Can AI Improve Your Finances?

Like much of the world, Wall Street is psyched about AI, specifically using its deep learning to help sell “advanced” advice and financial products. According to Investopedia, automated investment platforms are using AI to help build customized investment portfolios and recommendations. And ChatGPT reports that AI-based chatbots are being used by major Wall Street firms to provide budgeting, saving, and investment advice. “Customized” advice would seemingly include:

Hold 85 percent of your wealth in stocks.

Buy long-, not short-term bonds.

Put 15 percent of your assets in gold, silver, bitcoin, and alternative investments.

Take Social Security at 62 and buy a deferred annuity.

Take out a reverse mortgage.

A Disclaimer

As you’ll read, I’m not a fan of using AI to provide financial advice. I think it’s far too crude and is almost surely being trained to provide improper recommendations. This said, I’m not an impartial observer. As most readers know, my software company markets two precision, economics-based financial planning tools — MaxiFi Planner and Maximize My Social Security (MMSS).

MaxiFi covers all aspects of lifetime planning, including safely maximizing lifetime discretionary spending and assessing alternative portfolios in the manner prescribed by economics — lifetime expected utility maximization. MMSS focuses on maximizing lifetime Social Security benefits.

Tens of thousands of DYI households and financial advisors have used these super user-friendly and continually improving tools for decades. Given the hype over AI, many clients have asked what AI can do that our tools cannot. Here’s my admittedly biased answer: Make huge mistakes.

What Exactly Is AI ?

AI’s goal, according to ChatGPT, is “to create systems that can perform tasks that would typically require human intelligence.” Based on this description, an abacus, invented in Sumeria around 2700 B.C.E., qualifies as AI.

Hence, if your advisor (if you’re using one) claims they’re using AI to provide you with financial advice, ask them if they’re using an abacus. When they say “You must be joking.” ask them to explain precisely how their AI tool works. They will surely have no clue. That, by itself, should trigger alarm bells.

Neural Networks

In economics, AI references the use of algorithms comprising neural networks to make predictions. This includes curve fitting, which is a form of prediction. Economists in the predicting business have used neural nets for decades, but more so of late thanks to the availability of big data. Big data permits calibrating vast numbers of parameters whose inclusion can improve prediction.

Neural networks relate economic and other types of outcomes/decisions to complex mathematical functions of complex mathematical functions of complex mathematical functions of …, ultimately, underlying data.

Each set of functions of functions in the network’s architecture is called a layer. Adding more layers deepens the network, hence the term deep learning. The figure below illustrates a two-layer neural network, gn, of the data -- the x’s. The terms s1 through sd are complex mathematical functions of the data and the terms pd1 through pdK are mathematical functions of the s’s.

Each function includes parameters (coefficients) that govern the degree to which the function likely influences the behavior of interest. As indicated by the dotted lines, positing just a two-layer network produces a large web of interconnected impacts. Thus, the coefficients in the functions of pd,1 will impact how those functions alter the importance of the s1s2, s3, and s4 functions, which, in turn, alter how the functions process, for predictive purposes, the underlying data — the xs.

Black Boxes

Let’s say this network’s goal is to predict whether you’ll do better, by some measure, investing 80-20 in stocks versus 20-80 in bonds. Further suppose your advisor, following the network’s advice, puts you 80-20 in stocks and bonds. Finally assume the market tanks and you lose your shirt.

You’ll naturally and angrily ask your advisor to explain their advice. Given the system’s overwhelming complexity, one can’t trace a given prediction to a given factor. Hence, your advisor will have but one answer: “I was following the algorithm’s orders. What did it get wrong? No clue. And no one in the company knows either. The network is too complex to disentangle. But our black box is incredibly powerful.”

This sounds like a lawsuit in the making. This is particularly true given that the complex functions are non-linear. Extrapolating non-linear functions beyond the data on which their coefficients are based can lead to weird predictions. This, intuitively, is why neural networks often hallucinate without any clear explanation.

An example is drawing a picture of knights with no eyes and missing hands. Concern about what AI gets wrong led Elon Musk and more than 1000 other tech leaders, last year, to publicly call for a pause in AI development stating that AI tools represent “profound risks to society and humanity.”

What about risks to your portfolio? Can AI, even if properly trained, robo invest you in something with extreme risk and destroy your lifetime savings overnight? Sure sounds like it.

