{
  "video_id": "ui7kRlJMqjM",
  "channel_slug": "mitcsail",
  "channel_handle": "mitcsail",
  "title": "MIT Professor on How AI & LLMs are Shaping Financial Advice, Analysis, & Risk Management: Part 1",
  "duration_seconds": 1179.0,
  "url": "https://www.youtube.com/watch?v=ui7kRlJMqjM",
  "upload_date": "",
  "transcript": "- Hi everybody. I'm Andrew Lo,\na professor of finance at the\nMIT Sloan School of Management\nand a principal investigator\nat the MIT Computer Science and AI Lab,\nand I'm looking forward to\nanswering your questions.\nLet's take the first one.\nHow can large language models\nefficiently analyze financial reports\nto extract key insights\nsuch as identifying risks,\nopportunities, and emerging trends?\nLarge language models\nare actually designed\nto be able to read plain text\nand be able to digest and summarize them.\nEarnings reports, financial statements.\nThese are the kind of things\nthat large language models can now read\nand be able to generate very,\nvery short and pithy summaries\nto be able to focus on things\nlike risks, opportunities,\nand other issues.\nIn the financial sector,\nwe use certain keywords\nto be able to highlight\nrisks and opportunities so\nthat there are automatic tags\nthat large language models\ncan figure out very, very quickly.\nLarge language models\nwill have a huge impact\non the life of financial analysts\nbecause their job day in and day out\nis to be able to read these reports,\nand they can now do that\nusing these tools in a much\nfaster and more accurate way.\nCan LLMs identify subtle market patterns\nand anomalies that may\nelude human analysts?\nThe answer is yes, they can,\nbut they can also identify\nanomalies that don't exist,\nin other words, hallucinate.\nAnd this is one of the challenges\nthat we have to deal with with LLMs.\nRight now, we know that\nthey're very capable\nof being able to identify some\nreally interesting patterns,\nbut at the same time, they\ncan also identify things\nthat aren't there and require humans\nto engage in additional oversight.\nNow, by the way, humans\nare not perfect either,\nand we are prone to\nhallucinate on occasion.\nPaul Samuelson famously said,\n\"economists have predicted five out\nof the last three recessions,\"\nand I think that's part\nof what he was getting at.\nWe need to think about\nhow we can make ourselves more accurate.\nSo the combination of humans and LLMs\ncould actually be the sweet\nspot of being able to produce\nreally accurate forecasts\nfor all sorts of economic phenomenon.\nHow can we build trust in financial advice\nand decisions provided by LLMs?\nAnd what strategies can be employed\nto maintain human oversight and control?\nNow, this is a really interesting one\nbecause it gets at an\nissue that my collaborators\nand I have been working on\nfor the last several months,\nand we think we have an approach,\nbut it'll be several years\nbefore we come to it.\nAnd the problem that we're\nstudying is the issue of trust.\nHow do we get LLMs to be trusted\nand specifically in the\ncontext of financial advice?\nSo in that domain, we\nactually have a concept\nthat is known as fiduciary duty,\nand this really gets\nat the heart of trust.\nIn the financial sector,\na fiduciary is somebody\nwho puts your interests\nahead of his or her own.\nSo for example, if you\nhave a financial advisor,\nthat financial advisor\nowes you a fiduciary duty,\nmeaning that when they give you advice\nabout how to manage your money,\nthey have to be thinking\nfirst and foremost\nabout your welfare as opposed\nto lining their own\npockets with commissions\nthat might come out of the kind of trades\nthat they would do on your behalf.\nHow do we get an LLM\nto become a fiduciary?\nHow do we get it to be a trusted\nfinancial advisor to you?\nOur approach is to think about\nhow this happens in the industry\nwith human financial advisors.\nHow do human advisors become fiduciaries?\nWell, for one thing, they\nhave a code of ethics,\nso they have to follow certain rules,\nbut more importantly,\nthey have to follow a large\nnumber of financial regulations,\nlaws that we put in place in\norder to protect consumers.\nAnd at one point, I\nremember studying these laws\nbecause I had to take a securities exam\nin order to be involved\nin a particular kind of\na financial institution.\nAnd this exam called the Series 65\ntest you on your knowledge not\nonly of financial analysis,\nbut also of the regulatory infrastructure\nthat guides all of us in the industry.\nAnd so after spending hours and hours\nreading all of these rules\nand being frustrated that I was forced\nto memorize these things,\nit finally dawned on me\nwhere these things were coming from.