Hindsight Portfolios and Hidden Repo Plumbing

How each asset behaved

Gold had the most top-ranked years, 13 of 54, or about one year in four. It also tied with commodities for the most bottom-ranked years, 11 each. It was the most volatile asset, with a standard deviation of 24.7%. Its best year was +106% in 1979 and its worst was −37.6% in 1981.

The inflation question

Maggiulli split the years into high-inflation years (consumer prices up more than 4%) and the rest. Most financial assets did much worse when inflation was high. The S&P 500's average return was 11 percentage points lower and its median 17 points lower. International stocks, REITs, corporate bonds and Treasuries each fell by roughly 8 points or more.

Commodities were the exception. Their average return was about 2.9 points higher in high-inflation years, and their median was essentially unchanged. Gold's average rose by 5.6 points, the largest increase of any asset, but its median fell by 7.7 points. When the average rises and the median falls, a handful of extreme years are doing the work. For gold those were +106% in 1979, +59% in 1973 and +49% in 1974. Maggiulli warns these may not repeat. He points out that gold's +60% year in 2025 came three years after inflation peaked in 2022.

The optimal mix

The mix with the highest Sharpe ratio over 1972–2025 was:

• 36% U.S. homes
• 27% S&P 500
• 18% gold
• 16% 10-year Treasuries
• 3% REITs

T-bills are excluded by construction, because they serve as the risk-free rate in the Sharpe calculation.

The numbers show the trade-off. The optimal mix averaged 4.7% a year in real terms with a standard deviation of 6.6%. The S&P 500 averaged 8.3% with 16.8%. That is 1.77 times the return for 2.55 times the risk.

As an illustration, consider why swings cost money. A portfolio that gains 25% and then loses 20% has an arithmetic average of +2.5% but ends exactly where it started. Maggiulli's point is that avoiding deep holes is itself a source of return. The mix's worst decline was about 13%, in 2022. The S&P 500's was 48%, in 1974.

Leverage, and the mortgage loophole

Housing is a partial exception. A 30-year fixed mortgage is leverage that a lender cannot call in when prices fall. Maggiulli calculates from the Federal Reserve's 2022 Survey of Consumer Finances that the median American homeowner holds 51% of net worth in home equity. That isn't far from the optimizer's 36%. He suggests many households already run a partly leveraged version of the mix without realizing it.

He also accepts the objection that a home is something you live in, not just an investment. Removing homes from the analysis gives:

• 43% S&P 500
• 26% gold
• 22% Treasuries
• 9% REITs

Why the best portfolio is an illusion

The key test is to run the same optimization over only the first half of the data, 1972–1998. That gives 68% S&P 500, 24% gold, 6% international stocks and 2% corporate bonds. Homes, REITs and Treasuries, which make up 55% of the full-period answer, are completely absent. The result depends heavily on which years go in.

Maggiulli says this isn't a forecast against any asset. He simply doesn't know whether the U.S. will default, whether gold will fall out of favor, or whether ageing populations will hurt REITs. Optimization, he writes, "gives you the perfect solution to a problem that no longer exists." His conclusion is that the truly optimal portfolio is the one you can stick with through whatever happens.

What the piece leaves open

Several points go unexamined:

• Smoothed home prices. The 36% housing weight rests largely on home prices being recorded infrequently, which makes them look less risky than they are. Real homes also carry costs a yearly price index ignores: transaction fees, maintenance, property taxes and the fact that you can't sell a slice of one.
• Frictionless assumptions. The analysis appears to assume yearly rebalancing at no cost and no taxes.
• Annual data. Using one return per year hides declines that happened within a year.
• Out-of-sample testing. The article doesn't test how the 1972–1998 mix actually did from 1999 onward. That would show directly how much a backward-looking optimum costs or earns going forward.

The instrument

A repurchase agreement, or repo, is a sale paired with a promise to buy back. One party sells a security for cash and commits to repurchase the same security, or a closely related one, at a fixed price on a fixed date. The other party sees the same deal in reverse and calls it a "reverse repo." Economically it is a collateralized loan. The time between the sale and the buyback, together with the price difference, sets the interest rate, called the repo rate.

