelligent enough to do this yet, was my was my uh was was my conclusion.

All right. So I could give up there, or I could try to think about uh different areas. One of the areas that I started doing is like, "Okay, like how can I complete this task um in a way um in a way uh to have a human in the loop?" So if it's, I'm not going to spend 20 hours on it, it can't be done in an hour. Maybe I can do this in six hours, right? With some human-level assistance. So the pivot from that experience was, "Help me think about AI Excel as a as a partner, right?" Part of that is a step of chunking, chunking, chunking down the the model build, right? What are the steps I want to do? What are the steps I want to do um in building a model, right? We'll go through this in our forthcoming, you know, financial modeling course where we sort of put all of this into skills architectures. We do this by hand, and then we sort of do the chunking, chunking system. Um, but these are sort of like the 19 steps, right? I my my brain is either blessed or cursed with this sort of like decomposition thinking um to really kind of chunk things down into individual, individual uh steps. Uh, which is helpful. I'm going to assign this to an intern like, "Hey, these are the individual, these are the individual uh pieces." So that's sort of where I'm at right now on AIXL is it's a human in the human in the loop. That some of this can be done with uh modeling process uh sort of workflow architecture via skills. Um, but there's still a lot that needs to be done by the human uh for for for the time being. So what does that look like? It looks like, you know, I have to go in and add the rows, add the new columns, add the new segments, right? What I what I did in this instance is I went through and manually restructured, manually restructured all the segments, right? I went in, added rows, you know, put my old rows below, and I did the input for the first year and for for first quarter, right? So after this, I'm like, "Well, okay, that took me some time, but can I say after that?" I said, "The highlighted range is a is a structure that's reported in Danaher's press releases. Please go and pull this information from the highlighted range, update these numbers below." Guess what? It did a really good job at that, right? Once I sort of steered it and created the structure, I lowered my expectations for what AI could do, and it actually did a great, actually did a great job. This was again, Claude with MCP. It even did a situation where in um in Q1 2025, Danaher hadn't yet reported a segment breakdown of other operating profit, but there was a footnote in that quarter talking about a $15 million impairment related to a facility facility in the biotech segment. So, it took the liberty of inputting that number into the other operating profit. That's a pretty good job. Like, you know, clap clap clap it up for clap clap clap it up for Claude. So, it caught caught a footnote adjustment um in a situation that hadn't been reported on the face of the P&L. I'm like, "Okay, cool. Like that. I like that. That was pretty that was pretty um that was pretty uh pretty um uh pretty pretty pretty impressive." Um, so, you know, I I found it helpful to sort of think through the resegmentation. I went and built this on my own, but just a simple act of asking the questions of like, "Hey, when did Danaher change their reporting structure?" All right, it started in Q323. Thanks. That saves me three or four minutes from having to go into the filings, check to see check to see where where it's happened. All right, so the three segment structure exists today, only goes back to Q3 2023. So having this thing as like a, you know, a chatbot in the Excel structure where I can ask questions, you know, it's helpful just to highlight, "This is when the new reporting structure happened. This is when the spin happened. This is when the Cytiva acquisition. This is when the dental spin happened." I found that to be helpful. I found that to be like a simple, simple help helpful um help helpful struct structure, right? I asked this structure of the new reporting of Danaher to go back in history, and it said, "Hey, you know, before Q4 2024, can it be filled?" Because they didn't report it, right? We can either leave it blank, allocate the DA ratably, allocate the segment share and revenue per quarter. I don't really want to do either of those, right? But I think I said, "Go in and put NAs your NAs and all of these all all of these cells, right?" That's helpful. That's helpful as a PM if I'm looking in analyst models, like, "Why did they report biotech adjusted operating profit in Q424 but not before?" I could ask that question, like, "Oh, they didn't report that before. This was a new reporting uh reporting change." And so the ability to have a chatbot in Excel, I found quite helpful for the for this for this process. Even the process where I had to go in and build it on my own because the capability wasn't wasn't there. It reported, "Hey, it's a disclosure rule change, not Danaher or hide anything. ASU 20202307, which was a I guess a FASB rule improvements to reportable segment disclosures issued in 2023." Like, "Oh, cool. I thought that was pretty cool, right? To be able to go in and just see why the reporting has changed." So, I'd say I'd say it gave it a pretty pretty good there, right? You know, I said it can take this 20-hour process down to two, you know, an hour, but maybe could take it down to six hours. So, I I took that I took that first step of, you know, this process of build the first step for them, almost in a way of like, I had an intern. I say, "Update this then her model with no just grunting at the intern with no specific guidance." Like the output probably wouldn't be good. But if I was going to give this to an intern like, "Hey, I updated the first row for you, first column for you, go back and fill in the last three years with that," I'm probably going to get a better output, right? And so, um, um, you I always sort of scoffed at the humanization of of AI, you know, chatbots, word calculators, etc. But like agents, I kind of like bringing that back a little bit of like, "Okay, like I updated the first one for you, right? Take this process and go and and go and update update update those update those back." And so it did a did a pretty good did a pretty good job job of that, right? The ability to go in uh populate the populate the numbers were were great. I continued have to have some, you know, struggles even through this process though with Claude and MCP. Um, I asked Claude to go in and update five years of five years of balance sheets, and it said, "Given the volume, I'd suggest breaking this into a fresh conversation. Just focus on the 2025 to 2022 to 2025 um uh fill." I didn't really think that, you know, this is that much of a that that big of a lift. Certainly not relative to press a button, build a model from scratch. Um, but Claude had some issues here. The connection timed out. Um, you know, the chat sort of ended, and maybe I should have deleted this. This is just my this is my frustration. Frustration with uh with working with Claude, you know, the capital come on or the HUD, just do it. Um, so not probably not my one of my better moments, but uh sorry Claude for being a little bit rude to you. Still still not perfect. Um, I said this balance sheet populate is substantially larger than typical. Like, "What are you talking about?" I just, five years of history is for a pretty simple balance sheet. Um, for some for some reason, the honest assessment is like quite triggering to me. Like I want to give you an honest assessment. Like five years of balance sheet updates is just substantially larger than typical. Like, "What is typical?" Like, "This is exactly typical." Um, so I I have a mixed relationship with Claude right now. I wasn't too happy about this simple use case. Certainly if it can't do this, you know, it it reliably, a simple 5-year model update. Um, you know, the full model restructuring, you know, it's it's it's not it's not there yet.

