How Do We Make Financial Advice Independent of the Advisor? | Abhishek Taneja

Alex: Hello, everybody, and welcome to the Curiosity Code podcast. I'm your host, Alex Khomyakov, and my guest today is Abhishek Taneja an AI and machine learning practitioner with over two decades of experience across retail banking and wealth management, focusing on data, machine learning, and business transformation. He has worked at major institutions like Citibank and HSBC, and is now with a major asset manager in Asia. He's work centers on scaling data, capabilities, putting machine learning models into production, and driving large scale AI transformations in financial services. Abhishek, welcome to the show.

Abhishek: Thank you, Alex. Happy to be here. Nice to see you.

Alex: Likewise, likewise. Let's dive into it. I have so many questions to ask. Most of your career was in retail banking and then you moved into wealth about three years ago. What surprised you the most about how wealth uses data and models?

Abhishek: Yeah, I think those two business are quite different. Honestly, the thing that hit me right away was that actually where the information lives, like in retail banking is so data rich because every credit card transaction, your account transaction is captured and there are data teams that are kind of looking at and we are building models and everything that a retail bank does, right, from approving your a credit card or a personal loan, to onboarding you, doing your KYC giving you loans, or you're increasing your credit card limit, or giving you personalized offers. Everything is based on data, right? But, you know, like earlier in my career, right, I have seen you know, we actually built out an enterprise MLOps operation where we have 100 models running in production, right? So, you know, so you can understand the scale of that, right?

Abhishek: And then I look at the wealth client data set, right? It's kind of shocking, right? Because the data is very thin, right? Like, instead of millions of customers, you might have thousands of customers. The AUM might be huge, right?

Abhishek: the AUM might be billions of dollars. And also, you don't have that much data, right? Because, you know, once you onboard a client, and they start investing with you they don't have too many reasons to kind of reach out to you often, right? Unlike in banking, you might want to check your balances almost every day, right? Or your credit card limits and such like that so but the biggest change is where that data lives right like you know if you think about it right for in wealth business the most important data lives in advisors' heads, right?

Abhishek: It's not captured somewhere, right? It's in their physical notebooks you know where they have had a chat with the client who's worried about you know getting some cash flow for their business or they're thinking about their retirement, or they want to send their kids to US for studies that is the most important data. And that's not captured. So that, I think, is a major difference between the two businesses. If you think about the wealth and asset management business we always had the quants.

Abhishek: We had the quantitative firepower. But that was mostly used in portfolio construction, like building alpha-generating strategies. right? And there we have been using machine learning and you know we had the some of the smartest people working there, right? But when it comes to actual last mile, right, that touches to the client, right?

Abhishek: It was entirely left on humans, right? Like, you know, advisors to their clients, they used to understand their financial goals, where they are in their financial journey and give them advice, right? So I don't think that, you know, the gap was, you know, that we didn't have good quants or data practitioners is this where the data was getting used right and how much data that we have

Alex: You give me one decision like wealth still makes on judgment that the bank would have modeled years ago?

Abhishek: Yeah like you know let's talk about like the you know the last one the advisory right Now, all wealth and asset managers will have a house view, right? So they will have an asset allocation view on saying, hey, if you are a conservative investor, right, and you want to preserve your capital this is our house view, right? You should put 60% capital in debt, 10% in gold, right, whatever, 10, 20% in equities, and then you should be diversified across geographies and all that, right? But that advice is mostly in PDFs, right? You know the research team will keep on setting out all this research which hardly anybody looks at, right?

Abhishek: Ultimately, what happens is when the advisor meets their clients it purely depends on their judgment, right? For example, that you can't quantify. For example, if I have two competing goals like I might want to retire in next 15 years I might also want to send my kid for higher education in 15 years, right which goal takes precedence, right? There's no mathematical model for that. It's entirely on the quality of advisors that are meeting the clients, understanding the situation, and giving them advice.

Abhishek: So that last mile of where the advice lives and who gives you the right advice has been very human-centric, traditionally. And that's one of my biggest challenges is how do we create a baseline where we can use data and AI so that no matter what kind of advisor you get, right? And if you think about it, as retail advisors The major asset management companies will not give you an RM that's only working with five or ten clients. Your wealth manager or your RM might be working with 200 clients or 300 clients. So if you're not one of their top five clients, they will probably not give as much focus to you.

