The 48% Collateral: Auditing the American Household Balance Sheet as a Protocol Under Stress

Hook

# reconstructed from Federal Reserve Z.1, Table L.1
# Households and Nonprofit Organizations, nominal USD, latest quarter
E_direct    = 27.0e12   # corporate equities, directly held
E_fund      = 16.0e12   # mutual fund shares (equity share of funds ~0.60)
E_noncorp   = 15.0e12   # noncorporate business equity
E_pension   = 15.0e12   # pension entitlements (DC look-through ~0.65)
A_financial = 115.0e12  # total household financial assets

rho_a = (E_direct + E_fund) / A_financial # ~0.374 rho_b = (E_direct + E_fund + E_noncorp) / A_financial # ~0.504 rho_c = (E_direct + 0.60E_fund + 0.65E_pension) / A_financial # ~0.402 rho_d = (E_direct + E_fund + 0.65*E_pension)/ A_financial # ~0.459 ```

Four constructions. Four answers. All defensible. All drawn from the same quarterly release.

The headline circulating through the crypto press right now is 48% — the share of American household financial assets sitting in equities. That number is real. It also is not a primitive. It is the output of a definition: which accounts you fold together, how far you look through fund wrappers, whether you count noncorporate business equity as equity, and whether pension entitlements get treated as a bond-like claim or an equity-like one.

What the headline is doing, whether or not the author intended it, is picking the top of the plausible band and reporting it as a fact about the middle of the distribution. That is not a scandal. It is routine. But it matters, because the entire policy argument built on top of the number inherits the ambiguity of the number.

Here is the more useful framing. The American household is now the terminal holder of the residual claim on the US corporate capital stock. Not a portfolio preference — the output of a system. The household balance sheet is a protocol under stress. The 48% is its collateral ratio. And the collateral is marked to market by a price feed the household does not control, cannot audit, and has no governance rights over.

I have spent the last decade auditing protocols for exactly this class of problem. The failure modes rhyme.

Context

The Federal Reserve publishes the Financial Accounts of the United States — Z.1 — every quarter, roughly ten weeks after period end. It runs to several hundred pages. It contains tables L.1 through L.230 plus the flow tables, and it is the same document family that produces the headline net worth numbers the business press reports as "household wealth hit a record."

The accounting unit is not a household. It is "Households and Nonprofit Organizations." An NPISH sector in the System of National Accounts vocabulary. A university endowment and a family of four are aggregated into one line item. This is not an error; it is the shape of the accounting framework. But it means that any statement about "the American household" derived from this table is a statement about a blended entity that does not exist.

The second definitional move is the one that does the most work. Z.1 splits household assets into financial and nonfinancial. Real estate lives in nonfinancial. Owner-occupied housing, at rough replacement and market value, is somewhere in the vicinity of $48 trillion. That is the single largest asset class most American households will ever hold, and it is entirely outside the 48% denominator.

So the correct sentence is this: American households hold a record share of their financial assets — not their assets — in equities. If you put real estate back into the denominator, equity exposure lands closer to 25 to 30 percent of total household assets. The number is still historically elevated. It is not the number on the poster.

There is a third framing that neither the media piece nor the Fed's own commentary states directly. Every financial asset is someone else's liability. Household equity holdings are a residual claim on the corporate sector — the last claim paid in the waterfall, after wages, after trade credit, after senior debt, after taxes. When I write that the household is the terminal holder of the residual, I mean it in the strict balance-sheet sense. The household sector now holds the most junior tranche of the US corporate capital structure, in size. That is a structural position, not a sentiment.

The path to that position ran through three policy decisions that were never framed as asset-allocation policy.

The first was the shift from defined-benefit to defined-contribution retirement. ERISA in 1974, the Revenue Act of 1978 that quietly created the 401(k) subsection, and then four decades of plan sponsors closing DB plans because DC plans shifted the funding risk onto participants. The risk transfer was the point. It was also, in aggregate, a mandated conversion of household savings into securities-market exposure with no floor, no guarantee, and no pooling across cohorts.