\nThe historical body of case law\nthat has been used to\nguide lawyers, regulators,\nand consumers as to what\nthey can and cannot do,\nthat is a fossil record of all of the ways\nthat one human has decided\nto be able to try to take\nadvantage of another human\nand ultimately got caught\nand prosecuted successfully for it.\nSo if we trained LLMs,\nnot just with the body\nof financial knowledge,\nbut the full case law history\nof all of the various different lawsuits\nthat have been filed\nagainst certain bad actors\nin the financial system,\nif we do all of that,\nwe then can train LLMs\nto become fiduciaries.\nWe believe that we're\nstill a few years away\nfrom something that the SEC\nand lawyers would agree\nwould be a fiduciary,\nbut we at least know the\ndirection that we're going,\nand we're very excited to\nbe able to go that route\nand ultimately produce a piece of software\nthat can be fully trusted by humans.\nWhat role could LLMs play in automating\nand streamlining risk assessment processes\nfor banks and financial institutions,\nand how should these systems be governed?\nSo this is also a great question because\nthat's one of the most important features\nthat financial institutions\nhave to carry out\nis risk management.\nRisk management has two parts, in my view.\nThe first part is a quantitative part,\nand that's actually\nrelatively easy to automate.\nComputing things like value\nat risk and scenario analysis\nand worst loss scenarios,\nall of those things can\nnow be done pretty much\nat a push of a button\nand and are being done\nby all the big financial institutions.\nWhat is much more difficult\nis the second set of tasks,\nand that is taking all of those numbers\nand putting that into a\nnarrative that can be given\nto a risk manager,\nto a policymaker,\nto a customer,\nso that they understand\nwhat the implications\nof those numbers are for their\nown personal circumstances.\nImagine that the stock\nmarket just went down\nby 15% today,\nand that means that your portfolio of\nbonds are now in trouble\nbecause the stock market has gone down\nand there's a concern that the\nFed is gonna have to step in,\nand everybody is worried about\ndefaults across various\ndifferent kinds of scenarios,\nand you're holding a\nbunch of corporate bonds\nthat are risky to begin with,\nbut in the today's\nscenario, it's a lot worse.\nWhat do you do about it?\nAn LLM could analyze the\nnumerical data instantaneously\nand then weave that into a\nnarrative which says, perhaps\ntoday equity markets\nare down by quite a bit.\nThat means the Fed's gonna step in,\nthere's gonna be panic\namong a bunch of investors,\nand that means that certain\nassets will sell off\nand other assets will\nbecome much more valuable.\nIf you're holding corporate bonds,\nthat is likely to get hit hard,\ntreasury bills are gonna do really well,\nbut I would wait for\nanother three to five weeks\nbefore you do anything dramatic\nbecause over the course of history,\nif you looked at similar situations,\nthese kinds of moves are temporary\nand things will normalize.\nThat's the kind of narrative\nthat an LLM can provide\nthat is not available\nwhen you're just looking at\na whole table of numbers.\nSo over time, I suspect that\nLLMs will get much better\nat weaving those narratives\nand also being able to describe\nsome of the weaknesses in their narratives\nand where you might wanna use\nyour own personal judgment\nto make a call for your own portfolio.\nSo I think we're not that far away\nfrom the use of LLMs for risk management,\nand I'm hoping that\nthat will get everybody\nto focus on the right kind of risks\nthat are relevant to them.\nHow can the natural language\nprocessing capabilities\nof LLMs be used to\nperform sentiment analysis\non financial news,\nsocial media, and other\ntextural data to inform trading\nand investment decisions?\nSo typically, we believe\nthat financial markets\nare moved by two things, fear and greed.\nNow, that's the Wall Street\ntrader's perspective.\nIf you're asking an economist,\nit's all about the numbers, right?\nThe bottom line is it's both.\nIt's the numbers, but it's\nhow we interpret the numbers,\nand that's where sentiment comes in.\nSo the idea behind sentiment analysis\nis to try to understand\nhow human emotion is gonna\nreact to the numbers.\nAnd over the course of\nthe last several decades,\nthere's been a lot\nwritten on human behavior,\npsychology and the ability for us\nto manage our emotions\nin very, very difficult\nfinancial circumstances.\nObviously, most of us\nhave trouble doing so\nbecause we are all hardwired\nto engage in the fight or flight response.