Abhishek: So I think we need to create a baseline of right advice going to the customers in a seamless fashion. And we have all the tools, we have all the advisory, we have all the research teams that are doing that research. It's just how do we surface it to our clients. That's a challenge.

Alex: What's interesting is that, you know, while like AI is a technology is relatively new, the machine learning has been around for more than 20 years and banking leveraged that a lot in their production. In your opinion, why did wealth and asset management wait for that?

Abhishek: Yeah, so as I said that wealth management does use machine learning and AI, but they always use in portfolio construction and sort of alpha generating strategies, right? For me, I think personally there are like three or four things that stand out right? First is I would say is the feedback loop, right? Like, for example, if I build a fraud model, let's say a credit card fraud model I get the feedback almost immediately. Because what I do is I predict which customers are going to do a fraud in the next three months, or first payment fraud.

Abhishek: So within three months, I'll get to know whether my model is working or not. In wealth, the feedback loop is long. You might be in retirement, and the goal is 15 years away. How good is my strategy for you? How good is my portfolio construction for you?

Abhishek: I will not get to know, you know, like, you know, in short amount of time, right? So the feedback loop is very fast retail banking, right? So you can validate and retrain your model very quickly, right? So that's number one thing, right? Number two, I would say, is also an economic argument, right?

Abhishek: Because retail banking is a high-volume business, but the margins are very, very thin, right? So, you know, if you kind of track banking, right, it has huge impact on the rates, right, on the central bank rates. So you know, it's a very thin margin business, right? So automatic decision making kind of pays for itself, right? Because you know, if you have a good model, right, and you know, your NPA goes down say from 2% to 1.8%, that's like, that is hitting on your P&L, right?

Abhishek: So banks are always willing to invest a lot more money on their data research and building machine, any models which are like very accurate and you know kind of use all those kind of signals right. Third thing which is very specific I would say wealth management is that your data is so fragmented right like you have custodian banks, you have your transfer agents your you know customer data might stay in your core system right. You have your CRM systems you know like Salesforce which you are using for your sales team. So the data was very fragmented. And historically, wealth managers never put so much money in data plumbing, making sure that all the data is centralized, it's available for everybody, and you can start using that data.

Abhishek: And I think there's always this thing about regulatory accountability, because the banking regulators, they were very comfortable. If you say that I have this application scorecard, which decides which customers would I give a particular product, say, personal loans, right? The regulator were, you know, very, you know, like, very aligned, saying that, hey you can use an application scorecard, right, to decide which customers you want to give a loan, right? You can use FICO scores, you can use your own models, right? It's just, you know, regulators have to be comfortable with that, right?

Abhishek: In advisory business, right, I think you know, because, like it's very difficult to say whether your advice is right or wrong, right? Because it's so human dependent, right? That I think even the regulatory landscape has not progressed that much, right? So, you know, like in fund management business, for example, right? You have to publish your fund scorecard every month, right, you give your fund report, you kind of track it against the benchmark and all that, right?

Abhishek: But other than that regulators don't tend to kind of push you towards using AI or data a lot more right? So I think but you know like what's happening in recent years is that you know and all of us know this right the active fund management AUMs are you know dropping sharply and you know passive ETFs are capturing all that money right but if you think about passive ETFs right they work on margins of 0.06% 0.07% right so the same margin pressure that retail banking had those margin pressures are coming on asset managers as well right so I think that you know with that margin compression and the client experience has also changed. If you think about it, all of us are so used to having an Amazon kind of shopping experience. I get all this information, give me recommendations, I click on it, and I buy something. So the entire way of thinking about how do you sell products, how do you give advice is changing a lot.

Abhishek: And I think that's where generative AI can really help, because the other thing we to understand is that in our business, well management business, a lot of data is not in structured tables. It's not in your data warehouses, or data marts, anything like that. It's unstructured data. It's fund commentaries, it's meeting notes transcripts of earnings calls, and things like that, right? So I think that's where generative AI is so powerful because you can ingest all this data and start making sense of all this data that's available, which is outside your traditional data warehouses.