To try and do this, I experimented with a bunch of different tools. Uh, I up downloaded Duopa Scott. I'm not I'm not sponsored by Duopa or any uh vendors. No financial relationships. Um, um, and um, uh, sort of on the belief in the prior that the underlying uh quality is good. So then I said, "Okay, please update my balance sheet and cash flow statement here, all the quarters and the years up to full 2028." And that, you know, complete. I didn't show an example of it, but it completed it sort of perfectly and added click-click-through visibility to that, right? And so another example of a, you know, helpful co-pilot. Boston Scientific, I'm updating the model there. Had a resegmentation, similar similar thing. So just asking the questions in Excel, I found really um really helpful. Why not use FactSet, Cap IQ to do this? If I go back to the five use cases, you know, we talked about in a model, this sort of output is not going to reach institutional standard. For generalists, I get the feedback a lot. "Hey, I just use Canalyst models. They've gotten much better. The structure is more like a structure you'd see from one of my models. They have strong general adoption." For general covering a broad space, it's a strong solution in my opinion, and particularly if they can figure Alpha can figure out how to integrate the uh Canalyst modeling architecture with the agentic workspace, the agentic capabilities. You can see a vision where you're tying in the AlphaSense expert network calls into the key drivers and catalyst model, giving you a variance versus consensus. That could be pretty slick, right? That could be pretty slick. So, I'm I'm more positive on AlphaSense than I think most um um uh just given the pieces they have and I think what what what what they can do. You know, having sort of hit a hit a wall on a five-year balance sheet update, I had very low expectations for the investment banker with buyside experience, which is basically saying, "Hey, here's my Uber model. Go build me a model from like this from scratch." It was nowhere close. I didn't even uh do do an example, right? So, that's sort of where on the level three, I'd say mixed feedback, right? Not great. Certainly not. You know, if you're at a hedge fund saying, "We want to hand all all of our financial modeling over to Excel to AI," I would say, "Please don't do that. That's a very bad idea." Um, there's just still a lot of jagged edges here. Part of it, I think, is just the fundamental intelligence isn't there yet. Um, part of it is, you know, the context window, part of that is context windows. There's just too uh it's still a too token consumptive of a task to do with institutional grade. Um, as a human-in-the-loop structure, downloading Claude or GPT or Duopa, um, or, you know, Shortcut or Index as an app, having that as a sidecar, asking questions, updating certain elements of the model while you work in other pieces of the model, it's pretty slick. I think we are I think we are we are there for sure. After after doing this a few times and be able to ask questions, it would be kind of hard to go back, right? It's hard to go back if you can if you can oneshot five years of history with the Duopa Scout. It's pretty painful to go back and do it do it uh uh do it do it by h do it by hand. I think you can do it faster, and I think you can do it more, you know, sort of just as just as accurately.