The second was passive indexing. The first retail index fund launched in 1976. Passive share of US fund assets crossed 50 percent somewhere in the mid-2020s. Once passive exceeds active, price discovery stops being a market function and becomes a residual. The index provider — not the investor — decides what the household owns.

The third was the interest rate regime. From late 2008 through early 2022, the policy rate sat at the zero lower bound for most of the period. Deposit rates followed. Real returns on cash were approximately zero for over a decade. An agent maximizing real terminal wealth under a legally suppressed risk-free rate will hold equity. This is not a behavioral anomaly. It is a constrained optimum.

Those three decisions compose. DB to DC pushes savings into securities. Passive indexing concentrates those savings into a cap-weighted basket. Financial repression removes the alternative. The 48% is the residual of that composition.

Core

The wealth effect is a transfer function, not a law

The standard story says equity gains drive consumption. The mechanism is usually stated as an MPC on stock wealth of three to five cents per dollar.

Run the arithmetic before accepting the conclusion. Look-through household equity exposure — direct plus indirect plus the equity share of DC balances — is on the order of $50 to $55 trillion depending on construction. A 10 percent drawdown destroys $5 to $5.5 trillion of paper. At an MPC of three cents, that is $150 to $165 billion of annual consumption. Against a $29 trillion nominal GDP, that is roughly 0.5 percentage points.

At five cents the figure doubles to about 0.9 points. And here is where the popular treatment breaks down. The MPC on losses is not the MPC on gains. The literature on loss aversion and on liquidity-constrained households converging after a shock is consistent on this point: consumption responds more to a drawdown than to an equivalent mark-up. The transfer function is asymmetric.

Model it properly. Let C_t = c(Y_t, W_t) where W_t is household net worth. The partial derivative with respect to W is positive in both directions, but not equal. In the down direction, the derivative steepens because a fraction of households are simultaneously hitting liquidity constraints — margin calls, home equity lines, tuition timing, medical events. Those households are forced sellers or forced reducers, and their consumption function is not smooth. It has a step in it.

Five trillion dollars of paper losses do not translate into economic damage. Five trillion dollars of paper losses that arrive with margin debt at cyclical highs and a retail options complex sitting on short-dated convexity does. That is the version of the wealth effect worth worrying about, and it is not the version in the headline.

Reflexivity and the Fed's reaction function

Write the loop as a discrete system.

E_{t+1} = f(E_t, r_t, pi_t, eps_t)        # asset prices
C_t     = c(Y_t, E_t)                      # consumption
Y_t     = g(C_t, I_t, G_t, NX_t)           # output
pi_t    = h(Y_t, u_t)                      # inflation
u_t     = u(u_{t-1}, Y_t)                 # unemployment
r_t     = Fed(pi_t, u_t)                   # policy rule

The policy rule takes two arguments. Inflation and unemployment. It does not take E_t. That absence is the entire story.

For fifteen years the market has been calling that function with a third argument it does not accept. The 2020 repo dislocation, the 2018 fourth-quarter drawdown, the 2019 funding squeeze — each produced a policy response that was explained in terms of the two official arguments and motivated by the third unofficial one. That behavior has a name in the casual register: the Fed Put. In systems terms it is a runtime patch applied to a function signature that was never updated. The rule still reads Fed(pi, u). The observable behavior reads Fed(pi, u, E).

At 25 percent equity share, that mismatch is tolerable. Asset prices are a second-order input; the policy rule approximately closes.

At 48 percent, the missing argument is a first-order input. The household sector's spending is now coupled to the price of the residual claim on corporate earnings. That coupling means the policy rule is operating with an unmodeled state variable — a latent variable that drives both of its official arguments through the wealth-effect channel and is not itself in the equation.

Control theory has a term for a system with an unobserved state variable feeding back into the controller. It is not "robust." It is "marginally stable at best, and the margin is not measurable from the outside."

Concentration: the index is not a diversification primitive

The aggregate statistic says households hold equities. It does not say which ones.