Alex: So, based on your experience and how AI has been implemented into wealth management tools I'm curious to understand where the hesitancy is strongest today. Is it in portfolio construction, alpha generation, or advice? And is it justified at all? Or is it just a question of time when we get there and it's just a pushback Because of spherical reasons?

Abhishek: I think that's a very good question. I think the hesitancy is overwhelmingly on the advisory side, right? Right. Because I think portfolio construction has been heavily quantitative, right for decades. Right.

Abhishek: Like we, as you know, right, we had like very, very smart quants working on building models. Right. And machine learning is just a kind of evolution step. Right. Like, you know, quants are using very traditional, you know, like, you know and your differential equations, right.

Abhishek: Machine learning is kind of this, I would say, one more evolution step towards that right but advice I think is a very different game right because because when you know like if you have an LLM that is spitting out like a personalized investment recommendation right and that recommendation goes down right who takes a responsibility for that because you know your advice is going to say Hey you know you recommended this right and that's why in advisory still, even the way we are thinking about creating an AI layer for advisory is that your reference portfolios or what should be required to a customer should come from a specialized team. It should not come from an LLM. And I don't know if you track this, a lot of people are going to ChatGPT or Gemini and uploading their portfolio and saying, hey, what should I do with this portfolio? Right? Or should I buy Nvidia or whatever, right?

Abhishek: That is okay if you're doing it for yourself, right? And you know, you can take that risk, but as a regulated entity, right? We can't let LLMs to give you a portfolio advice, right? So right now, the hesitancy is that advice has to come from humans. So we have this whole team sitting that their job is to create model portfolios for each kind of customer, right?

Abhishek: So broadly, we think about it, right? We divide your customers into what is your risk profile, so whether you are conservative, moderate aggressive, highly aggressive, something like that, right? So it could be like a three to five scale, point scale. And then what is your investment objective? Is it you want capital preservation, you want capital growth, you want an income portfolio, you want capital preservation and income both, right?

Abhishek: So they could be mix and match, but ultimately, we look towards our research team to give us model portfolios within each of this group of investors, right? And we don't want to use LLM at all to give any sort of advisory, right? Because, you know, these teams are trained to do that, right? So, you know, they make sure that, you know, whenever they recommend a portfolio right, it is well diversified. They take a consideration of max drawdowns.

Abhishek: What's the volatility? What's the alpha? What's the Sharpe ratio? So they'd look at tons of metrics. Now, you can argue you can do that with an LLM with an AI as well.

Abhishek: But I think, you know, right now, neither we as an organization and neither our regulators are comfortable that you kind of create advice using LLMs. I think that it will evolve, right? But I think, you know, using AI for all the surrounding layer outside the advice of how to do your portfolio construction, right? Like, why should you use this portfolio? Why is this portfolio better than your existing portfolio?

Abhishek: Or what should you do next? All the surrounding areas, which are a little low risk I think AI layer can be built around that. But the core, the advisory part, I think we still would probably for some time have humans building that advisory piece.

Alex: Yeah, that's interesting, especially how you draw the line between asset managers and hedge funds. And that's actually one of my next questions, because what I'm seeing happening more and more, that hedge funds hire machine learning and AI engineers alongside quants. And I'm trying to understand, what are they willing to do with AI that traditional asset manager is not due to regulations or risk management? Where is this line? I'd like to understand it better.

Abhishek: Yeah, hedge funds is an very interesting case, because hedge funds, if you think about it, right? The whole model is, you know, it's you know, they, they grant you a certain return. So they'll say, I'll guarantee 20% return. Then anything I make above 20%, you give me half of that, right? And they have it, you know, 2%, you know, AUM charges or whatever, right?

Abhishek: So the best, just because of the business model is such, that you're generating alpha and you're getting paid a lot for that. Right. But you know, I think fundamentally the way it is built is it's kind of what is in statistical edge, right? And that statistical edge can come from anywhere, right? Like, you know, hedge funds, they use so much non-traditional data, right?