The authors cite Office of Financial Research (OFR) estimates of about $13.5 trillion in repos outstanding on a given day, roughly 40% of U.S. GDP. About 70% are backed by Treasury securities, and most mature overnight.

Several factors move repo rates:

A worked example

The authors give a funding example. An MMF has $98 million to invest overnight and lends it to a repo dealer. In return it takes a basket of Treasuries with a 2% haircut, so the collateral is worth about $100 million. The next day the MMF receives $98,009,800 and returns the securities. The $9,800 of interest equals 3.6% annualized. You can check this with the money-market convention of a 360-day year: $98 million × 3.6% ÷ 360 ≈ $9,800. If the dealer failed overnight, the MMF would hold bonds worth more than the cash it lent.

Why lenders trust it

Two motives on the lending side

The borrower in a repo always wants cash. The lender can have one of two reasons.

The borrowers

The next three largest borrowers are U.S. branches and agencies of foreign banks, U.S. depository institutions such as commercial banks, and real estate investment trusts. Together they have borrowed between $800 billion and $1.3 trillion since 2013. That is a much flatter path than the hedge funds'.

The lenders

MMFs lend the most. Their repo lending rose 90% from January 2013 to January 2021, from under $600 billion to $1 trillion. It then nearly tripled in five years to $3 trillion by January 2026. The authors explain the appeal: repo is a short-term investment that meets MMFs' regulatory limits on maturity and asset type, and it stays flexible enough to handle investor withdrawals.

What the post leaves open

This is a descriptive primer and does not assess risk. A reader may notice that the two fastest-growing sides of the market match up: money fund cash increasingly funds leveraged hedge fund bond positions, with dealers in between. Whether that concentration is a problem is a question the post does not raise. It also does not explain why MMF lending accelerated so sharply after 2021. Dealers, who channel cash from lenders to borrowers, are left for the second post. The Fed's own repo operations are covered in the third.

Bill McBride of the Calculated Risk newsletter has published the first half of a two-part look at the US housing market as of mid-September 2026. This part covers inventory and sales. Part 2 will cover prices, rents and mortgage rates. The free section is short, and the more detailed supply and sales data sits behind a paywall. What's public still gives a clear picture of a market where homes are piling up while buyers stay scarce.

Sales have been stuck for four years

The article doesn't say why, but the usual explanation is well known. Millions of owners refinanced or bought at rates below 4% in 2020–21. Once rates moved into the 6–7% range, selling meant giving up a cheap loan for an expensive one. Owners stayed put. Since most sellers also buy their next home, low listings meant low sales on both sides.

Supply is now outrunning demand

Here is an illustrative example (not the article's numbers). If 1.2 million homes are for sale and they sell at 4 million a year (about 333,000 a month), that's roughly 3.6 months of supply. Hold sales steady and add inventory, and the figure rises. Hold inventory steady and let sales fall, and it also rises. In 2026 both forces are working together: sales are flat to down and listings are creeping up. Analysts generally treat about six months as balanced. Readings well below that favor sellers, and readings above it favor buyers.

What the listing data shows

McBride cites Realtor.com's August 2026 housing trends report:

These figures show something important. Inventory is growing even though new listings aren't rising. That points to homes sitting on the market longer instead of more owners deciding to sell. Months of supply climbing to a decade-plus high while inventory is still below pre-pandemic levels tells you the weak part is demand, not a flood of supply. McBride also hedges, saying inventory growth "might have picked up again." One month of acceleration isn't a trend yet.

He notes that listings are seasonal. He expects them to stay roughly flat for the next couple of months and then fall toward year-end as sellers pull homes off the market for the holidays. A drop in November or December would be normal and wouldn't mean the market is tightening.