The top ten names in the S&P 500 have at various points over the last two years summed to well over 35 percent of index weight. A cap-weighted index with that weight distribution is not a diversified basket in any meaningful sense. Run the Herfindahl index.

weights (illustrative, top 10):
7.0, 7.0, 6.5, 4.0, 2.7, 2.5, 2.5, 2.0, 1.5, 1.5   # percent
sum of squares (top 10)  ~ 183  (%^2)
remaining 490 names, ~62% weight, decaying distribution:
sum of squares           ~ 60-90 (%^2)

HHI ~ 0.024-0.027 Effective N = 1 / HHI ~ 37-42 ```

A portfolio that says "500 stocks" on the label behaves, under concentration metrics, like a 40-name book. The diversification the household believes it bought is a wrapper, not a property of the underlying.

This matters more than it looks, because of a second property. Passive vehicles are price-insensitive buyers. They buy whatever the index says, in the weight the index assigns, on the schedule the index defines. In market microstructure terms, their demand curve is vertical. There is no price at which a passive flow stops.

A vertical demand curve does not stabilize. It transmits. When the price moves, the flow does not respond by leaning against the move the way a value investor's flow would. It responds by mechanically replicating the new weight. On the way up, that is a momentum amplifier. On the way down, it is a momentum amplifier. The index structure has removed the natural counterweight that active capital used to provide, and replaced it with a feedback term.

The household is not holding a diversified portfolio. It is holding a levered expression of a concentrated one, with the leverage embedded in the correlation structure rather than in a margin account.

The denominator: financial repression, not greed

Every commentary on this topic eventually reaches for a moral frame. Households chased yield. Households got greedy. Households speculated.

Set that aside and look at the relative price vector.

From 2009 through 2021, the effective real return on transactional deposits in the US was at or below zero for most of the period. The real return on short-duration Treasuries was negative for long stretches. The earnings yield on the S&P 500, while not spectacular, was structurally higher than the risk-free alternative. Under those constraints, the rational allocation to equity rises. Mechanically. Without any change in preferences.

The 48% is not a preference vector. It is a price vector.

The policy literature has a term for the deliberate maintenance of negative real rates to reduce the real burden of public debt: financial repression. It is not a conspiracy; it is a well-documented toolkit, and it has a documented side effect. It pushes savers up the risk curve. Every household that would have held cash or bills in a 4 percent real-rate world now holds duration and equity in a zero real-rate world.

Math doesn't negotiate. An agent maximizing real terminal wealth under a suppressed risk-free rate will hold the risky asset. The only variables are how much and with what wrapper.

This reframes the policy question entirely. If the objective is to reduce household equity concentration, there are exactly two levers. Raise the real risk-free rate, which raises the cost of servicing the debt. Or build a non-market savings instrument with a real return, which is politically indistinguishable from a fiscal transfer. There is no third lever. Telling households to diversify more is not a policy; it is an exhortation.

The wrapper: 401(k) as an unauditable fee-bearing contract

I have audited a lot of vault contracts. The 401(k) is structurally a vault. It has a deposit function, a menu of permitted strategies, a withdrawal schedule gated by age, a fee layer, and an administrator with discretionary authority over the menu.

The differences from a well-built on-chain vault are instructive.

A good vault publishes its fee in the contract, has a timelock on parameter changes, and lets the depositor exit at any block. The 401(k) has an expense ratio that is disclosed in a document most participants do not read, a menu that changes at the administrator's discretion with notice periods measured in weeks, and an exit function that is penalized until age 59.5. There is no on-chain equivalent of a 10 percent early-withdrawal penalty, and if there were, it would be classified as a honeypot.

On fees, the arithmetic is unforgiving. A 100 basis point annual drag against a 7 percent gross return consumes roughly a quarter of terminal wealth over a 40-year horizon. That is not a rounding error. That is a permanent transfer from the household to the wrapper.

On hedging, the standard glide path assumes bonds hedge equity. That assumption held for four decades and then, in 2022, did not — both legs of the portfolio fell together, and the diversification the model was priced on evaporated precisely when it was needed. A glide path that mechanically de-risks into duration is not a hedge. It is a bet on a correlation that the model treats as a constant and the market treats as a variable.