Abhishek: From satellite imagery to, you know, call volumes you know, they have people who actually track how many trucks are going out from factories, right? They kind of look at all the supply chain data. So, you know, and, you know, like, there are hedge funds which kind of use petabytes of data right? And they build very, very sophisticated machine learning models, right? To generate that alpha.

Abhishek: But because, you know, the regulator doesn't care about that, right? The hedge funds are for very sophisticated customers, right? Mostly institutions or high-net-worth individuals, right? It's not for your another middle retail customer, right? So, and, you know, the regulator doesn't care if a hedge fund is not able to explain their model in English, right?

Abhishek: Like, but, you know, a bank has to explain their model in English to a regulator. Like, if you don't approve a credit card for you and you approve a credit card for me, the bank has to explain why did your credit card application got rejected Right? Or something like that, right? So, you know, the hedge funds, the regulatory burden is much lower than banks or traditional wealth and asset managers, especially if you're taking retail money right? Because retail money is money from your mom and pop shops.

Abhishek: It's for salaried people who are saving money for retirement, right? So, you know, there is a lot of difference on what a hedge fund can do versus what we can because we have different social and legal obligations and contract with our customers. So, yeah, so I think the, you know, like if you think about even like a new product launch, like hedge funds can create products like immediately, right? When we create a new product, it has to go through governance committees, you know, we take you know, like view from our compliance team, from legal team, right? Is it all okay, right?

Abhishek: Will it pass through the regulatory or not? So like you can't go to a regulator and say, you know, our deep learning model has come up with this portfolio and we are going to use it, right? Because they will need explainability. They will say, hey, explain to me how does it work, right? So I think, you know, neither model is wrong.

Abhishek: I think, you know, these are like two different games altogether, right? I don't think that, you know, we will ever go to the hedge fund route of kind of building this black box models, right? Which we are not able to explain but I think we will also start learning from them in using a lot of non-trivial data, whether It's satellite data, weather information, traffic data, things like that. It's just we need to find a way that when we get all this data, we should be able to go to a regulator and say, hey, this is the data that we have got, and this is how we are using it, and it has to be something that is going to be beneficial for our end customers right so I think that's that's broadly you know what I think the difference is

Alex: Got it. So most of our listeners are in the US and Europe and I'd like to take advantage of this conversation and tap into specifically South Asia, and Understand how does a person in Asia get a financial advice? Can you walk me through the channels from bank relationship managers to do-it-yourself platforms to tied advisors?

Abhishek: Yeah, sure, and I think for your Western listeners, I think they need to understand Asia is hugely diverse, right? So you have like, you know, like, ultra mature markets like Hong Kong and Singapore but you also have, you know, rapidly growing emerging markets across Southeast Asia there's India, there's China, right, which are huge markets, but not as sophisticated as Western markets, right? So, I would say just one defining thing that kind of separates Asia, like Asia has a broad category, right, is that both advice and broad distribution is kind of bundled together, right? Because, you know, in the U.S., you have this fee-only fiduciary RIA, registered investment advisors. That model is very dominant in the U.S., right?

Abhishek: And I think in the U.K., also is something like that where you know you have fee-based advisors right so because they have a fiduciary duty right so they don't care about what product they're selling to you they they are supposed to work on your best interest right so if for example Fidelity S&P ETF works the best for you is the cheapest available they will just recommend that to you because their incentives are not linked to what product they're selling right their incentives are linked that because you're paying fee to them right for advice it's a totally different model in Asia right because the incentives are so different so because you're selling both the advice and the product right so there's always a tendency to sell products which are higher margin right so that's why you know it's a very different kind of model where you know most of the relationship is owned by the relationship managers at your bank so you know the banks they not only sell you banking products they also sell you investment they sell you insurance they sell you retirement products right they sell you everything right and because the RM has such intimate knowledge about you right because they know how much money you have right where you stand and all that right so they are the first port of call right secondly you know for like a broad middle class and emerging affluent families in Asia the second biggest channel of where they get their advice and product is from the tied insurance agency force right and if I can explain to you a little bit a tied insurance agent is somebody who works for a particular insurance company. So, this person only works for a particular insurance company and they sell their products. Again, the tendency is that they tend to sell products which are of higher margin. They might or might not have the best interest of their clients in mind when they're selling products to you and give you advice. But for the large swathe of people in Asia this model really works because this is the only way they invest.

