In 1968, Garrett Hardin wrote an influential paper named “The Tragedy of the Commons”. It was published in Science and was kind of a big deal. The paper was about the problem of common resources that can be rationally over-used by individuals such that the resources are destroyed – making everyone worse off. The general idea is a great insight and we can see examples everywhere. People being loud in public spaces, odiferous foods microwaved in the breakroom, spoiled food that remains in the refrigerator. These are relatively innocuous. But more substantial parallels include dumping waste into common water sources, overfishing/hunting, or over-harvesting trees or scavenging in common forests.
The Original Problem
Harden was specifically concerned with human overpopulation. He wrote in the same vein as Paul Ehrlich, who published “The Population Bomb” in the same year. Harden explains that the universe, and earth specifically, only has so much matter and carrying capacity for humans. He left the limiting constraint open. Maybe there would be inadequate clean air, water, nutritious food, etc. The specific inadequacy doesn’t matter, they are all due to the finiteness of our world.
From Hardin’s perspective, there was no natural prophylactic that would prevent us from bumping into the natural and horrifying constraint that inadequate resources would impose. He thought that people would continue to reproduce at greater than replacement rates and death would ensue. He also astutely pointed out that if good people agree to limit their number of offspring, then we will be left with only bad people reproducing. They’d place everyone at risk with their irresponsible, though entirely rational, folly. He concludes that the only hope for humanity is a mutual consent to a coercive entity that would prevent this commons problem of over-reproduction.
The Typical Response and its Shortcoming
The most popular argument against Hardin comes from Julian Simon, who argued that humans are the source of innovation and ingenuity and that more humans implies more innovation. Against Paul Ehrlich, Simon famously bet that commodity prices would fall over the course of the 1980s, asserting that human ingenuity would increase abundance.
Putting the wager aside, Simon’s idea that humans figure out ways to make resources more accessible is entirely correct – except for when it isn’t. The Malthusian booms and busts in population and persistent subsistent income levels did occur historically. But it’s been a long time since then. For a few hundred years we’ve had increasing standards of living *and* an increasing population. There are exceptions. There’s been starvation in Venezuela, Cuba, Ethiopia, China, North Korea, and Ukraine. These countries were plenty populous with plenty of human minds to tackle the challenges of scarcity. More people = more abundance isn’t exactly what Julian Simon said, but it’s how a lot of people answer Hardin and Ehrlich’s predictions.
That turned out to be the commonality among the four papers I discussed at the Association for Public Policy Analysis and Management last weekend:
In “Subsidizing Data Centers”, Vijay Prakash found that US states subsidizing data centers saw no significant difference in unemployment (or electric prices) compared to those that didn’t.
2. In “Spillover Effects of Million Dollar Projects on Wage Distribution and Skill Demand“, Alaa Abdelfattah (with Layla O’Kane) found that big new business projects (mean investment $333 million, mean subsidy $130 million) led to no increase in total job postings in the counties that ‘won’ the project relative to counties that were reportedly the ‘runner-up’.
4. Will the jobs remain once the subsidies are gone? They mostly didn’t in Puerto Rico. In “Corporate Tax Incentives Retraction: Evidence from Puerto Rico”, Dayanara M. Diaz Vargas found that the end of Section 936 tax credits in 2006 led to a 28% decline in manufacturing employment.
The results reminded me of subsidies for sports stadiums, which always promise to bring jobs and economic activity but rarely deliver:
The latest CPI-U data suggests inflation was around 4 percent in the past year (I say “around” because due to the late 2025 government shutdown, we have no October 2025 baseline data to calculate the rate). The persistently high inflation — even using the core CPI, it is still close to 3 percent — and currently high interest has made some investors consider the safer, more guaranteed returns of good old bonds.
One type of bond you may want to consider: Series I Bonds. These bonds issued by the federal government pay a rate that is set for 6 months, and the rate is based on the most recent 6 months of inflation as measured by the CPI-U (plus a fixed rate, more on that in a moment). These bonds are currently paying 5.51 percent, since the most 6 months of CPI-U inflation was 2 percent, which you essentially double to get an annual rate, and then a fixed rate (currently 1.4 percent) is added to that inflation rate. When you buy the bond, the fixed portion of the rate is fixed for the life of the bond (30 years). That 5.51 percent is currently higher than the 10-year Treasury bond, and even slightly better than the 30-year bond! But the rate is only guaranteed for 6 months, after which it will change.