The most important point about the wrapper is not the fee. It is that the risk was transferred to the participant without transferring the actuarial infrastructure that used to absorb it. A DB plan pools longevity risk and market risk across cohorts, and the sponsor bears the shortfall. A DC plan does neither. The participant bears market risk at the specific moment in their life when sequence-of-returns risk is highest, which is the decade on either side of retirement, and there is no pool to spread it into.

That is not a design flaw. It was the design.

No proof of reserves for the household sector

Here is where the balance-sheet framing and the cryptography I actually work on intersect.

We can observe the aggregate. We cannot verify the distribution.

The Fed publishes a sum. It does not publish a proof. That means we cannot verify, from the outside, three things that determine whether 48% is a stability risk or a stability fact.

We cannot verify who holds what. The distribution across the wealth scale is the load-bearing variable. If the top decile holds the overwhelming share of the equity, then the consumption channel is narrow and the political channel is wide. If the exposure is broad, the reverse. The aggregate cannot distinguish these cases, and they have opposite policy implications.

We cannot verify leverage. Margin debt, home equity lines used for securities, structured products with embedded leverage, and short-dated options convexity are all outside the equity line item. A household with $500,000 of equity exposure and $200,000 of margin debt has a different beta than one with $300,000 unlevered, and the Z.1 table reports both the same way.

We cannot verify cost basis. A household sitting on a 300 percent unrealized gain behaves differently from one that bought at the top. The first has a tax reason not to sell. The second has a margin reason to.

A zero-knowledge proof is the obvious primitive here. Commit a Merkle tree of accounts. Prove the sum. Prove each individual balance is non-negative. Prove the aggregate ratio without revealing any individual position. The cryptography for this has existed since the late 2010s; the arithmetic is not the hard part.

The hard part is completeness. You can prove that every account you included sums to a given number. You cannot prove that you included every account. That is the exact gap that made exchange proof-of-reserves reports of limited value in the years before and after the failures that made them famous — a commitment scheme with a biased or incomplete witness proves nothing about the population it claims to describe.

Privacy is a protocol, not a policy. A policy that says "we will protect your data" is a promise. A protocol that proves a statement about encrypted data without revealing the data is a construction. For household balance sheet aggregates, we have the policy and not the construction, which means every claim about household fragility rests on an unverifiable sum published by the same institution whose policy created the configuration.

That is not an accusation of bad faith. It is an observation about observational equivalence. The measurement and the mechanism are not independent.

Oracle latency: the price feed problem

A price feed is not an oracle. It is an opinion with a timestamp.

I have written before that oracle latency is the soft underbelly of DeFi, and that the decentralization framing around node-operator sets deserves more scrutiny than it usually receives. The same criticism transfers to the equity market with only cosmetic changes.

US equity prices are produced by a consolidated tape system with exchange-level feeds, a securities information processor, and a set of index providers who publish methodology documents that are the de facto specification for what most household wealth tracks. The index provider is a centralized oracle. It has discretionary authority over inclusion, free-float adjustment, weighting, and rebalancing schedules. It publishes a methodology, which is a whitepaper, which is not code.

The household has no governance token in that oracle. It does not vote on the methodology. It cannot fork.

The comparison to the on-chain equivalent is direct. A decentralized oracle network is a set of node operators staking reputation on reported values, with aggregation rules and deviation thresholds. Whether that is meaningfully more decentralized than a committee of index providers is an empirical question about operator overlap and correlated incentives, and I have not seen a satisfying answer to it. What I have seen is that both systems fail in the same way when it matters: not by reporting a wrong value, but by being unable to update the right value fast enough during a discontinuity.

Now add the 2025 wrinkle. Tokenized equity exposure trades continuously against a feed that is only live during a 6.5-hour session. During the closed window, the tokenized instrument is priced by whatever liquidity is present in its own pool. That price is not the equity price. It is the pool's opinion about the equity price. When the session reopens and the underlying gaps, the pool price and the reference price separate, and every position with a health factor computed against the closed-session feed becomes a liquidation candidate.