\nWhen we are threatened,\nwe will react in a very\npredictable way physiologically,\nwhich is great if you're being chased\nby a saber tooth tiger\non the plains of the African savanna\na hundred thousand years ago,\ndoesn't work so well on the floor\nof the New York Stock Exchange today.\nSo sentiment analysis is an attempt\nto try to look at the various\ndifferent financial indicators\nto get a sense of\nwhether or not the market\nis overreacting or underreacting.\nAnd I think this is where\nLLMs will have a field day\nbecause they're gonna be\nable to look at the numbers,\nbut more importantly,\nthey're gonna be able\nto read the literature\nof what's being written at that\nvery moment by news sources\nthat are freaking the rest of us out.\nWhen the news stations tell\nus, is there something in milk\nthat could be hurting your infants?\nMore at 11.\nYou're gonna feel compelled to watch\nthat news story at 11 o'clock.\nSo we are very easily moved\nby those kinds of threats.\nAnd so LLMs will be\nvery good at picking up\nthose kinds of threats\nand coordinating the\nanalysis of the language\nwith the numbers to be able\nto produce sentiment analysis.\nIn fact, I wouldn't be\nsurprised if certain hedge funds\nwere already using LLMs\nto be able to detect\nthese kinds of patterns\nand start making use of them.\nThe hope is that the\ntypical retail investor,\nthe rest of us consumers\nwill have access to those tools soon.\nOkay, next question.\nHow can we mitigate bias in\nLLMs for financial applications,\nand what other ethical considerations,\nlike algorithmic transparency\nand accountability,\nshould be prioritized?\nSo the first thing to note\nis that LMS absolutely do have biases.\nAnd I know this because my students\nand I documented that in a paper\nthat we wrote recently\nlooking at hiring decisions\nthat an LLM might make\nwhen confronted with a variety\nof different candidates.\nSo it definitely suffers from gender bias\nin a variety of different contexts.\nAnd it's not surprising because\nwhat are LLMs reflecting?\nThey're reflecting the sum\ntotal of the literature\nthat they're using as\ninputs to be trained.\nThe first step in dealing\nwith bias is to document it.\nWe need to understand, depending\non the nature of the LLM,\nhow it's trained, other\nsupplements that we use with it,\nwe have to understand exactly\nwhat those biases are.\nWe have to quantify them.\nOnce we quantify them,\nthen we can start asking the question,\nhow do we decide to change the biases\nto make it more appropriate\nfor the purpose at hand?\nSo that's a question\nthat requires domain specific knowledge.\nIn certain areas, the\nbiases may be very small.\nIn other areas, the biases may be huge.\nSo for every single application,\nI believe that we need to think carefully\nabout the implicit biases\nin the LLM that we're using.\nAnd once we document that,\nto be able to then start\nengaging in recoding\nor adding various different supplements,\nretrieval, augmented guides, rags,\nthat would actually lean\nagainst those kinds of biases\nto the degree that we wish.\nOver time, we need to understand\nhow these biases are changing.\nThey change across time,\nacross culture, across country.\nSo understanding just exactly\nwhat these LLMs are doing\nis something that is\ngonna be a prerequisite\nto us being able to put any\nof these things into practice.\nYou have to measure before you can manage.\nCan LLMs be used to enhance the detection\nand prevention of financial fraud?\nI think the answer to this is\nunambiguously absolutely yes.\nRight now, there are a number\nof machine learning tools\nthat are already being used by the SEC\nand other agencies to\nidentify potential fraud.\nA long time ago, it was\nsuggested by a mathematician\nthat there are certain\nstatistical properties\nthat have to exist among\na table of numbers.\nAnd so if you see a departure\nfrom that statistical regularity,\nthat's an example of fraud.\nSo using these kinds of distributions,\nalready we can detect\ncertain types of fraud,\nbut now with LLMs, with\nmore sophisticated ways\nof processing natural\nlanguage and numbers together,\nwe can actually do even better.\nSo there, there's no doubt in my mind\nthat this is gonna be a\ntremendously powerful tool.\nThe dark side of this\nis that with these LLMs,\nwe can also create fraud\nthat is harder to detect.\nFor example, imagine prompting\nyour LLM by asking it\nto take a look at your tax returns\nand suggest ways of putting\nin certain kinds of deductions\nthat will give you a much lower tax bill.\nAnd even if they break the rules,\nto do so in a way that makes\nit virtually impossible\nto detect by a typical IRS auditor.