Training a Neural Network to Produce the Wrong Answers

A different and far greater concern with Wall Street’s use of neural nets is training them to give bad advice. Conventional planning dominates the advice industry. As I’ve previously described, its 70-year-old, sales-oriented, methodology advises clients that investing in risky assets is safe provided the probability of remaining financially solvent, while spending at one’s desired target, exceeds 80 percent.

The practice starts with a seemingly benign question: How much to do what to spend in retirement? Since there is no apparent cost to providing a high number, the “planning” begins with a target that’s likely unaffordable. A few questions and calculations later, bad news arrives — “You can’t meet “your” retirement-spending target given how you are investing. But if you invest in our high-yield funds you’ll have an 80 percent plus chance of spending what you want without running out of assets.” Within minutes you’ve signed up for a portfolio with far greater risk than you want or understand.

Whatever Wall Street says and its regulatory body, FINRA — the Financial Industry Regulatory Authority — condones, leaving a household with an up to 20 percent chance of becoming financially destitute — something that could occur early in retirement — is the opposite of safe. As any decent economist will tell you, it’s not only extremely risky. It violates a reasonable fiduciary standard — something FINRA is supposed to enforce!

Since Wall Street’s neural networks are surely being trained (its parameters are surely being tweaked) to provide conventional advice, they are being trained to, in my view, provide awful advice.

Financial Advice Is Either Precisely Correct or Wrong

The highest quality Swiss watches contain over 400 parts, take a year to produce, and cost tens of thousands of dollars. Their purchasers seek to know the time, not predict the time. The answers in personal finance require equal precision. E.g., either a Roth conversion of $X for Y years lowers lifetime taxes and raises affordable lifetime spending or it doesn’t. This answer requires calculating taxes and all interdependent measures, including internally consistent paths of spending, assets, and taxable asset income, correctly (within the dollar) for each future year, not making refined guesses.

Unless Wall Street’s neural networks incorporate all the details of our myriad fiscal policies, including Social Security with its 2728 primary rules, the clock, so to speak, won’t run on time. But neural nets simply aren’t cut out to make precise tax, Social Security, or other such calculations. Neural nets are akin, in this context, to using a sledge hammer to access the workings of a Rolex.

Deterministic Versus Stochastic Financial Planning

Wall Street’s focus is, to be clear, on investing your savings and charging you for the privilege. It’s not helping you answer what are often far more critical decisions. E.g., when can I afford to retire. How much should I be saving for retirement? Does getting an MBA make financial sense? Should I leave California and take a lower paying job in Tennessee (which has no income tax) and a terrific pension plan and healthcare plan?

These are precise questions requiring precise answers. To repeat, they aren’t questions neural nets are remotely prepared to handle. If you are trying to add 2 plus 3, you use arithmetic. You don’t use an algorithm that gets “close” when the answer matters.

Such questions are deterministic. If I make conservative assumptions, including how well my assets will perform, what’s the story? Take the last question — should I move to Tennessee? The answer depends not just on the specifics of the two jobs, but also on the worker’s age, marital status, spouse’s earnings, both spouse’s Social Security year-by-by earnings histories, presence and ages of children, regular and retirement account (Roth and non-Roth) balances, the couple’s special expenses, such as alimony payments, whether they own or rent and, if they rent, the terms of their mortgage, and the list goes on. MaxiFi incorporates all these factors and answers the question in a half second. That’s real AI.

As for investment guidance, MaxiFi adheres to economics’ lifetime expected utility maximization. This too involves precise measurement. It entails precisely calculating, not predicting, a household’s range of future living-standard trajectories arising under alternative investment strategies. Given these living-standard Monte Carlo trajectories, economists determine which investing strategy delivers, on average, the highest lifetime utility (welfare).

Lifetime utility is described in mathematical terms. It takes account of how a household’s living standard evolves through time as well as the household’s tolerance for risk — its willingness to experience a lower living standard in order to have a greater chance of a far higher living standard.

Every top economics and finance department in the country teaches its graduate students expected lifetime utility maximization in conveying optimal portfolio choice. The framework is the bread and butter of the field of finance and underlies multiple Economics Nobel Prizes. It’s not something neural networks are built to measure. In contrast, MaxiFi does precision expected lifetime utility maximization inside a minute.

My admittedly biased question to Wall Street is this. Rather than use neural nets to provide patently wrong financial answers, why not use MaxiFi to produce precisely correct answers?

Related: The Real Gold Rush: Investments Beyond Bitcoin