The biggest reason to buy these now is not the headline rate but the fixed rate of 1.4 percent. That’s the best fixed rate they have offered since 2006-2007! I purchased some of these bonds back in 2022, and I still think it was a good choice: they were paying over 7 percent at the time, which increased to 9.62 percent the following 6 months. Not a bad return in a year when the stock market was unstable and actually fell almost 20 percent that full year. But when I purchased the I-Bonds in 2022, there was no fixed rate attached. That means the 2022 bonds I am still holding are currently just 4 percent, rather than the headline 5.51 percent.
Even if you believe that CPI-U inflation will eventually get back down to 2 percent, and stay there into the future, I-Bonds purchased now will pay almost 3.5 percent for the life of the bond — not a stellar return, but not bad for a guaranteed return. Of course, you could always lock in almost 5 percent for 10 years or over 5 percent for 30 years with ordinary Treasury Bonds. Those aren’t bad ideas other. But if you think inflation will remain elevated, or especially if you think it will ramp up again, putting some of your cash in I-Bonds might make sense.
As a historical point of reference, the best time to buy these bonds would have been in the late 1990s, when they were first issued. At the time they paid a fixed rate of over 3 percent! That means today those bonds are paying 7.5 percent and had been paying as high as 13 percent in 2022 at the peak of CPI-U inflation. Not bad for a guaranteed return! The cumulative return on a $10,000 investment in 1998 would be about 5x, today worth over $50,000. That’s only about half the return of the S&P 500 — but with much less volatility. Those high fixed rates are unlikely to return soon, but the 1.4 percent fixed rate is not too shabby.
Important final thoughts: don’t take this as professional investment advice, just as an idea to research more on your own and with your financial advisor. You can only put $10,000 per year into Series I Bonds, so no need to think about whether you should move all of your money into I Bonds — just think about the margin dollar you are investing. And if you sell these bonds within 5 years, you will lose a small amount of interest (3 months’ worth).
Everyone loves a trilogy, so let’s wrap this up. I told you why agent-based modeling never broke through in economic research. I told you how it could eventually matter. Now I’m going to tell you how AI could, and maybe should, actually make it happen. And as someone who is not fully “AI-pilled”, I will confess that there aren’t that many avenues in which I am particularly excited about AI contributions, particularly in the arts, and even where I see opportunities for great things, I remain pessimistic that Gresham’s Law is going to hit us hard and fast while we are still learning to culturally and institutionally separate the wheat from the slop-riddled chaff.
But let’s put on our more optimistic glasses for the moment and think through what the opportunity for AI to guide agent-based economic modeling into a new (and first) golden age.
First, let’s define a few terms. When I refer to Agentic AI I am specifically considering an LLM model that is capable of autonomously pursuing goals, taking actions, and adapting their behavior in response to new information, all based on a set of pre-defined rules and reward functions and with little to no need for continuing human instruction. When I refer to an Agent-Based Model (ABM) I referring to a computational model of an economic or social system in which individual, heterogeneous agents (consumers, firms, workers, etc) make decisions according to specified behavioral rules, interact with one another and their environment, and collectively generate aggregate economic outcomes.
I don’t want to be wholly redundant with my previous posts, but it’s important to revisit a key theme that ran through both posts: the greatest strength of agent-based modeling is it’s ability to accomodate incredible depth of individual and institutional heterogeneity. Which is a nice way of saying that it is parameter rich. Which is itself a nice way of saying that the models can generate any result without the prospect of credibly investigating the mechanism behind those results and, more importantly, suffiently culling enough noise from them to learn anything about how the world works. For those less familiar with the curse of dimensionality, let me explain.
If a model has 2 parameters you can simply run a grid search of every possible combination of those parameters and observe what kind of outcomes are generated. If you had an agent-based model with 100,000 agents randomly interacting you could still investigate the parameter space by “running” the model 1000 times for each possible combination of parameters and look at the full distribution of outcomes, reporting the mean, variance, and individual quantiles of every outcome of interest for each combination in your grid of parameters. Your results might look like this:
This is a useful way to design and investigate an agent-based model. But you probably already see the problem. Sure, toy models with two parameters are cute, but how do I answer complex economic questions? Toy models can only take us so far. And that’s the rub.
Once you start really start trying to advance economic theory of large, complex economic systems with an agent-based model you’re going to have a lot of visible parameters and lot more less visible parameters. An ABM will be composed of:
Agents (people)
Collectives of agents (firms, clubs, mosques, families, states)
Rules for agent decision-making
Rules for collective decision-making
A context, network, and/or phystical environment with it’s own rules
Every single item on that list will in turn have attributes with both
Initial conditions
Distributional characteristics of those inititial conditions i.e. mean, variance, etc
Are they fixed, exogenously varied, or endogenously varied over time?