Last point I'll make is really two parts of a of a financial model. There's the there's the historicals, which we've mostly talked about today. Those are table stakes, but like really important table stakes, like having accurate, clean uh financials, normalizing for one-time items, segment KPI, unit economic views, GAAP to adjusted. That's it's very, very important because that's the foundation of what your forecast situation, the business momentum, business quality sits. So very important. Also like table stakes, like alpha doesn't necessarily come from just the historicals alone. You'll get insights around the business. More of the alpha comes from a differentiated forecast, right? The three D three driver discipline of like, "What are the three drivers that really matter here?" And so much of the parade optimization approach, this business is like, "How can I run at these three drivers with incredible rigor, develop a differentiated view on those three three drivers?" Right? So we show this framework of what's possible. I have a stock, I'm going to break that into three key drivers. Um, I can develop seamless, augmented, automated key driver differentiation work. I can overlay an expectations engine, which is one I think one of the holy grails of quantamental investing. I set up pattern recognitions and base rate probabilities. All of my models, I have KD key drivers and sensitivities. I can do the thematic betas and narrative betas, etc. Implied revisions, you know, proactive alerting system, customized risk. "Hey, looks like from this analysis, from what you know, Pier said at the London conference, and this data, and one of our channel checks, that the key driver and market is rolling over, you know, flashing flashing yellow light. Check on that. Maybe we want to punt the idea." This is sort of like, yeah, I wouldn't necessarily the holy grail. Um, but what we're working towards, what I'm working towards, and trying to make a reality in this in this in in this space, right? So, a company like HCA, how do you code in three key drivers? What's the sort of typical work stack you would do on those key drivers manually, right? And how do you scale that? Like fundamental investing has never scaled. I've never feel like I had true, like generally was like beyond idea 12, beyond idea 15. Like I always felt like my 16th best idea was like really bad. Um, part of the reason I did, you know, felt like I was a better Tiger Cub investor than multi-investor. Um, because a lot of this work just takes a long time, right? Like there's no, there has been really no shortcutting um all of this work that is required to have differentiated insights on a broad set of companies. Maybe that's changing, and I think one of the interesting workflows, one of the the abilities we can have is to scale fundamental investing more, to create an automated Excel structure where for each idea under coverage, 157 names, I can go in, human in the loop process, identify the key drivers of all names, right? Yes, right? And any name, this is adaptive over time. These three drivers today may not be the three drivers in six months. Market's a Bayesian adaptive approach, but you maybe monthly, I can go in and sort of revalidate, revalidate and configure my system. These are the three drivers that will matter. I can weight those into, hey, 50% here, you know, 10% here, you know, 40% here, etc. So I can have a re-weighting algorithm. Then I can create agentic processes to go in and evaluate all these key drivers with external, internal, alternative data, you know, management notes, sell-side research, web scraping, etc. I can flow those conclusions back into the financial model through the financial model architecture with my operating leverage assumptions and unit economics and P&L flow. Um, you know, connectivity of three statements and the output of that model, right? Which is sort of the the the manual output today is an investment view, right? My my estimates relative to consensus, the sort of adaptive impact on my risk-reward, uh, the business momentum considerations, is business getting better or worse, right? The model starts to emit an investment view, not just not just numbers, right? This has like been sort of sci science fiction and still science fiction, but like you I think we're getting closer. Some of the pieces are there to to start to do that. You know, some of this comes from sort of core frameworks. One of the core frameworks you we teach in Analyst Academy is incrementalism, right? Trying to always identify the incremental change. You lose a contract, win a contract, revenue, how does that flow through? What's the incremental, what's the EPS impact? You know, how does that change numbers? What should the stock do based on that number? So part of this is building an expectation engine. What's the expectation on this data point? What do we think's going to happen? Is that already in the stock? How does it change numbers? How should the price move off of that? Right? Automated training. You know, there's opportunity for automated trading associated with this, too, where it's like, "Hey, something happens, the stock's off six, it should be up three." Like that's a big alpha gap. Certainly in earnings, you see a lot of those moves where stocks gap down three, finish the day up three. And those, you know, four or five-point alpha opportunities are really a function of compre speed, speed of comprehension, and speed of speed of behavior, right? U and so a lot of it is just simple signal versus noise and incrementalism that that this agentic system can help you to be sit on your shoulder, developing that wider research funnel, understanding attribution, the signal, you know, the stock price linkage to KPI, the exposure materiality, and seeking the inflection. This has been a deeply artisanal process for the history of investment investment research, and I think we're we're at an interesting threshold where that may be uh changing. So I'm starting to build some of the architecture around this, some of the skills um, and putting this into actually the workflow or orchestration to say, "Okay, where is this going to impact the model?" Putting that incrementalism framework into a skill. Um, "Okay, take this news and try and quantify it into an EPS number. Is this signal? Is this noise? You know, is is does our thesis maintain integrity?" Um, etc. So this really feels like the future state, which is why I'm running after it so aggressively.