This is exactly the stale-price and thin-book manipulation pattern that generated a whole category of exploits between 2021 and 2023. The instrument is new. The failure mode is not.

The hedge that isn't

If the household is over-concentrated in equities, the obvious trade is to add an uncorrelated asset. Crypto has spent a decade marketing itself for that slot.

The marketing does not survive contact with the correlation data. Reconstructed BTC-NDX rolling correlation has repeatedly printed in the 0.5 to 0.8 range during stress windows. In March 2020, cross-asset correlations converged toward one across the board. In 2022, the drawdown in crypto and the drawdown in long-duration equity were not independent events; they were the same event, expressed through the same liquidity function.

The reason is mechanical and worth stating plainly. In a margin call, you sell what trades. A hedge has to be either negatively correlated with the losing leg or uncorrelated and liquid enough to monetize. Crypto is liquid, twenty-four hours a day, in size, which makes it a superb source of funds and a poor hedge. The 24/7 property that its advocates cite as a structural advantage is the same property that guarantees it gets sold first when a leveraged holder anywhere in the system needs cash.

There is a second-order effect that matters for the 48% question specifically. If a household holds a concentrated equity book and pairs it with a crypto allocation it believes is uncorrelated, and the true correlation in the tail is positive, then the household has not diversified. It has increased its exposure to a single global liquidity factor while believing it has reduced exposure to a single-sector factor. The perceived risk is lower. The realized risk is higher. That gap is where the losses live.

None of this is an argument about valuation. It is an argument about the correlation matrix under stress, which is the only correlation matrix that matters.

Terra as the fast-motion control case

In the six months after the 2022 collapse, I stepped out of public writing and did nothing but read consensus code and mechanism designs from failed layer ones. The output was a long paper on algorithmic stablecoin instability. The relevant finding here is not about stablecoins. It is about topology.

The Terra mechanism was a reflexivity machine. Anchor's subsidized yield pulled deposits. Deposits burned LUNA. The burn reduced supply. Reduced supply supported price. Higher price supported confidence. Confidence supported more deposits. Every leg of the loop was endogenous; the only exogenous input was the subsidy, and the subsidy was finite.

Write the same diagram for the American household and the labels change, the loop does not.

E_t up -> W_t up -> C_t up -> Y_t up -> earnings up -> E_t up
r_t low -> discount rate low -> E_t up
E_t up -> r_t pressure -> Fed response -> r_t low -> E_t up

Both systems have a reinforcing loop with no exogenous anchor. Both have a finite backstop. In Terra's case the backstop was the LUNA treasury and the reserve assets, and it ran out in about a week. In the household case, the backstop is the central bank's willingness to expand its balance sheet, which is not finite in nominal terms but is finite in credibility terms.

The structural difference is time constant, not shape. Terra's loop completed in days. The household loop completes in quarters. A slower loop looks more stable for longer and gives the participants more time to build position size on the assumption that the stability is structural.

I want to be precise about what I am and am not claiming. I am not claiming the US household balance sheet will collapse the way Terra did. I am claiming that the topology is isomorphic and that the difference in time constant is a difference in observability, not in kind. A system that takes four years to express its fragility is not safer than one that takes four days. It is only easier to describe as safe while it is happening.

Prescriptive note for builders

If you are building a lending market, a vault, or any protocol that consumes a household-adjacent balance sheet as collateral, there are three things the Z.1 story should change about your defaults.

Cap correlated collateral at the portfolio level, not the position level. If your borrowers all post the same two assets, your effective collateral ratio is not the weighted average of individual ratios. It is the ratio of the whole book against the worst joint scenario, which is lower by a factor you can compute and probably have not.

Treat your price feed's downtime as a solvency event, not an availability event. A feed that stops updating during a discontinuity is not a degraded feed. It is a feed reporting a value you cannot use. Design the liquidation path so that it halts rather than executes against a stale reference.