Take the first list and add up the total agents, collectives, rules, etc. Then multiply those by each other and by every decision that has to be made in the second list. That number will likley be in the thousands. Congrats, that’s how many true model parameters you have.
With modern macroeconomics, and I am speaking here as a non-macroeconomist, your goal is to reduce that parameter space using basic logic, empirical observation, and modeling decisions that collapse heterogeneity to a small number of representative agents or types. You produce equilibria predictions at each stage of your model and compare them to objective reality. Once the model is sufficiently micro-founded and validated, you can then generate predictions and investigate mechanisms.
The whole point of agent-based modeling, however, is not to collapse the model to a small numbe of representative agents. It’s to endogenously grow aggregate outcomes from the individual endogenous decisions and how, through interactions within the model, they aggregate into predictions for a social system. So what’s the problem? Let’s assume every agent type and collective is characterized by 5 attributes and that each attribute is characterized by 3 statistical properties. Each rule comes with 3 functional variations, characterized by 3 weights each. We’ll then say that there are 5 total agents ( 2 individual types and 3 collectives), for each of whom half of attributes are fixed and half are endogenous. They manner in which they are endogenous has 3 functional variants. If we want to do a true grid search, we’re looking at least at a 5000 combination parameter grid space.
And we are no where near done yet, because we haven’t considered how time works. How steps and sub-steps are ordered. What decisions are synchronous, asynchronous, hieracherially ordered or random. We haven’t considered the social or physical environment What’s you start digging in, your parameter space starts gaining orders of magnitude. And then you realize the real problem is that true grid search of the entire parameter space would take until the heat death of the universe.
And that’s what research expertise really is – competently making decisions about what can and should be varied, what can and should be assumed, etc, until we have a tractable question and well-executed research endeavor to answer it. The ABM space is too big for a researcher, however, because too much of the competent execution of research is unobservable, possibly even to the researcher themselves, as the model and the interactions of each decision become intractably complex. Too big for a human researcher. But what about an Agentic AI?
What if we created a base ABM, handed it to an Agentic AI, and told it to recursively evaluate and update the model until the model and the results it generated met a pre-defined set of necessary meta-conditions that the researchers defined as i) a falsified theoretic prediction, ii) generated via observable, definable mechanisms, iii) dependent on a maximal set of modeler decisions, and iv) and independent of set of unlimited, but fully observable modeler decisions?
This is going to take a long, but not necessarily intractable amount of time because we can also define a series of “golden section” adjacent rules that allow the AI to decide with a parmeter, rule, or decicsion is sufficiently irrelevant to ignore or sufficiently constrained to tractably investigate as key outcome determinant or source of outcome sensitivity.
As long as I am giving unsolicited advice to graduate students everywhere interested in agent based modeling, I would suggest forcing such an Agent AI modeler to exclusively use a constrained, pre-defined modeling language or library. Further, you probably want to force a literal, numerical maximum on the number of lines or characters of code. And, lastly, you’ll want to come up outcome rules by which the Agentic AI has to itemize and defend every single deviation of the final model from the base model handed to it. By constraining AI to a limited variety and quantity of coding objects, you are in turn ensuring that the resulting model and the internal mechanism will be transparent to the researchers managing the AI and the broader discipline evaluating them.
Agentic AI offers researchers the ability to iterate through the prospective solution space for modeling problems otherwise intractable via brute force. Further, it allows researchers transparently reveal the instructions that guided the agentic AI dimensions on which the generated model is sensitive to both varied parameters and modeling decisions. It allow for two hierachical levels of meta-research transparency.
That’s what I would be coding write now if I was a PhD student interested in contributing to economics and other social scientific disciplines via agent-based modeling.
For the past two years, we’ve had a unified government in which the Republican party has populated the White House and had majorities in both chambers of congress. Once the new legislators are in office we’ll have a divided government. That means that the political party controling the White House won’t also control the senate and house of representatives. In the spirit of saying something of substance, I want to know what the implications are for our fiscal picture.
One can measure fiscal health in many ways. Does a divided government bode well for shrinking the deficit? Rather than spit-ball about politics, I take the data since 1960 in one hand and a model in the other. See the end notes for model details.
Since 1960, the most popular form of government split has been a Democratic president and unified government, followed by Republican presidents and a divided government. Democrats have controlled at least one chamber of congress for 62% of the years. Do these splits matter for federal outlays, receipts, and the deficit? Since a unified democratic government was most common, I use that as the baseline fiscal case and compare the cases to it.
