Lots of people are turning to LLMs for financial advice. But what does this advice actually look like?
@TahaChoukhmane, @wdwlin, @AkuzawaMatthew, and I have a new paper that develops and implements a new methodology to describe this advice and why it varies!
Thread ⬇️⬇️⬇️
The oddest intellectual position I have encountered over the last decade is the refusal of so many to believe that the current version of Pangram is incredibly good at avoiding false positives (i.e., flagging a human-written text of some length as AI-generated). It is not terribly hard to trick Pangram into false negatives (i.e., classifying an AI-generated text of some length as human-written) with “humanizing” tools and some careful editing. But the other way around? Nearly impossible.
For two weeks, I have been running experiments in many different situations for a new paper I will circulate tomorrow (and because, as an editor, I care a lot about the answer). Not in a single case did a human-written text of reasonable length score above 10% for AI generation. And, yes, many the texts I used (including mine!) are from ESL writers.
Frankly, the counterarguments I have read over the last week here on X (and the papers linked to “back them up”) are childish in the extreme.
If you think that false positives at Pangram are, as of November 16, 2026, a concern, you are lying to yourself.
The more home ownership or marriage seem out of reach for young adults, the more they turn to financial gambles such as crypto currencies.
nber.org/papers/w36010
What monetary policy maximizes efficiency—the joint surplus of all households—in heterogeneous-agent New Keynesian economies (HANK)? In a canonical HANK model, the answer is the same policy as in RANK. Read on for why (1/9) nber.org/papers/w35967
Economics has incorporated insights on behavioral consumers (or individuals) for some time; it has been slower in thinking about "behavioral firms," but that has been changing too of late, partly as imperfect competition has become more of an acceptable benchmark in product and labor markets.
Market power mitigates size of losses in profit from optimization error, so more easily allows a lot of "clunky" behavior to survive in a competitive equilibrium. [Examples from labor market: aeaweb.org/articles?id=10.12…; nber.org/system/files/workin…]
1/ I agree the axiom is innocuous. Completeness and transitivity on their own say very little. That is the point: the ideology isn’t in the axiom. It’s in what the axiom is then asked to carry. Three moves usually follow it, and none of them is innocuous. 🧵
The assumption is that people have a preference ordering. They can rank different consumption bundles. A is preferred to B, or B is preferred to A, for example. People are assumed to choose their most preferred feasible consumption bundle. This sounds pretty innocuous to me.
Apropos systemic vulnerabilities to breakdowns at the micro level:
@ChadJonesEcon & @ChrisTonetti have a new (early-stage) WP "The Weak-Link Asymmetry: AI’s Peril and Promise" that is rather topical.
The paper highlights that if tasks are gross complements -- aka "weak links" / Kremer's O-ring -- not only do the gains from automation arrive slowly, but the economy is also very vulnerable to shocks at the task level. With an elasticity of substitution of 0.2, boosting productivity 10x in tasks worth 1% of GDP raises output by <1%, whereas cutting their productivity 10x drops output by 68% on impact. The paper applies this to AI -- motivating by the risk of AI-enabled attacks on the financial system, electricity grid, and other critical infrastructure. But the FAA ground stop is a pretty vivid illustration of the underlying fragility...
🔗 christophertonetti.com/files…
A point I emphasize again and again is that the relation between income and fertility is changing quickly, both within and between countries.
To make this point between countries, I take the total fertility rate (TFR) of up to 158 countries from the UN World Population Prospects and their GDP per capita from the Maddison Project, measured at purchasing power parity and in constant 2011 prices (that is, the data already correct for differences in price levels across countries and for inflation). I am not a big fan of the UN World Population Prospects, but this way nobody can accuse me of being selective with my sources.
Then I plot TFR against GDP per capita for 1950 (the start of the sample), 1985 (the midpoint of the sample), and 2022 (the last year in Maddison’s dataset).
The left panel includes all countries except the six Gulf oil exporters (Saudi Arabia, Kuwait, the United Arab Emirates, Qatar, Bahrain, and Oman), whose wealth comes from oil rents, not from their level of economic development. In the right panel, I also remove sub-Saharan Africa, a region with very poor data on TFR and GDP per capita.
Start with 1950 and 1985. Plot the two curves, and you will have trouble telling them apart. Countries with $2,000 of GDP per capita had 6.1 children per woman in 1950 and 5.9 in 1985. Countries with $5,000 had 4.3 in 1950 and 4.6 in 1985. Countries with $10,000 had 2.8 in both years. Countries got richer and had fewer children, but they moved down the same curve. In 1985, as in 1950, fertility dropped fastest between $5,000 and $10,000, and the typical country reached replacement (about 2.1 children per woman) only at around $14,000. Mexico, with $10,000 of GDP per capita, had 4.1 children per woman. Argentina, with $11,000, had 3.1.