Uh, so what's uh what what's next? We'll have one more of these free webinars next uh next uh Thursday to sort of cap uh cap this exercise. Process for me, a lot of this like, you know, webinar series has been just getting my thoughts on paper. Uh, we have a uh our our podcast "Invest with AI" coming soon. We put a date on AI Accelerator. So we will go more deeply into this uh all of this stuff June 8th. If you take this webinar series is what's possible uh to do, Accelerator will be how to do it. We'll sort of get into these individual workflow labs month by month, month by month, as really the core core part of it. And we're taking a lot of this IP, workflow, context, training, and implementation into enterprise partnerships as as well too. Uh, so that's really the core of what we're we're uh going to be doing. This foundational seminar, which is sort of a three-hour on-ramp to deploying agents and the investment process, and then the workflow labs where we'll take these individual pieces like, "Okay, this month, let's go into, you know, short ideas and how you can augment a short short sale process, short identification process with with agents." It's really interesting stuff, at least I find it very, very, very, very interesting. And we're in development. We slow this down a little bit just because this piece I think still needs a little bit of seasoning. U but a core part of what we'll be doing at Fundamental Edge is the, you know, is the is the multiple pillars. We want to start with the foundations, build a model from scratch, how to use and think about a model. And this is sort of the old way, right? But we want to start bringing all of our program to sort of accelerate thoughtfully with with AI. I believe maybe in old school, the foundations are critical. You don't want to you don't want to jump right to pillar four with anything because you don't have context and and etc. And part of like, you know, reason I struggle with coding agents.

I couldn't debug, and I couldn't tell what's good. The reason I can make these Excel agents work is if they break with the Danaher model. I need, I know how to go in, take over control, and rebuild the model in a way that I can sort of get through that, get through that gap. So, you know, my old-school mindset is the foundations of how to do this job are probably going to be more important than ever. Certainly, the ability to articulate it and have a process around it will be more important than ever because if you don't have a process, there's really nothing to accelerate, nothing to augment. But if you can take a great process that took you 20 hours in the past and take you six hours in the future, all of a sudden that starts to get pretty interesting.

All right. So, those were all 97 slides I had. Uh, we will take some Q&A now. And, uh, let's see what we got. Record. Will there be recording? Yes. Will there be recording? Yes.

Um, uh, say you cover oil and oil prices as a driver of each company's model. Is that a good use case of AI? Yeah, I need to think through this. I need to think through how the actual technology works. Um, but the way I've seen industrial models built in the past is like historical correlations of all the macro KPIs sort of embedded in a financial model. So if I'm covering an airline, right, I want to see all the operating statistics, RASM and chasm, etc. But I may want to see, you know, TSA volumes in there. I want to see any other data that's publicly reported. Um, and, um, maybe in the forecast period, I want to have an explicit, you know, cost per seat mile tied to an oil price for forecast. Um, this brings me way back because I was an investment banker for Delta back in the day. Uh, way back in the day. Um, so you'd want to have like different scenarios, but in theory, you could have an agentic update of that data series in your financial model. So every time you open your model, now things would move around a lot, but maybe you could press a button to give you an updated number to see, okay, based on this oil price, what happens to my number? Now, I don't know if there's alpha in that necessarily. That's maybe more beta because I think that stocks that have that sort of like obvious input get coded pretty quickly and they tend to be fairly efficiently traded visa v macro input.

All right. I have someone who was 1:30 a.m. from Europe. So I'll take his question. How do you see this with private equity sourcing to investment committee? They don't need so much of modeling, right? What a software which covers this proper industry-specific IC memo before making a bid? Will PE buy this? Um, I don't know what private equity will do. I think the interesting part about agents is agents are a flexible substrate. Um, so I think what we're focusing mostly on is like helping people understand that substrate, helping people understand sort of adopt an engineering mindset around their own investment process, mapping their own workflows. Um, and the reality is like if we work with a hundred people, like they're all going to have 100 different ways they want to do that, just as 100 different investors will have 100 different Bloomberg launchpads. Um, so, um, that's the promise of agents is it's not there's one right way to do things. Um, you might scoff at the way I build my models, and I might scoff at the way you build models, and that's great. Um, that's part of how this industry operates. Um, so, um, I think each firm will have to think about how they adopt this in their own way.

How much of the balance sheet do I model? I don't have a model here with me, but typically I model every input historically. So I can run the analytics, your cash conversion cycle, net debt. It really depends on the type of company I'm looking at. If working capital is a big, I generally will model three pieces on a decomposed piece. I'll model the working capital cycle out in a cash conversion approach. I'll model out the net debt to EBITDA and I'll model out the capex cycle. Um, and I think there's sort of a prey to optimization approach in public companies because when you're covering 15-50 companies to model out, to take eight hours to model out the balance sheet of a net cash business is not a great use. So, if I'm looking at something like Amazon, I probably want to spend a lot of time modeling out the capex cycle. I want to see like how many dollars are going in relative to revenue, when that could peak, etc. But maybe I'm not that worried about the net debt, right? Because it's like there's not, you know, Amazon's balance sheet is healthy. Um, so I think every model looks a little bit different. You're adding customizations to try and develop insights on the key drivers that matter.