Prices: still up, but past the peak

The median existing-home price was up 1.6% year over year in August. McBride says it "has peaked for the year." Median prices normally peak in early summer and ease into winter. The more telling point is how small the annual gain has become. A 1.6% nominal increase is likely below general inflation, which would mean a small decline in real terms. Part 2 is promised to cover prices in detail.

A hard year for builders

McBride describes 2026 as "a difficult year for homebuilders." Three things stand out:

• There are still a large number of completed new homes for sale, meaning finished houses with no buyer.
• More homes than usual are unsold while still under construction.
• Builders have been cutting prices to clear inventory.

A finished, unsold house is costly for a builder. It carries financing and holding costs and ties up capital. That is why builders discount, often by paying to lower buyers' mortgage rates as well as by cutting list prices. New homes compete with resales, so builder discounts also put pressure on existing-home sellers.

Regional differences

McBride stresses that there are "significant regional differences," but the details are behind the paywall. In past years his coverage has typically contrasted the South and West, where inventory recovered fastest and prices are softer, with the Northeast and Midwest, which have stayed tighter. Whether that pattern still holds this year isn't shown in the free section.

What stays open

The public portion doesn't give the months-of-supply figure, the August SAAR, or the regional breakdown. The broader questions are also unresolved. How much of the inventory build-up comes from buyers stepping back, and how much from sellers slowly coming back? And would lower mortgage rates unlock sales or simply bring more listings? Until Part 2 appears, the takeaway is supply rising against weak demand, slowing price gains, and builders doing the discounting.

Nick Maggiulli, who writes the personal-finance blog Of Dollars and Data, set out to answer a question many tech workers face: is the equity a startup offers worth the lower salary that usually comes with it? He builds his answer from published data on how often startups sell and for how much, adds a model of what happens to a grant along the way, and ends with a blunt number. For a typical early employee, the expected payout is smaller than the pay they give up.

The headline number is not the payout

His example is employee number 10, who takes a below-market salary and gets 0.25% of the company. The tempting math says a $100M sale pays $250,000. Maggiulli lists six things that shrink that figure.

He deliberately leaves out some complications: restricted stock units (RSUs, shares granted outright, which later-stage firms use instead of options), harsher preferences, "acqui-hires" that wipe out common stock, and secondary sales that let employees sell some shares early.

Back to the opening example. Two rounds of dilution leave 0.17%, which is $170,000 on a $100M sale. Subtract a $7,500 exercise cost and a 40% tax and $97,500 remains, 61% less than the naive number.

How often startups pay out at all

CB Insights, a startup-data firm, tracked 1,119 US companies at the seed stage, their first outside funding. Only 30% had an exit. A working paper from Adigital found that more than 80% of US tech startups acquired between 2002 and 2020 sold for under $50M. Large IPOs are more common than large acquisitions, but acquisitions far outnumber IPOs: in 2016 there were 3,260 global acquisition exits against 98 IPOs. From these figures Maggiulli estimates:

Each payout assumes the grant is fully vested, a flat $7,500 exercise cost and 40% tax. Weighting each outcome by its probability gives an expected value of $35,820. Maggiulli compares that with a salary $20,000 a year below market. Over four years that is $80,000 of forgone pay, or $48,000 after tax, and the equity's expected payout arrives years later.

His model leaves out liquidation preferences and discounting future money to today's value, and he notes that either would push the expected value lower. Doubling the grant to 0.5% roughly doubles every figure. Leaving early without the cash to exercise sets the payout to zero. In his base case there is roughly a 5% chance of a six-figure payout.

The median employee

Most startups never exit, but most startup employees still go through an exit, because companies that sell tend to be bigger. Maggiulli estimates headcounts from data by Carta, a company that manages startup ownership records. He assumes 20 employees at companies that never exit, 70 at small exits, 250 at big ones and 1,000 at unicorns. Every 100 startups then employ about 5,150 people, and roughly 73% of them work at a company that exits.

The median employee lands in the small-exit group: roughly employee #35 at a 70-person company. Maggiulli estimates a grant of 0.05% for that person. After one round of dilution and a $50M sale, the gross is $20,000. After a $1,500 exercise cost and 40% tax, the payout is $11,100.