Abhishek: If you think about Do-it-yourself kind of investors that group is very small sliver in Asia. There are very few customers or investors who are confident enough in their skills that they can do the investments themselves. They can open a brokerage account and start investing themselves. Most people don't do it. The model has flaws, but I think at least it gets people in the door, right?

Abhishek: Like, they start investing, right? And that's a huge win. So, you know, and insurance agents typically will be somebody in your relationship, right? It could be your uncle, it could be your neighbor, it could be a family friend, and they come and sell you insurance, and they'll give you advice, say, why don't you start saving money? The third model is basically your bancassurance, right?

Abhishek: As I said, banks not only sell you their own products, but they also sell you insurance, right? And investment products. So, banks, they can be producers of the products so they can have their own funds or their own insurance or they could be third-party distributors right where they you know they have a shelf of products from various insurance companies and investment companies and they come and sell it to you right so it's a very different kind of model right I think the younger generation is becoming more and more comfortable with do-it-yourself kind of brokerage apps and the robo advisory where you know the costs are far lower but I think that self-directed channel is still very small and I think for a few years to come I would say another decade or so these channels right the bank channel the insurance advisor channel and the bank digital channel will remain the three big channels for both insurance companies as well as asset management companies.

Alex: In our introduction call, you tap into something interesting that I'd like to explore here. You ask a question, how do we make advisory independent of the advisor? What does it look like in practice?

Abhishek: So when I said that I didn't mean that we need to get rid of advisors, what I meant to say was that how do we make sure that there is a guaranteed floor of advice right? Like nobody can go below that. Like, you know, nobody should give you wrong advice because You know, like on ground, what happens is that, you know, people always try to catch the previous wave, right? Like today, right? If you're investing semiconductor funds, the semiconductor run-up has already happened, right?

Abhishek: You're not going to catch another Nvidia or Micron, right? But what happens is people look at last three years performance of a fund and they say I should buy a semiconductor fund. They've done really well. That's not a good advice because our view is you should always be diversified, you should always have multiple asset classes in your portfolio, you should be diversified across geographies and you should just focus on your goal. So if your goal is 20 years ahead, you want to retire in 20 years and if your portfolio can give you 8% return, you will meet your goal.

Abhishek: So, you don't need to catch all the hot trends, you don't need to get into bitcoins or whatever. So, when we look at closely at how advisors actually prepare for client meetings there's no analytical layer, there's no shared data layer that they can go and say hey, what should I recommend to this customer based on customer's life stage, the risk profile, the investment objectives, right? So it really depends on your luck, right? Like what kind of advisor you get, right? Or what kind of advisor gets assigned to you, right?

Abhishek: If it's a very professional experience advisor, they will give you good advice. And it's not even a question of intent, right? It's a question of capability as well as understanding of the markets, right? So what I meant to say was that there should be a base level advice that everybody should get. No matter how big is your portfolio, no matter whether you are a new investor or old investor, what is your asset size, we should be able to give you a fairly robust advice which helps you to meet your goals.

Abhishek: And then the job of advisor, instead of going through all your previous records and making an Excel sheet for you to kind of put where should I contact you, how should I explain your extreme portfolio to you, not all that is kind of, you know, we have an AI layer that does all that, and your job just becomes as an advisor to build a relationship, right? And kind of be a sounding board to your customers, right? Like, you know, go through the emotional journey, right? Because markets go up and down, and you know, you need somebody on your side who can keep you calm, who make sure that you don't sell at the low, right? Or don't buy at the high right make sure that you know you are consistent and you know you just focus on your goal right so what I think will happen is that you know all the things which are add-on to for an advisor to do right like all the data they need to collate for portfolio checks right like how's your portfolio doing what's the next best action what should I tell you next right how do I rebalance your portfolio all that you know there'll be data in AI layer that will do it for you and there'll be audit trail that will say hey what advice did I give to a customer and what was the outcome of that advice.