And write down the completeness assumption on any proof you publish. A Merkle-sum commitment over a set of accounts proves the sum of the set. It does not prove the set is the population. Say so in the specification, because if you do not, someone will read the proof as a claim about the world rather than a claim about your data structure.

Contrarian

The consensus reading of the 48% is that households are dangerously overexposed and a correction would be catastrophic. That reading is directionally right and analytically wrong in three places that matter.

The first error is treating the number as a risk measure when it is a policy measure. A high equity share is not a bug in the household's behavior; it is the equilibrium output of a suppressed risk-free rate, a mandated DC retirement architecture, and a passive infrastructure that converts savings into index exposure automatically. If you want to lower the number, you have to change the inputs. Warning households about concentration while the deposit rate is below inflation is telling someone standing in a rainstorm that they should be drier.

The second error is the direction of the asymmetry. Most commentary frames the wealth effect as bidirectional — stocks up, spending up; stocks down, spending down. The correct framing is that the down direction has a steeper slope, because losses interact with liquidity constraints and gains mostly do not. A system with an asymmetric feedback term is not a system with a symmetric risk profile. It is a system that looks calm in the up state and accelerates in the down state. Every model calibrated on the full sample will underestimate the drawdown.

The third error is the one I find most interesting, because it is a blind spot shared by the bulls and the bears. The 48% figure is produced by an institution whose policy produced the 48% figure. The measurement and the mechanism are not independent. The Fed sets the rate that pushes households up the risk curve, and the Fed publishes the table that reveals how far up the curve they went. That does not make the data wrong. It makes the data endogenous, which means that any inference of the form "48% is unsustainable, therefore policy must change" has to clear a higher bar than usual, because the same institution supplying the evidence also has standing to dispute it.

Here is what I think the real blind spot is, and it is not on either side of the bull-bear divide.

The 48% is an average that does not exist. The distribution underneath it is almost certainly extreme. If the top decile holds the overwhelming majority of household equity, then the aggregate is a statistic about a small number of balance sheets wearing the costume of a statistic about all of them. The consumption channel that makes the number dangerous is narrow. The political channel that makes the number unstable is wide. Those two facts push policy in opposite directions, and a policymaker reading the aggregate cannot see either one.

Which brings back the completeness problem. We do not lack a way to prove the distribution. We lack the political will to create the proving system, because the proving system would reveal the distribution. That is the same reason exchange proof-of-reserves reports kept failing on exactly this axis. The cryptography was never the bottleneck. The willingness to commit to a complete witness was.

And the crypto industry's response to all of this — that 48% equity concentration is the argument for BTC as a hedge — has the causality backwards. If household equity concentration is a fragility, then the correct response is a lower-correlation asset. Crypto's realized correlation in every stress window since 2018 has been higher, not lower, than the correlation between equities and the traditional hedges. Marketing an asset as uncorrelated does not make it uncorrelated. Math doesn't care which deck the slide came from.

Takeaway

The signal to watch is not the 48%. It is the first quarter in which it declines, because a declining share in a rising market is a different event than a declining share in a falling one. The Z.1 release that shows the ratio breaking trend will be the first hard evidence that household allocation behavior has changed rather than household wealth. Until then, every claim about de-risking is a claim about prices, not positions.

The second signal is the indexing infrastructure. A methodology change at a major index provider — concentration caps, a narrowing of free-float rules, a change to rebalancing frequency — is a multi-trillion-dollar reallocation executed without a vote by the people whose assets move. Watch that document set more closely than the equity tape. It is a governance surface with no governance.

The third signal is the tokenized-equity complex. If it grows to the point where a closed-session gap produces a cascading liquidation that transmits back into the underlying, the loop between the on-chain price and the reference price becomes a genuine systemic channel. That has not happened yet. The infrastructure to make it happen is already deployed.

One question to hold. If the American household is the terminal holder of the residual claim on corporate earnings, and the household's willingness to keep holding it is a function of a policy rate that is set with no reference to that holding — then when the household becomes the seller, who is on the other side?

The 2020 answer was the central bank. That answer is available again. It is not infinitely available.