Even that assumes the person exercised. Carta found that as of late 2024, employees exercised only 32% of vested options that were worth more than their strike price.

Making the deal better

Maggiulli concludes that equity mostly disappoints. He also notes that its long right tail, the small chance of a very large payout, is something almost no salaried job offers. He argues the stronger reasons to join a startup are experience and network, especially for someone young with low expenses. He also warns that your paycheck and your equity depend on the same company, a concentrated bet he would not normally recommend.

His practical advice:

His exit probabilities and headcounts are his own estimates stitched together from several sources. The 40% flat tax and fixed exercise costs are simplifications. Treat the results as orders of magnitude, not forecasts.

The problem: correlations that break

Smith's simplified version of the 1970s is that this failed. Policymakers pushed inflation up and unemployment stayed high, which gave the economy stagflation. A correlation is not a cause, and once people expect a policy, the old pattern can disappear. Economists began asking which parts of the economy stay fixed when policy changes.

Hall's 1978 consumption test

Hall gave this approach a key piece of evidence. Milton Friedman had argued that people plan their spending ahead. If they do, then apart from a smooth trend, quarter-to-quarter changes in consumption should be unpredictable. Any known information, such as last quarter's income, would already have shaped spending last quarter. Hall tested this and found that past income did not predict changes in consumption.

The result wasn't airtight. Stock prices did predict changes in consumption. A surprise like a government stimulus could still move spending at once. Smith says the paper still mattered for two reasons:

• If people plan ahead, policymakers can change behavior today with promises about tomorrow.
• If economists can't forecast how spending will change, consumers plan about as well as economists do. Policy therefore works best when the public understands what is being done.

A wrong turn, then a messy boom of ideas

The first major model built this way was real business cycle (RBC) theory, which later won its own Nobel. It held that booms and recessions come from changes in the pace of technological progress. The data didn't support it: productivity booms don't line up with economic booms. RBC also implied that interest-rate policy has no effect on output or jobs. Its creator, Ed Prescott, compared monetary policy to "rain dancing." Smith says the best evidence shows Prescott was wrong.

In the 1990s many economists tried to explain why monetary policy does affect jobs. Most of their ideas were about sticky prices. When demand falls, for example after a Fed rate hike or a banking panic, why don't firms simply cut prices and keep producing? If prices fall too slowly, firms produce less instead. So many people contributed (Mankiw, Romer, Blanchard, Yellen, Calvo, Galí, Gertler, Taylor and others) that Smith thinks it was unclear who deserved the prize, which helps explain the delay.

The name is somewhat misleading. Keynes focused on government spending and taxes, which economists call fiscal policy. Early New Keynesians focused almost entirely on the central bank, which owes more to Friedman.

Hall added a missing piece. The basic model linked interest rates to output but not to jobs. His 2005 paper showed how slow-moving wages force firms to lay workers off in a downturn, because they can't simply cut pay.

Woodford's synthesis

Woodford turned this mass of ideas into a model simple enough for central banks to use, first in a paper with Julio Rotemberg and later in a well-known book. It has three main parts:

The model explains how a central bank can cause a recession, as Paul Volcker did in the early 1980s, but can't keep unemployment low forever with low rates. It also explains how cutting rates can cushion a collapse in private demand. It became the standard tool at central banks around the world.

The zero problem and the return of Keynes

In 2008–09, the slump was so deep that cutting rates to zero wasn't enough. Smith describes three responses:

The acid test: 2021–23

Smith calls the recent inflation the theory's real test. Inflation peaked around 9% in 2022 and fell to about 3% by early 2023 after the Fed raised rates. Unlike in the early 1980s, unemployment didn't rise much. Smith admits the cause isn't certain. His leading explanation is that firms stopped raising prices once they believed the Fed would do whatever it took. A credible commitment plus forward-looking expectations did the work, without a recession. Those are the same mechanisms Hall and Woodford built into their models.

What stays open