Below are the average impacts of government-party types on federal spending, revenues, and the deficit. The blanks are insignificant differences. On average, the only party split that results in less spending is a Democrat president with divided government. Those and unified Republican governments raise more revenue and reduce the budget deficit from its trend (there was a small decline in in discretionary spending at the start of Trump’s 2nd term with a unified government). The bad news for deficit hawks is that Republican presidents with a divided government – the type that we’ll have for the next two years – perform poorly. On average, they reduce government revenue and increase the deficit. Again, all of this is ‘on average’.
I predicted a ‘moderate red wave’ here in 2022. I predict a blue wave for 2026 that is bigger than the red wave of 2022, though not a ‘mega tsunami‘
That post looks good now with the exception of those last five words- it was in fact a blue ‘mega tsunami’ by Kalshi’s definition (Democrats hold 240+ House and 54+ Senate seats).
This comes after a longtime WSJ editorial writer hoped for a big Republican loss, something I’ve never seen from the WSJ before. But if they speak for the markets, it makes sense- US stocks opened up 2% the day after the election.
Crypto and gold prices are also up, an odd combination. Oil prices are down, perhaps seeing an end to the Iran war as more likely. Polymarket shows the odds of a US invasion of Iran this year down 5%. John Ossoff and Ron Desantis saw their odds of becoming President in 2028 rise at the expense of JD Vance and Gavin Newsom.
What will we see change on economic policy? Congress may lower tariffs and take back some of the tariff power they have long delegated to presidents. The Federal Reserve board will feel more free to raise interest rates if they think it is warranted. Many state-level results mean data centers will get harder to build, which will hurt the ‘picks and shovels’ companies literally building them but not necessarily the frontier labs who won’t be able to spend as much on racing ahead. NVDA and Micron stock is down while Google, Meta, Anthropic and OpenAI are up.
But the fact that President Trump still has veto power will limit how many big changes we’ll see over the next two years for anything that would require a new bill. Welcome back to divided government, to gridlock, and to checks and balances.
This is my last post before the midterms Tuesday. What will they bring?
Midterm elections are usually bad for the President’s party:
I don’t think this year will be an exception to the rule. I see many indicators pointing in the same direction.
The largest US prediction market Kalshi currently gives Democrats a 91% chance to take the House and a 66% chance to take the Senate. The international Polymarket and play-money Manifold currently give slightly higher chances, 93/67 and 94/69. Forecaster Nate Silver is higher still at 95/76.
The best odds I’m seeing for Republicans among the models I have followed for years is Ray Fair giving Dems only a 60% chance to take the House (he doesn’t forecast the Senate). In Fair’s model, the economy is working in favor of Republicans; inflation is a negative for them but economic growth is a stronger positive (including the 3.4% Q3 GDP estimate just released today).
In practice though I think economic issues are working in favor of Democrats. The economy is good in many ways, like 4.2% unemployment- but most people aren’t happy about it. Consumer sentiment is at record lows. Economic growth is solid but largely driven by the AI boom, which gives people plenty to be nervous about. Inflation is only a bit high (3.4% vs a 2% target) but easy to tie to administration policies like tariffs, war, and pressuring the Fed to lower interest rates. Ad spending indicates that politicians on both sides think tariffs have gone from a minor issue that voters have mixed feelings about to a major issue that voters don’t like:
Gas prices have long been especially salient, and are currently high because of the Iran war. I said here in March:
if the US administration wants to avoid a midterm wipeout driven by high gas prices, I see only 3 options:
Get lucky and see the Iranian regime fall quickly
Negotiate an end to the war quickly (which might itself be unpopular if they can’t get a good deal) or just declare victory and go home (but its not clear whether Iran would re-open the strait now just because the US stopped bombing)
Restrict Exports
But they took none of those three options, and now it seems the midterm wipeout is coming.
I predicted a ‘moderate red wave’ here in 2022. I predict a blue wave for 2026 that is bigger than the red wave of 2022, though not a ‘mega tsunami‘. The main things Republicans have going for them are the Senate map (Dems need to win in multiple 2024 red states to win) and the possibility that angry anti-incumbent voters will take down some Dems too.
Based on all of the good predictors that I track, GDP growth for the 3rd quarter of 2026 will likely be right around 3.2%:
Table 1
Kalshi, the Atlanta Fed’s GDPNow model, and the WSJ survey of economists are typically the best predictors, and they are all right around 3.2%. I’ve also added a new column to the table, a Nowcast from the St. Louis Fed. This one tends to be a pretty poor predictor, so you might wonder why I’m adding it to my list. Read on below to see why, but for now it’s interesting to note that it is very close to the three other good predictors (the NY Fed model is generally not very good).