Abhishek: But outside of that, human-to-human connect will be the key job of an advisor. So that's what I meant when I told you that.

Alex: Do you think people will generally accept AI advice as long as a person signs off on it?

Abhishek: I think the challenge is that there's tons of good advice out there. If you think about when internet first came, all of us thought it's a huge boom. People will learn everything that there is to learn and we'll go through YouTube through all the tutorials and we'll learn new things. None of that happened, right? In fact, social media happened.

Abhishek: And what I'm trying to say is that, you know humans are not very good at kind of doing the right thing, right? So you might have the best advice available, right? But how many people actually action on it? Right? Like, I think, you know especially in Asia, right?

Abhishek: I still believe that it's not that people don't have knowledge, right? Or they can't find information, right? It's, you know, they still need hand-holding right they still need a somebody who can listen to them who can understand where they are at their life stage right and you know who can make you know like make judgment call for them right like it requires a lot of human conversations right and you know you need to you need to guide your customers on what the right thing to do is right like for example right everybody knows that you should do dollar-cost averaging. You should keep investing a certain amount of money every month because when markets are down, you get more units, and it kind of makes the volatility much lower. But when we look at the portfolio that we have very few customers do dollar-cost averaging.

Abhishek: Very few customers consistently invest every month. So whenever they get their bonus or they get some money, they invest the money, and they take it out very soon, like if they see 20% returns, they take it out. So it's not that they're not aware. These are professionals, they run businesses they are very smart people. But if you think about it, investment is the easiest thing to do, and like arithmetically, but the most difficult thing to do emotionally.

Abhishek: Because arithmetically, if I put, say, $500 every month in S&P 500, I just buy VOO. And I keep on doing it, right? You will reach your goal, right? But how many people actually do it, right? Like the advice is very simple, right?

Abhishek: You know, save 20% of your money, keep on investing it, buy low-cost ETFs, and you are set, right? But people don't do it, right? And it's more of an emotional decision than a mathematical decision, right? Because if you look at the mathematics, right? We know that over 20 years, like US markets have been on like 10 to 12% returns right?

Abhishek: But that's with broader volatility, right? Markets have gone up 40%, gone down 30%, right? So you still need at least in Asia, I feel, you still need that human connect to kind of guide you through all those market turbulence and make sure that you're on the right path.

Alex: Yeah, that actually leads to my next question and it's interesting to explore it because it could be related to cultural differences, generational differences. So I had a guest here on the podcast and he argued that an AI co-pilot can give people directly what an expensive advisor gives them. And your route goes through the advisor. So I'm trying to understand, like is that like just two different points of use or there is a difference and how people actually using technology and perceive the advisory in general. Like, as I said, depending on their culture, depending on the geography, depending on the generation.

Alex: What's the point of that?

Abhishek: Yeah, I think, you know, there'll be segments of customers, right? Like today, you are a digital savvy high literacy customer, right? You have high financial literacy and you are very sensitive to fees. You don't want to pay any high fees, right? You can do it yourself, right?

Abhishek: Like, you know, in US, you can go to Robinhood or, you know, in Asia also, we have these platforms which are super cheap. You can buy, you know, you can build your own portfolio, right? Can buy multiple asset classes and start investing. But I think the focus I have on empowering our advisor first is it's kind of a sequencing choice, right? Because we still have all these customers who have grown in that era of banking, RMs coming to you and giving advice and selling you products, right?

Abhishek: And they're still working, they're like my age, right? So they still need that And they are not that sensitive to fees, they are willing to pay some money out if they get good advice. So if you look at across every Asian market, and I can't speak about European market or American market You see a very interesting trend. People start with robo-advisory or they will have these discounted online brokers where they start putting money. But as soon as their money hits a certain amount, let's say $100,000 or $150,000 know, a 20% drop in the market when you made your own decisions feels very different compared to when you have an advisory.