All right. Um, how do we get the fundamental edge modeling guide? Uh, you don't. Um, these are sort of my proprietary IP that I'm putting together. I'm putting these into skills architectures. Um, these are things I'm not going to release broadly. I think we're working on a handful of more enterprise-grade relationships where we'll come in and do the full enablement and bring those guides and bring the skills architecture. And, um, you know, let's, I think everything's changing a lot. My core business trains humans. Um, so we're trying to figure out, you know, what the future looks like for us as well too. How will those humans be trained differently? Um, can we start to train computers more? So, we're thinking through that as well too. And part of that is being thoughtful about sharing our IP too broadly.

From Antonio, hello Antonio, nice to see you. You mentioned connecting to the MCP a lot. Have you tried simply providing all the full PDF filings and asking the model to fill in historical data? Does that work? If not, would you recommend any MCP or is the DUPR the one you would go to or others to use? I don't, I'm just going to give you a little bit of secondhand what I've heard. Um, you know, PDFs, you multi-document retrieval, even simple things like tables have been an issue in the past. You've heard a lot about like OCR, optical character resolution. You hear about parsing, that's a little bit above my pay grade in terms of understanding anything beyond what I hear. Um, you know, I hear that MCP is sort of a workaround to that, that workaround to OCR, workaround to PDF parsing. Um, there also is this debate which, like I'm not, I also can't really opine on, of like, is MCP2 brittle? Do we CLI and APIs, etc.? Um, so I'll continue to observe that debate, but I don't really have a great viewpoint on that. There are some other MCPs emerging. Um, it's financial modeling. Is it financial modeling data? Just ask, ask like ChatGPT with the MCPs. Finhat is another one. Um, I think they changed their name. So, there's a few other like relatively low-cost MCPs. FactSet has an MCP. I'm a FactSet user. That was a $10,000 add-on to use their MCP. So I did not pay that extra money. Um, so I think there'll be more MCPs out there.

From Gary. Can you give more color on the different AI, ChatGPT versus Claude versus Perplexity versus Gemini? So this is my, not everyone agrees with me. I'll give you my hypothesis on where this is all going. When Sam Altman talks about the super app strategy, that to me is like that resonates with me. Um, you know, the, um, I think this is all going into the agentic workspace where you can very easily get your connectors set up and your skills uploaded. Cursor was a delightful experience to get that set up. Claude was a really painful experience to get that set up in Claude co-work. I set all this up in Cursor yesterday in like under an hour. Um, so I think that's what you're seeing. You're going to have these agentic workspaces that have the tool, that have the hands of a coding agent. They have web scraping. They can update documents on your computer if you want them to, or you can turn that off if you don't. Um, I sort of wonder if this is going to be single model or multimodal. Um, you know, Perplexity gets some hate amongst finance use cases for the years of spamming that Perplexity is putting Bloomberg out of business. Um, you know, Perplexity Computer has been a delightful experience. I think it's the most user-friendly. It's the most consistent tool. Like Claude will be very nerfed. Like Claude will be great one day and not so great the other day. That's sort of a challenge, particularly, you know, if you have this great system, you put all this work into it, you upload your skills and your data on Claude Enterprise Claude, and then you have earnings season, and it's nerfed on earnings season. That's kind of like a problem, right? So, this nerfing issue, this throttling issue, I think is a real, I think is a real not great if I run out of tokens and I can't upgrade. Like some of that gets handled on an API, like enterprise-grade API basis, but there's a debate on whether the harness API, blah, blah, blah. So, um, so that's, I'm still sort of in the Perplexity Computer camp. Um, my, like, if you had to push me, I think the winners are going to be the vertical players here, just as we're seeing. Harvey has been a little bit of a preview in the legal environment. You know, people like, why would you use Harvey when you use Claude? Well, it's like, actually, when you get into regulated industry data, enterprise-grade security, usability, forward deploy, forward deployed engineers, like all of that focus, right? A simple thing is like, you know, ChatGPT is hiring investment bankers, but not hedge fund people, right? So, some of this is just speed, some of this is focus, some of this is the specific build-around. And, um, so I think that, I don't know if it's, it may be alpha, it may be alpha sense, it may be someone like Rogo. Like I think that one of these vertical players will probably be, you know, listen, like even in the terminal space, you have Cap IQ and FactSet and Bloomberg and etc. ETA. So this is going to be, I think this is going to be a massive market. There'll be room for many, many players. But if you had to press me, that's my belief. I believe that just like Harvey has now built a very big business in legal AI wrapper, I think that there'll be a very big business built in finance AI wrapper. I didn't believe that in, for the record, September 2025. I believe that now.

Have you tried Shortcut AI? One of the only Excel integrated LLMs that have processed complex toggles and formulas, but also very good at following natural language instructions. I try not to. I, yeah, I try not to give too much feedback on like the individual because sometimes when I say bad, when I get bad feedback, then it like overwhelms my inbox and the vendors want to. So I did not like Shortcut when it first came out. I tried it recently, like it has improved a lot. And so I think Shortcut now is something to pay attention to. Um, and, um, and it did a good job in all of these tests. I just didn't, I don't want to get to the point where I'm showing six vendors and one gets an F, and then the F is always mad at me. And, um, so, um, yeah, I think Shortcut is one to keep an eye on for sure. Um, they've definitely been, I think focusing and iterating pretty quickly. Some of this again, I think comes down to this focus area where if you're just going to do Excel for everyone, you're going to miss the nuances. So, some of it is how do you pick a lane or have a flexible substrate where you're pulling in these skills easily.