\nNow imagine if you gave\nthat prompt to an LLM\nand imagine if it could actually\ndeliver on that request.\nThat's the danger.\nAnd one of the reasons why I\nthink we're in an arms race\nbetween the regulators\nand the perpetrators,\nand one of the reasons why I believe\nthat we ought to increase the budgets\nof regulatory authorities\nbecause they need to have\nthe same type of equipment\nand sophistication to be able\nto address these concerns\nso that they can stay\nahead of the fraudsters.\nIn what ways can LMS\nassist in the development\nand testing of more\nsophisticated trading algorithms?\nI believe that they're\nactually already being used\nfor just that purpose.\nSo it used to be the case\nthat machine learning algorithms\nreally had to be focused\non the specific feature\nthat you were giving it\nto be able to detect\npatterns in financial data.\nAnd that's really what\ntrading algorithms are.\nIt's really just pattern matching\nso that you can predict\nwhat's gonna happen tomorrow\nand trade today to take\nadvantage of that prediction.\nBut now imagine being able to\nmake those predictions based,\nnot just on numerical\ndata, but on textual data.\nAnd to be able to combine the two,\nto be able to create this kind\nof a sentiment analysis score\nand be able to understand how it is\nthat certain kinds of\nlanguage yield predictions\nthat will ultimately come to pass\nin financial stock prices, bond prices,\nand other instruments.\nSo large language models now can analyze\nlots of different sources\nof text, including news.\nNews is one of the really key\naspects of financial markets.\nMarkets are always reacting\nto current information,\nand the information does not\nnecessarily have to be accurate\nfor financial markets to react.\nThey will react to rumor in many cases\njust as quickly as they'll react to\nsubstantive true information.\nAnd so large language models can combine\nthe kind of information\nacross various different news sources\nand distill it into a\nsingle prediction, which is,\nwill the asset go up or\ndown in price tomorrow?\nI believe that more sophisticated methods\nof prediction are possible,\nbut the challenge is to come\nup with the correct prompts\nto make those predictions,\nand then to be able to deal\nwith the issue of hallucination.\nSo that's one of the\nreasons why hedge funds now\nare experimenting with\nthese large language models\nand why those of you who are\ninterested in having a career\nin financial analysis,\nI would urge you to start\nplaying around with LLMs\nfor exactly this purpose.\nWhat regulatory and\ncompliance considerations\nshould be addressed when\ndeploying LLMs in this field?\nWell, I mentioned that\nthere's an arms race going on\nbetween the regulators and the fraudsters,\nand I think that that's\nsomething that we really need\nto consider over the course\nof the next few years\nas the pace of innovation\ngets faster and faster\non the side of the practitioners.\nWe need to give regulators the tools\nto be able to fight this kind of a battle.\nAnd unless we pass\nlegislation to help them,\nI think it's gonna be\na very one-sided race.\nLet me give you an example\nof one piece of legislation\nthat I think we have to consider.\nData is an incredibly powerful currency\nin this business.\nAnd so who controls the data?\nIf I'm a consumer\nand I put my data at the disposal\nof a particular vendor that's\nproviding a service for me,\ndoes that vendor have\nthe right to use that\nfor any purpose whatsoever,\nincluding purposes that\nare detrimental to me?\nSo this legislation that\nneeds to be formulated\nas to who has the rights to the data,\nand if vendors are making\nuse of customer data,\nwhat can and what can they\nnot do with that data?\nWe need to pass legislation\nto make this clear\nso that we can allow the\ndata to be used in broader\nand more effective ways,\nwhile at the same time\nprotecting the interests\nof those who need protection.\nSo I think that's the biggest\nset of issues right now,\nis that the regulators\ndon't have all of the tools\nthat they need and they don't\nhave the budget that they need\nto be able to deal with\nthese kind of issues.\nWe need to to bite the bullet\nand make an investment in\nour regulatory infrastructure\nto deal with this brave new\nworld that we're entering.",
  "transcript_chars": 17513,
  "ingested_at": "2026-05-15T10:56:39.268338+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 121248,
    "like_count": 2802,
    "channel_id": "UCBpxspUNl1Th33XbugiHJzw",
    "categories": [
      "Science & Technology"
    ],
    "tags": [
      "MIT",
      "Massachusetts Institute of Technology",
      "CSAIL"
    ]
  }
}