Abhishek: So what we see is that once people have a certain amount of money, they start going to advisory because now they have much more to lose, right? And they don't mind paying like 60 basis points or 70 basis points for advisory. If they can get this piece of mind that, hey, you know, like somebody professional who's looking at is my portfolio and you know you know it can work for the volatility and all that right so the amount of money that is with you know these discounted online brokers or you know like these robo advisors very small it could be five ten thousand dollars fifty thousand dollars but if you look at the amount of money that banks have or you know the traditional broker houses have that's massive right because the losses become huge right like if on a 20% market drawdown if you're going to lose a million dollars right then you you don't worry about the advisory fees right you want somebody to make sure that you know you're on the right path right yeah at the second I would say also there is a regulatory reality right I think like a fully autonomous personalized financial recommendation engine, I don't think anybody has been able to crack it, even the, you know, the online brokers or, you know, the robo advisors, right, they will give you a portfolio, but they will still make the customers choose it, right? Because they will never take a financial liability that you know, I give you this portfolio and they will give all that disclaimer saying hey, you are choosing it based on your own volition, right? You can change it, right?

Abhishek: You can do whatever, right? Because I don't think the regulators are comfortable with any AI based or even machine learning-based advice going directly to customers and customers so this there's always this friction that the ultimate decision to buy something is still with that customer right because none of the intermediaries will take that risk of taking a decision on your behalf right so they will say hey you this is my recommendation to you but you have to make the choice you have to buy it yourself right so that you know so that we have say body key things that human behavior there's a regulatory reality to it and I think there's a trust factor, right? People don't trust their judgment so much especially when the judgment is wrong right? So when you have much to lose, you start going back to the traditional advisory channels because you get some comfort of mind, right? You get some, you know, peace of mind that, hey, I'm talking to professionals and you know, they're looking at it.

Alex: Yeah, yeah, that makes sense. The end of this episode, I'd like to Get a bit technical and explore things that you wrote about tree models and tabular foundation models. Can you, first, can you, in plain terms for the listeners explain what the difference between the two and for a decision about someone's money, which would you trust in production and why?

Abhishek: So, you know, the way they work are similar, but, you know, the traditional gradient boosted decision tree models, right? They are basically a bunch of decision trees, right? And you know, you had a bunch of techniques to kind of see where you do early stopping, or how do you cover losses, how do you do gradient descent, all that right? So if you think about models like XGBoost, right? Or like GBM, they are all basically tree-based models, right?

Abhishek: And they are very good for structured data. So if you have structured data, right, which is labeled data. So ultimately, for example, you know, let's take an example of say, credit line increase, right? So, you know, I have a base of two million customers. I want to give some customers a credit line increase, right?

Abhishek: Say of 30%, right? And my target objective is very clear. I want to give, so either you'll get approved for credit line increase or you'll not get approved for credit line increase, right? So it's a binary decision, you know, one or zero, right? So whenever you have objective function which is one and zero and you have all your tabular data properly structured in data warehouse the tree-based models are the best models, right?

Abhishek: Because they're very explainable. They'll tell you exactly what is the weight of each other variable, why did you make this decision, right, and like that. Now, the way tabular foundation model works is that they are kind of pre-trained right? So basically, you know, they have kind of used tons of synthetic data and they've created these models, which will do the same thing because ultimately this is a classification model, right? You classify yes, no, one, zero, right, whatever, right?

Abhishek: So but because they are pre-trained right you can do you can do one shot like for example you can upload your base and you can say hey I want to you know build a credit approval model right and it'll it will try to build a credit approval model you know it's like prompting right so you you upload your data because it's already pre-trained on large amount of structured data it will come with a result right but I think you know these models are still not very mature right so if you think about your traditional risk models They are still completely based on tree-based models, like XGBoost or LightGBM. Because they remain the, I would say they are the workhorses for banking and financial services. Because they're very fast. They're deterministic. And they're very stable.

Abhishek: And they're easily retrainable. So for example, when your model deteriorates, because the underlying population is changing, you can retrain those models, right? And you'll exactly know how much iteration happened on your model and why did you retrain it, right? And you can put like hard constraints. For example, if your regulator says that, you know you can't have any you can't use gender to make your decision, right?