When I have tracked these predictors in the past, I have been comparing them to the advanced GDP estimate, the first one the BEA puts out. Being able to predict this release is a good and hard enough task, but ultimately we don’t actually care about the number that will come out tomorrow. This number will be revised many times with more data, not only the 2 months after it, but through subsequent benchmarking and revisions in later years. Ultimately, we care whether these predictors are getting the true number, not BEA’s initial release (even though they are calibrated to predict that initial release, for the most part as I understand them).
What does it look like if we compare those predictions to the current best data on GDP growth going back to 2023? On this measure, the St. Louis Fed Nowcast looks pretty good recently:
Table 2
While the St. Louis Fed Nowcast looks pretty bad over the entire time period — the worst of these 5 predictors — it has done very well in the past 3 quarters, including nailing the 1st quarter of 2026 exactly (though in one quarter, they got the sign wrong). So I’m now tracking it in my table to see if it will continue to perform well — maybe they improved the model? The 3rd quarter data won’t actually tell us much about the St. Louis Fed model relative to the others, since they all are saying the same thing, but it will be worth watching in the future.
If we tally up the best predictors using the final or most recent GDP growth estimate, the best predictor is Kalshi’s betting markets. Kalshi has the lowest average error, which is just slightly better than the WSJ survey of economists. But Kalshi was also the best predictor in 6 of the 14 quarters in Table 2. The WSJ survey was only the best in 3 quarters — and in two of those it had the same prediction as Kalshi. It was only best as a standalone in one quarter (the oldest one in these 14 quarters). The St. Louis Fed was the best predictor in 5 quarters, all of these by itself, but its average error (as well as standard deviation and variance) is the highest of these 5 predictors.
Based on different combinations of the predictors, the best pair of predictors is Kalshi and WSJ. Averaging those together gets you an average error of 0.68, better than any single predictor. That means my prediction based on all this information is 3.15% growth in the third quarter for the final figure, once we eventually have it. The average of the Atlanta Fed and Kalshi is the best pair for the advanced figure, which gives a slightly different prediction of 3.24%.
Let’s talk about disability. The American Community Survey (ACS) identifies seven broad categories, not including the baseline case of ‘No Disability’. I’ve written before about how disability designations are difficult because many qualifying conditions are a matter of degree. Surely you’ve met a person who is legally blind but can see *some*. How much sight ability can remain while still being considered blind? If a person can identify the sun’s location, but nothing else, then is that person still blind? Similarly, who’s to say whether some physical or cognitive characteristic is disabling? Why not enabling – depending on the task?
Wisely, the ACS side-steps this entire discussion and simply asks respondents whether they have “difficulty”. We can be agnostic about the cause and degree while still trying to get at the heart of the matter. One downside is that people tend to have more difficulty as they age. Since that’s not relevant for earlier life economic outcomes, I only examine ages 24-45. Difficulty with “self care” is too small of a category in the 2024 ACS for the analysis here. Below are the other six difficulties identified:
We should say early on that many of the people who participate in “Deaf” (capitalized) culture do not consider deafness to be a disability and that any such suggestion is motivated by bias and hearing culture. I won’t engage with that here. But I can say that there is a long tradition going back through the 1800s that rejected alms-giving for deafness and instead supported independent living for deaf people. Historically, this manifested as outlawing panhandling while featuring deafness and rejecting government programs that provide simple transfers due merely to deafness. More recently, the push to reject the ‘disability’ label and its baggage has even been the motivation for some deaf people to oppose cochlear implants being covered by medical insurance. For ease, I’ll be using the colloquial terms for all of the above difficulties. Now that you have the greater context, let’s jump into the data (unweighted).
The baseline income chart is above. Given the above description, it should come as no surprise that deaf people have the highest average income among disability groups, conditional on employment. However, it’s also surprising that most of the above disabilities have incomes that I wouldn’t exactly call destitute. Many people who have a disability look much like everyone else. From an economic point of view, we all have a comparative advantage and it’s perfectly consistent with gainful employment and flourishing otherwise.
Certificate of Need laws require healthcare providers to obtain government approval before building new facilities or expanding services. The original intent of the laws was that less building would mean less spending, but critics worry that it means less competition and higher prices leading to higher spending. Our paper finds that the critics are right, as states repealing their Certificate of Need laws see their healthcare spending fall relative to those that maintain them.
Longtime readers may recall that back in 2022 I posted about another paper of mine with a similar result, and even that wasn’t the first word on the subject. So what’s new in this paper? Besides newer data, we measure health spending as a percentage of the state economy rather than in raw dollars, to try to account for how CON changes the true burden of healthcare spending. We also use newer regression techniques that are probably more accurate but definitely help make prettier graphs.
Here’s what we found happening to health spending when states repeal Certificate of Need:
Full published paper is here, ungated working paper version here.