And the relationship between TFR and GDP per capita held despite many other changes. For example, child mortality collapsed at every income level, literacy and life expectancy rose, again at every income level, modern contraceptive methods spread widely, and international organizations and local governments sponsored aggressive family-planning programs across Asia and Latin America from the 1960s. Nothing seemed to matter.
After 1985, the slope of the curve became much smaller (in absolute value) as the left end of the curve started falling quite fast while the right end of the curve barely changed. Interestingly, most of the fall came between 1985 and 1995, well before cell phones and social media (everyone’s favorite explanation) reached anyone outside a small elite of wealthy and highly educated persons.
At $2,000 of GDP per capita, fertility went from 5.9 children per woman in 1985 to 4.0 in 2022, a drop of 1.9 children. But at $20,000 of GDP per capita, fertility was 1.7 children per woman in 1985 and the same 1.7 in 2022.
Similarly, countries with 2.1 children per woman (roughly the replacement rate) were at $14,000 in 1985 but at $11,700 in 2022
Take Sri Lanka. Sri Lanka’s GDP per capita is $11,200, roughly the income at which the 2022 curve crosses replacement. According to the UN World Population Prospects, Sri Lanka has 2.0 children per woman in 2022, while the United States, with five times Sri Lanka’s income, has 1.7. The United States last had Sri Lanka’s income in 1923, when American women had about three children. And according to the most recent data from their national statistical agencies, Sri Lanka is in 2025 at 1.37 children per woman and the United States at 1.57. Sri Lanka has undershot American fertility by a considerable margin (yes, 1.37 is much lower than 1.57) a century before reaching American income.
I also decomposed the change in average TFR across countries into two parts. The first is the change from countries getting richer while moving along a fixed curve. The second is the change from the curve itself shifting, holding each country’s income fixed.
From 1950 to 1985, almost all of the fall in TFR comes from countries getting richer. Remember: the curve barely moved.
From 1985 to 2022, about 40% comes from countries getting richer and about 60% from the curve shifting down. That is the big surprise.
Is sub-Saharan Africa driving this result? Yes and no. In the right panel, we can see that the 1950 and 1985 curves barely change when we drop sub-Saharan Africa. Interestingly, the 2022 curve becomes almost flat. Outside sub-Saharan Africa, 92 of the 108 countries with more than $2,500 of GDP per capita have a TFR below 2.5. Most of the income gradient left in 2022 is an African gradient.
Why did the relation between TFR and GDP per capita hold for 35 years (in fact, for at least the early 1800s) and then fall? I do not have a proof, but my conjecture is that low fertility is driven not by income itself but by “modernity.” As I have explained in many posts, by “modernity” I mean the combination of phenomena such as large bureaucratic organizations, mass education, credentials, formal labor markets, and mass media. Until the 1980s, these features of modern life arrived together with economic development and high income. After 1990, the “modernity package” began to reach countries well ahead of income: secondary schooling for girls, television, and integration into global supply chains. Income stopped being a good proxy for “modernity,” so the curve drawn against income fell.
We are only beginning to glimpse the future of “modernity” without economic development. Mexico or Sri Lanka now have the fertility of Japan or Germany, but without the incomes of rich countries, without their state capability, and without their capacity to attract immigration. Becoming rich is no longer a precondition for having few children. Instead, for most of the planet, getting old before getting rich is the new normal.
The United States has a total fertility rate of 1.57 children per woman. Many people ask why it is so low. That is the wrong question. The right question is why it is so high. At American income levels, the typical country has about 1.34. In fact, the United States has always had exceptionally high fertility given its income.
Let me explain. You can decompose a country’s fertility rate into three components: an income curve, a country-specific component, and a timing component. The income curve shows the fertility rate of the typical country at a given GDP per capita. That is, we look at all countries and calculate the average fertility rate for a country with, say, $10,000 in income per capita (in purchasing power parity, to control for differences in price levels). The country-specific component shows how much a country sits above or below that curve because of its own features: religiosity, family structure, social values, etc. The timing component captures the booms and busts that move births from one year to another (e.g., an economic boom, a depression, or a war).
To build the income curve, I use the 1985 cross-section of countries up to $19,500 in GDP per capita and the 2022 cross-section above that level (both in 2011 dollars, that is, controlling for inflation). The two curves meet almost perfectly at $19,500: the same fertility rate and the same slope. I drop the six Gulf oil exporters because their wealth comes from oil rents, not from economic development.
The curve uses no time-series data for the United States. It comes from a cross-section of 157 countries in 1985 and 2022, with the United States as just one more point. I use 1985 data because, last Monday, I showed that the relationship between fertility and income across countries, which had barely changed for many decades, shifted after 1985. I use 2022 data for very rich countries because, in 1985, there were almost none, making it impossible to accurately predict a “typical” fertility rate.