Abhishek: You can't have different outcomes for say, men and women, right? You can put that constraint in, that you know, don't use gender as one of your criteria or one of your parameters right or you know if there's some adverse selection like for example in some regulatory you know markets there's a constraint on unfair bias towards elderly population right like I can't reject somebody who's more than 50 years old just because of their age right so I can put all those constraints in the model and it's easily explainable so I can go to my regulator and say hey look this model is totally fair it doesn't you know treat any segment of the population unfairly, and I'm just using other parameters, right, to kind of build that model, right? Now, I have spent some time on experimenting with the, you know, tabular foundation models, right? And they're very impressive, right? You know, because you don't need to do hyperparameter tuning, it's very easy to build fast, right?

Abhishek: But I don't think they're ready for the prime time, because the explainability is still not there, right? So I would use the tabular foundation models as a challenger model, like, you know, where I'm building something new, I can challenge, I can build two models, one traditional XGBoost, one tabular foundation model, and use it as a challenge model and see, it does better separation, it does better classification. But if I'm building a model in production, where I need to go to a regulator and explain the model to them I would still kind of prefer to use gradient-boosted decision tree models for now. Until this model has become ready for prime time. You can put it into production, and you can explain it, explain to regulators, right?

Abhishek: But I think this is a very interesting research area, and I think I'm sure there are more and more research happening in this. Because, you know, that structured data, right, the tabular data, that's not going anywhere, right? There are tons of workflows that are still based on tabular data you know, where you're building classification or regression models. So I think that area will keep on evolving.

Alex: Abhishek, my last question is five years from now, what does the advisor's job look like?

Abhishek: Oh, I see. I think the job will become with a lot less administrative friction. Like today, when an advisor goes to office, the first thing they do, suppose you have 300 customers, you go to an Excel file, look at which customers have what portfolio, what's the performance, you look at their NAVs their three-year returns, five-year returns, things like that, right? All that you're doing manually, right? I believe five years from now, right?

Abhishek: What will happen? Your advisor will go to the office. They will have, hey, these eight customers you should have a conversation with them because somebody is making losses, right? Or somebody is about to retire or whatever, right? So you will know which clients to talk to, right?

Abhishek: What's happening with their portfolio? Is it going through a drift? Are they making unnecessary risk, right? Or, you know, their liquidity is changing, maybe they're retiring, so, you know they can't invest money, they need to, you know, take money out from their portfolio, right? So, you will have all your pre-meeting diagnostic, you know, alignment with your reference portfolio, everything ready, right?

Abhishek: And, you know, you will get English language talking points that you can just go and talk to your clients, right? So, all these people will go away, right? And your job will become as a professional is, you know, Now, how do you make sure that the clients are doing the right things for their long-term goals? You know, like maybe I want to leave some legacy for my kid, right? How should I think about doing, you know should I invest in property or you know, should I, you know, you know, what should I put in my will right, thing like that, right?

Abhishek: So I would say that a lot of this administrative overhead will go away. They will not need to memorize all the sheets or pushing whatever they know best about. All that will be taken care of. Their job will be just to make that relationship grow and grow with that relationship, right? Because one thing if you think about right now, right, is that if I'm an advisor and I have 300 customers in my book I will only focus on my top 10 clients who are like maybe 70% of my portfolio right?

Abhishek: That Pareto principle works everywhere. And the bottom 290 are just there. But whenever I need to kind of increase, I need to meet my quota, right? I go to those 10 clients because I know they will give me money, right? And I give them the best service, right?

Abhishek: I think that, you know, that long tail problem, that you have a long tail of customers that you can't service because you don't have the time, bandwidth and their assets are not enough for you to justify to spend time with them, right? That long tail problem will go away because all that will be automated. And you know you will know at the right time which customer to reach out to what should you tell them how should you rebalance their portfolio and what should be the next best action for these customers right so that's my hope that's what we're working towards and I really think that in next 10 years that probably will be possible

Alex: I love that outlook into the future. Abhishek, thank you very much for your time. It's been a great conversation.

Abhishek: Great. Thank you, Alex.

Alex: And for the listeners, thanks a lot for listening till the end, and don't forget to hit the like button on YouTube, subscribe to the channel, and leave feedback on the podcast platform you're listening to, and see you in the next episodes. Bye-bye.

Abhishek: Thank you.