Record Highs, Real Lows, Hidden Gaps

Lower-FICO households. As FICO scores fall, four measures all rise steadily: estimated risk aversion, risk premia paid, moral-hazard costs and retained exposure. The authors call this a paradox. The households least able to absorb a disaster loss hold contracts that leave them most exposed to one. They stress that this is a correlation, not a causal finding, and it may reflect factors other than financial constraint.

High tail-risk areas. Places most exposed to severe disaster losses are also where homeowners keep more of the risk.

Weaker insurers. Insurers with lower risk-based capital ratios cover riskier properties and collect higher dollar risk premia. Measured as a share of the total premium, though, the risk premium is roughly the same across insurers. The authors read this as weaker insurers being matched to riskier homes and different policyholders, not charging higher margins.

Checks on the estimate

Moral hazard can't be observed directly, so the authors tested whether their estimate behaves the way theory says it should. Estimated moral-hazard costs are lower when homeowners have more equity (lower loan-to-value ratios). They are also lower in states that require property inspections or verification. The results hold when flood exposure is removed, when non-disaster claims are accounted for, and when the model adds insurer competition and lender insurance mandates.

A final post will model what would happen if insurers had to offer full coverage. The open question is that the gaps that expose households are the same gaps that give them a reason to protect their homes.

What comes next

McBride points to a past pattern: after a sharp run-up, real prices usually take several years to reach new highs. He has written about this before as "7 years in purgatory," but says there is "nothing magic" about the number. It just made a good headline.

His forecast for 2026 was "mostly flat to a small decline in prices nationally." He cites two reasons. Inventory is rising a little, and higher mortgage rates are reducing demand. As a result, he expects nominal prices to end the year roughly flat. If prices stay flat while consumer prices keep rising, real prices fall further.

The full post also includes charts of the price-to-rent ratio, an affordability index, and five-year real returns on the National index. Those sections are behind the paywall and are not covered here.

• The highest-poverty districts spent $17,208 per pupil and the lowest-poverty districts spent $15,921, so high-poverty districts spent $1,287 more nationally (excluding New York City).
• If every district spent its own state's average, high-poverty districts would have spent $1,084 less, because they tend to be in states that spend less.
• Within states, high-poverty districts spent $2,371 more than low-poverty ones.

At the national level, spending by poverty or racial makeup forms a U shape rather than a straight line, with the lowest quintile below the top. New York City distorts the top of the range. It moved from spending below New York's state average to well above it, in a state that spends nearly double the national average.

The Conversable Economist post concludes that the claim "poorer districts get less" is generally wrong within states. Between states it is more complicated, because staffing a school costs more in higher-income states. The author argues that studies finding fewer resources for poor schools often use idiosyncratic cost-of-living adjustments and leave out federal funding, which is weighted toward low-income students.

The report does not say whether any group of districts has enough money, or whether more money would improve student outcomes. It only describes who spends what.

Data centers also pay property, sales and corporate taxes plus fees. Loudoun County, Virginia, the world's biggest data center hub, has used that money for top schools, libraries and roads. Smith points out that governments can raise taxes after a facility is built, when moving is costly. They can also negotiate "community benefit" agreements, which Jared Bernstein wants made stronger and enforceable. Smith grants that grid strain and noise are real costs.

Demand and consumers

Smith's main argument is about the whole economy. Tariff whiplash, deportations and erratic policy drove policy uncertainty very high, which normally depresses business investment. Data center construction worked as a positive demand shock that offset it. A Bridgewater chart shows the buildout accounts for a large share of GDP growth in Trump's second term. Smith admits crowd-out is real: the boom absorbs scarce labor and raises rates. But he argues that denying its effect on demand requires "heroic assumptions." Under Okun's Law, losing one percentage point of growth would cost almost a million jobs.

On the consumer side, Smith says AI has spread through households faster than any consumer technology on record. Pew polls show people use it for many things, especially medical advice. Brynjolfsson et al. estimate $172 billion a year in US consumer surplus. That is more than the combined revenue run rate of OpenAI and Anthropic, and about half of Nvidia's annual profit.

Smith agrees that Americans dislike AI. He says the reasons they give are job loss, existential risk and over-reliance on the technology, not oligarchs keeping the gains.

• Self-selection: below random.
• Random choice: the baseline.
• Peter Principle (promote the best performer in the current role): better than random.
• Emotional intelligence: about the same as the Peter Principle.
• IQ: clearly better than both.
• Economic decision-making ability: best of all.

"Economic decision-making" here does not mean commercial instinct or chasing revenue. It refers to a measure developed by Andrew Caplin and co-authors: how efficiently a person allocates limited resources. In practice, the people who score well hand each task to whoever can do it most productively.

The Peter Principle result has a catch. Laurence Peter's 1968 idea is that people keep getting promoted until they reach a job they can't do, and then stay there because firms rarely demote anyone. The experiment suggests promoting on past performance works reasonably well. But it was a one-off lab exercise, so it can't capture that long-run trap.

Klement argues the finding is intuitive but rarely acted on. He doubts many firms promote people for their ability to allocate resources well. One limit is worth keeping in mind: this was a short lab task with reasoning puzzles, not a real company over years.