For the United States, I use two series. The first is the period total fertility rate (TFR) from 1820 to 2022 (for white women before 1933, from Michael Haines, and for all women after 1933, from the Human Fertility Database). The second is completed fertility, the number of children women in each birth cohort actually had over their lives, for cohorts born between 1828 and 1974 (for married women before 1918, from Jones and Tertilt, and for all women after 1918, from the Human Fertility Database). I place each cohort at the GDP per capita in the year the cohort turned 28.
In the figure, you will see that the cloud of points for the United States is nearly always above the income curve. So I multiply the income curve by a single number, 1.22, to best fit the US data. With that one number, the curve predicts 6.8 children per woman at the US income of 1820, when white American women had 6.7, and 1.68 at the US income of 2022, when American women had 1.67. Two centuries apart, and the same curve, scaled by the same number, gets both ends right. The United States has followed the income curve remarkably well for over two centuries, but it has stayed 22% above it!
That is, if I tell you that the average country on Earth was doing in terms of fertility in 1985 at $10,000 (none of which was the United States), and you multiply that number by 1.22, you get the United States’ fertility when US income per capita was $10,000. In 1985, the average country with $10,000 of GDP per capita had 2.82 children per woman. Multiply by 1.22, and you get 3.44. In 1910, when US GDP per capita was $9,916, white American women had 3.42 children. You got to love econometrics!
What do I learn from this?
First, nothing about the United States’ fertility is exceptional in terms of its movement along the income curve. The United States has followed the fertility pattern it was supposed to follow.
Second, what is exceptional about the United States is that American women have had about 22 percent more children than women in the typical country at the same income, and they have had them since at least 1820. My guess is that the reasons are cheap land (first the frontier, now suburbs and housing far cheaper than in Europe or East Asia), relatively high religiosity, and an economy where having a third child does not end a career. Given the United States’ income, its fertility rate should be around 1.34, not 1.57. The extra 0.23 is the United States’ country-specific factor.
Now the timing component. In the TFR, the Great Depression puts the United States at 2.0 children per woman below the scaled curve in 1933, and the baby boom puts the country 1.5 children above in 1959. But a lot of this is just timing. In completed fertility, both gaps shrink: women born in 1908, who had their children during the Depression, sit 0.9 below the scaled curve, and women born in 1934, the mothers of the baby boom, sit 1.1 above. Much of what looked like a collapse of fertility in the 1930s and an explosion in the 1950s was American women postponing births and then catching up (with some under- and overshooting).
So, if you think that either the Great Depression or the baby boom somehow contradicts the downward-sloping income curve, you are confusing a high-frequency movement with a low-frequency one.
In summary: the United States is not exceptional in how its fertility responds to income. The United States has followed the same income curve as every other country, with a higher level. The big swings everyone remembers (the Depression and the baby boom) were short-run deviations that one should expect given historical events.
The @nytimes sketches out a type of "bank run" as people in their 60s are signing up "too early" for Social Security payments as they are thinking that the system will be insolvent soon.
nytimes.com/2026/10/31/well/…
What determines markups in production networks? If firms compete in markups: market shares and 4 network statistics, as US data confirm. Markup responses also dampen TFP gains.
New paper by Wu
restud.com/competition-in-ma…#REStud#EconX#EconTwitter
Can a chatbot do what 50 years of lectures by financial economists couldn't, and persuade people to buy index funds?
A new job market paper by my NYU Stern student Valerie Baldinger finds that the answer is yes
valeriebaldinger.github.io/p…
🧵
One ongoing frustration is the lack of effort I see being put towards near-term, achievable crash projects to do good with AI at a national scale, whether those efforts are by the AI labs or by policymakers & NGOs.
Some obvious ideas to explore suggested by the data we have so far: universal tutoring to support classroom learning, second opinions on healthcare, financial literacy/scam avoidance, and making government services easy to access.
These are hard problems, and I know because over the past 20 years I have worked extensively on building tutoring and financial literacy programs, which were always limited by existing technology. Now, real gains are likely achievable with AI, which should be a big deal. Yes, you would want to do careful design and studies before rolling things out. And yes, those efforts might fail. But, if done right, national programs could lead to big gains in public good for seemingly intractable problems. And they don't require AIs beyond current models to pull off.
My latest article is an exploration into a question I get all the time:
Why are poor people fatter than rich people?
Go check it out: cremieux.xyz/p/obesity-and-p…
Since most of my followers are finance and macro, I’m highlighting the JEP symposium on private credit (in another thread) and, for empirical work, a crystal clear guide to the Bayesian bootstrap. Imma bootstrapping everything now.
aeaweb.org/atypon.php?return…
Don't get too confident about prediction markets! MIT finance department PhD job-market candidate working on prediction market shows miscalibration. The long-shot bias exists and exploiting has a historical Sharpe ratio in the neighborhood of 2!
jiagengliu.com/