Alex Castro, have you seen any use cases for industry supply demand dynamics outside of just rhetoric? For example, aggregating hyperscaler capex plans or utility capex spending since that can later arrive? That's a really good point. Um, I spent so much, one of my other, you just gave me an idea, Alex, so thank you for that. One of my jobs when I was a junior analyst on a consumer team, a seven-person consumer team, was to aggregate the comp sheet, same sort of sales sheet, across like 120 global consumer businesses. Go in, go in. Now, they had different reporting, some reported half-yearly, some had year-ends, you had to quarterize that. Oh, it's such a pain in the butt. You gave me an idea, Alex, to be able to do that now would be really great. Um, across consumer, to have just an updated same-source sales tracker, industry P&Ls. Like in the Analyst Academy, we walk people through building an industry P&L, and you see all sorts of dynamics that on an individual basis you don't necessarily see, but you see it in aggregate. You see that certain industries go through peaks and troughs, predictable peaks and troughs in margin and ROIC. So the ability to do that, not only do that, but update that on an industry basis is really interesting. And the ability to aggregate hyperscaler capex plans, put that into an industry model, show that over time, update that, that feels like you probably have to again, sort of design that a little bit, but then once you have it designed, having it updated every quarter. Yeah, it's a good idea. Thank you for the idea, Alex.

See, have you started with the Q123 model and tried to update with Q223 and then repeated that as an eval that could recur to get your current? Yeah, just do one quarter at a time. Yeah, that would be one way. That'd be one way to do it.

How do you recommend codifying the investment process as a junior analyst such that when you encode it into skills, it can interpret the process and help extract insight? And how do you think we should iterate on this process as the models improve? This is my advice for junior analysts. Don't try and do too much with AI, especially if your PM and supervising analyst are truly there to support and develop you. I think the more severe risk is trying to do too much too early. Um, you know, find ways to go deep. I think the obvious use case for junior analysts is to use AI as a deep dive tool, sort of like almost even more the chatbot use case, less the agent, more the chatbot use case of deeper understanding. And again, there's sort of three ways to read a book. One, read it the normal way. Two, cheat and read the cliff notes. And three, read it, but debate it with a really smart friend, right? Or a professor, right? And what you're trying to do is turn AI into that professor, right? And so that at the end of that experience, you have a much deeper comprehension. And so I would, my advice to you is focus less on codifying and augmenting your process as a junior and focus more on really understanding the businesses, understanding the craft deeply. Because if you start to codify and augment too early in your investment process, you really risk a critical, you really risk bypassing the lived experience, the muscle memory of doing this job. One of the things I've started to develop too is our research playbooks for a junior analyst. And so if I'm a PM, let the PM do more of the agentic work. And on certain ideas, I've given like a five-page checklist like, hey, go read the 10K, check these calls, talk to these three customer checks, etc., to give some more guardrails around what a junior analyst can do. So there's more structure in the junior analyst process. Um, so I think that's, I think there's a rethink coming on how junior analysts are deployed in the investment process. If we can, if as a PM I can update all of my cash balance sheets and cash flow statements, that creates more capacity for a junior analyst to go do more value-added things, but it's less on the junior analyst to figure those out in my opinion, and more on me to deploy that resource effectively in my opinion. That won't be everyone's experience, but experience.

Routing, hello Routing, nice to see you, analyst academy graduate, not sick of me yet. In your future state example of skilled MD translating qual, what aspects do you perceive the LMS need to evolve to make this a reality? I don't know. Part of the reason I'm excited to start this podcast is just to ask people this question who are smarter than me. Um, it's, I think some of it is the engineer, like the harness agent harness, you know, building sub-agents, some of these architectural workarounds. There's a lot of things. Even when we did our version 1.0 of Analyst Academy AI Academy in fall, we were like, these things are going to take years to figure out, and it's like they took months to figure out. So I'm kind of like, my prior now is like these things get figured out. Um, I'm trying to be a little bit more optimistic than pessimistic. But, um, yeah, really what I'm, the real answer is I'm just going to go with the flow and check these things from time to time and follow people on Twitter who are doing YouTube who are doing similar experiments and just get the flow from the community. Like I get a lot of DMs of people showing me things that work, and I'll check them out. Damn, thanks for the tip. So that's how I'm trying to stay current on all of this and then hopefully share it with you as well too.

Have you used the enterprise version of Claude that some financial institutions have used? Heard through the grapevine it can produce a proper three-statement model that's useful for the buy-side, at least for the PE side, but have never used it. I have the max version of Claude, which I don't think was different than financial institutions, but, um, yeah, if you hear so, I have not heard that, Alex, but if you hear something different, that would be, I'd be curious. I hadn't heard that.

Trying to understand the value proposition of your new course, Modeling Academy. How would it help students who have attended your Analyst Academy and AI Academy? Analyst Academy, we have a three-hour module on modeling. I think the Modeling Academy is closer to 20 hours with, kind of 12 hours of just a real, a much more detailed decomposition of the modeling process, if that's your thing. Um, but then really an 8-hour, much more in-depth model chunking. Really, the output of that Modeling Academy isn't necessarily to teach you to build models from scratch, but to have all the grounding so you can accelerate your modeling process with AI. Again, I don't know exactly the timing on that because I think there is some improvement curve that's needed in intelligence to really, I don't want to do a course that's just science fiction. I want to have something that can be done.

Antonio Forado, do you have a PM course coming out? Yeah, I think we're thinking about that in the fall. So stay tuned on that. In all of the courses, I think we're thinking about the three or four pillars where it's like these are the core ways to do it, and these are the AI elements to do it. So there's, I think as many fascinating agentic use cases that exist on the analyst side, I think there'll be even more interesting things on the risk and PM side. Certainly for orchestrating team research and building a coherent operating system as a PM. I think it'll make the PM job more effective, interesting in a way.

Anonymous, are you sure you should not start running client capital again, given your insights is training others your highest and best use case? Thanks for the question, anonymous attendee. I quite enjoy what I'm doing now. Um, I do think that, you know, I get the bug from time to time, and I think there's, I think I don't know if I would go back to like the typical pod PM approach, but as like an AI Director of Research, to sort of bring in some of these use cases to help firms actually deploy capital, starting to think about that, doing that on a contract basis. But I think that's where we'll see where the world goes in five years. I'd say what I learned attempting to launch a hedge fund in 2018 is that by far the hardest part is raising the capital. So, that's that. I also quite like my life as a flag football coach and basketball coach and trying to be present with the kids, although they are getting older and they seem to need me less, which is sad, but frees up more time. But right now, this AI thing is just like, it's pretty fascinating.

Do you see a point in the future where stock calls, P&Ls benchmarked against AI decisions, an alpha over AI, so to speak? This assumes we get to a form of AGI. I love that. I love that idea actually. I was talking to someone today about building trading signals with an AI system. So I'm doing a webinar tomorrow for Wall Street Prep where we took all 172 of Warren Buffett's selling decisions and put that into a skills architecture and created a Warren Buffett sell agent. And I think you can do that certainly if you have all the internal data at your fund. You know, one of the first things I would do, there's all sorts of, I go down the quant rabbit hole. We talked about this in our Factor Academy. Like pattern recognition is much more powerful on losers than winners in stocks, right? Quant fundamental intermediate duration quant is very powerful on the downside, more powerful on the downside than the upside. Why is that? Stocks tend to fail for more common reasons. They tend to win for more idiosyncratic reasons. It's just sort of the nature of enterprise. What that means is that you see this common thumbprint of losers over time. I certainly see that in my career. There's like five or six mistakes I just sort of made over and over in my career. So that's great news actually, because I can start to turn that into codified structures. If you have 20 years of history, you know, trading history, you can go in and study your trading history. You can study your big losers. You can put your big losers into patterns. You can put those patterns into an agent. You can have that agent run risk analyses on your portfolio and you can not make those same mistakes twice. I think an interesting framework or benchmark to then use that or build conviction in that is to track how those signals perform. The same way that Steve Mandel used to at Lone Pine, I don't know if they still do this, is they would track saved P&L after they saved after they sold the name. So they sold the name at 80 and it went to 40, right? Often you sort of forget about that, but they would track that and celebrate that and show the P&L saved. And so I think some of that incentives restructuring in businesses is really interesting.

Alex, hedge fund recruiting market is slow right now. I'm not sure if it's war volatility. Yeah, I think P&L hasn't been great. That certainly during tariffs that tends to sort of press it. How do you think is attributable to AI getting much better, single manager in particular? Yeah, I think multis aren't really, I don't think you're going to see like you're not going to come in and see a big multi-manager cut headcount by 20% this year for sure. Again, like these agents can't do anything an analyst does. And I think like Jevons' paradox will most likely hold in these investment use cases. I tend to view like institutional hedge fund seats as secure for at least the next three years, maybe even growing. I think the roles will shift a little bit. The same way that hedge funds brought in data scientists to sort of work on all data, I think you'll see more of that sort of data science AI role, even maybe see sort of incremental growth. I think the near-term play is not headcount reduction, but it's alpha capture. So building these systems to actually drive better, certainly on the junior side. You know, ask me again in 18 months or 36 months, and I may be pressing the alarm button on labor. But I think that I think the recruiting market and headcount market is going to be fine if I had to guess. It's a little bit different on the emerging manager side because I could run a $100 million hedge fund by myself now, generalist, small cap hedge fund, like I don't think there'd be any problem running that by myself now. Simple models, etc., or analyst models, etc. So I think some of my friends that run smaller funds don't need to hire as much. But that being said, like one of my friends that runs a small fund paused on new generalist hiring, and now he wants to hire an AI analyst, and so maybe one bucket moves to the next.

How do you identify when a skill should be subdivided to limit context rot? It's a good question. I'm trying to figure that out. Um, I'm working with a friend to think about that, like the actual engineering structure of the skills a little bit. Right now, I'm kind of just like, it's experimentation to see what works or not.

All right, I think that was all the questions I had in the box. Let me see. Do I have anything else? Any hands raised?

What does the skill set and background at a small shop look like compared to traditional? It really depends. You have small shops, of an AI analyst. Yeah, small funds are generally going to be, you know, a junior analyst is going to do a lot of things, where you'd have a more defined role at a larger fund.

All right. Well, we just did about a two-hour poke, so I think that's a good stopping point. Thank you for the time. I almost guarantee the next session will be shorter. I hope it's shorter. I said that every time. We will close up this free webinar series next Thursday, and then I'll walk you through, I'll walk you through a few. Not that we will continue to try and share a lot of this publicly, but we also are a for-profit business. We have a team and we'll be running programs around this. So, if there's interest, no hard sell at all. If there's interest, I'll sort of walk you through some of those opportunities next week. And certainly, if you're at a fund that wants some help on this, we are really launching into mid-day the Inerson Foundation seminars and some of the workflow labs and helping firms think through this. I find that very intellectually interesting. So that's how I want to spend a lot of my time over the next 18 months.

Last one from anonymous. Are you backtesting skills yet? No. I want to get to the point where I can have that automated inest variance output. And then I probably will actually run like an automated unautomated book, an automated book. But it's, I need to get the process sort of dialed in before I do that.

Was this one answered? How much can AI replace the reading stack? 10K, 10Q, earnings for analysts initiating coverage on a business or trying to establish a position or to find ideas for investment? This has been a common use case. Hey, I'm looking at a new business. Don't read the K's and Q's, earnings. Just press a button, get a primer. That's, depending on your context, I think that's a dangerous use case. You know, yes, it can accelerate a little bit, but you need to be really careful not to move into the mode of speed reading, right? You know, AI can do your thinking for you, but it can't do your understanding for you. So, you really are driving towards understanding at the end of an up-to-pe process. It's effectively an open-book test on how deeply you understand the business. How deeply do I understand Great Gatsby if I sped-read it or read the cliff notes? How deeply do I understand the business if I read an AI-generated primer, right? So I'd say be very careful about bypassing comprehension. There's no rule that you just because you have an AI system that you have to speedread through the K's and Q's and transcripts. So, I think just be thoughtful about that process. That means different things for different people.

And Mark, hello Mark, analyst academy grad as well. Seems AI is more helpful on the process front rather than the judgment front. How are you thinking about making sure the process leads to good judgment and not get lost in the process with AI builds and usage? It's a very common approach, like AI doesn't have judgment. I actually more and more disagree with that because what is judgment? Judgment is informed by all of these outputs of your investment process. I think more consideration is better. And so building a wider, more detailed research funnel should drive judgment. Judgment comes out of that detailed comprehensive process. I think the one thing that's obvious is AI is helpful in process. It's helpful in enforcing process discipline. I remember one time I walked through an Analyst Academy with a student. I walked through like, hey, you could do this for 60 hours, another 60 hours, and he's like, Brett, like this sounds great. You just gave me a roadmap to spend 200 hours on a company, and you also said 3,000 in a year. So like I can do this like 15 times in a year. Like that's it. Like that's a really good point. So I'd say like having been in a seat where I could spend 500 hours on one name, like there is that degree of rigor that you can apply to this process. The ability is like, okay, like I can spend 60 hours, but then I can do the next 100 hours with an AI-augmented approach and I can scale that across names now. So by the time I'm making a decision on a name, I've pulled up all of the key things management has said and all customer comments from a web scrape of different discussion boards on products, etc. I'm just making judgment with much better raw input, raw process output, and I can also turn the patterns of my past trading, etc., into that judgment engine. Hey, flashing red light, like you've been right 40, only 45% on this side of this type of trade. Flashing yellow light, this type of trade, you've been consistently six months too early. Like maybe you want to size it smaller, wait for three months before you lean in, etc. And so I think that AI can really facilitate. AI is a stateless sort of being, doesn't fundamentally have judgment and investing context. If anything, has bad judgment natively, sort of the opposite of what you would want. It can speed you to a consensus view. I think used correctly, adapted to your process, I think has an immense ability to facilitate better judgment and drive quality decisions. So that's what's interesting.

So, hello Mark, thanks for the question. Maybe that's a great place to stop now as well. So thanks everyone for the time today. And I will see everyone next Wednesday, if you're interested. Thank you.