QE and the Estimated Neutral Rate

TL;DR

Money for long-term borrowing moves freely across borders, so liquidity is a global phenomenon. QE bought bonds from asset holders and paid in cash. With this cash they bought other assets, everywhere and in every class. Asset prices rose worldwide, including in countries that never ran a QE programme of their own.

That was inflation, in a market the consumer price index does not cover. And bidding an asset's price up is bidding its yield down, so the same event was a global fall in borrowing costs.

None of this began in 2009. Bernanke saw low rates in 2005 and diagnosed a global savings glut as the cause, favouring a structural explanation over a liquidity one.

Most econometric models that estimate the neutral rate are fitted to consumer prices, output and interest rates. Asset prices are not among the inputs. So the figure they report is the rate that is neutral for groceries, and it has nothing to say about whether borrowing was running ahead of the capacity to service it.

There is a second problem underneath. The models infer the rate from two relationships that a decade of successful stabilisation had flattened, leaving the history of the policy rate itself as the main thing to lean on. Policy rules then took that estimate as an input, and naturally recommended something close to the policy already in place.

Estimation needs variation, and stabilisation exists to remove variation. A central bank that succeeds gradually erases the evidence it needs to keep doing its job.

The liquidity story does the work of several. It explains why asset prices boomed where no programme ran, why consumer prices stayed quiet and then surged in 2021 when the same mechanics reached households instead, and why every country's estimate fell together despite different demographics, deficits and growth.

For a decade, the rate that would have kept borrowing serviceable sat far above the one the models report. That is not a measurement problem. It is a specification problem, and a specification problem returns a confident number to the wrong question.

Introduction

For most of the 2010s economists argued about why interest rates were so low. The answer usually offered was that the neutral rate of interest had fallen. Something structural in the world economy meant money was simply worth less than it used to be, and central banks were following that down rather than leading it.

There is a simpler possibility. Central banks pushed rates down, held them there for a decade, and the statistical models that estimate the neutral rate read those rates back and reported that the neutral rate had fallen.

Two claims follow and they are worth keeping apart, because the second does not depend on the first. One is that a decade of intervention contaminated the estimate, so the models could not tell a structural shift from a policy stance held for long enough to look like one. Much of that damage was done by the policy rate itself rather than by bond purchases. The other is that even a perfectly estimated figure would answer a narrower question than the people quoting it believe, and that is where the asset purchases matter most, because of what they did to where the money went.

What QE actually did

Quantitative easing is a central bank creating money to buy bonds, mostly government bonds, from whoever is holding them. Everyone calls it QE. The mechanics are worth following slowly, because most of the argument about it is really an argument about where the money went.

The central bank buys a government bond from a pension fund. It cannot pay the fund directly, because only banks hold accounts at the central bank, so the payment routes through the fund's bank.

Two entries are created, not one. The bank gets reserves as an asset. The fund gets a deposit. Banks cannot lend reserves to anyone outside the banking system. When a bank makes a loan it creates a new deposit, which is why the familiar story about QE handing banks money to lend out, the money printing version, was always wrong. The deposit belongs to the pension fund, and the pension fund can spend it on anything it likes.

Where the money went

The fund did not want cash. It owes pensions in twenty and thirty years and needs assets that match those liabilities. It has just swapped a long-dated bond for a deposit paying nothing.

So it buys something else. Corporate bonds, equities, property, foreign debt. Whatever it buys, it bids the price up. The seller of that asset then holds the deposit and faces the same problem, so the deposit passes along a chain of asset holders, lifting prices at each step, until it reaches someone content to hold cash.

Too many dollars chasing an insufficient supply of assets. Textbook asset inflation, in a market most people do not think of as a market.

A second effect ran alongside. Every bond the central bank bought was a bond private investors could no longer hold. With less long-dated paper available, investors accepted less compensation for holding what remained, which lifted bond prices directly. That one works even if every seller sat on their cash forever.

The Bank of England measured all this in July 2012 and published the numbers. Its £325 billion of purchases had raised household financial wealth by just over £600 billion. Asset purchases had, in its own words, pushed up the price of equities by at least as much as they had pushed up the price of gilts.

It also published who received it. The top 5% of households held 40% of the assets. The median household held about £1,500.

The market was already global

Nothing obliged that pension fund to buy British assets.

By the time quantitative easing arrived, the developed world had a single market for long-term borrowing. That did not happen because of QE. It happened a generation earlier, when capital controls came off across the OECD through the 1980s, when Australia floated the dollar in 1983 and dismantled its exchange controls, and when inflation targeting spread through the 1990s and dragged inflation rates together.

The evidence is in the yields themselves.

The chart takes every developed economy the OECD has a long series for. Wildly different fiscal positions, demographics and growth rates. Their average ten-year yield ran near 13% in 1982, across the nine with data that far back, fell to under half a per cent in 2020, and sits near 3.6% today. They moved together the whole way down and the whole way back, mostly keeping their places in the queue. Switzerland and Japan stayed cheaper throughout, Australia dearer. Italy broke ranks during the euro crisis and rejoined. Otherwise they turned at the same moments.

Italy and the Netherlands do not have similar public finances. Japanese demographics look nothing like Australian ones. Korean growth is not Belgian growth. If the price of long money were set by national savings, national investment opportunities and national populations, these countries would not share turning points for forty years.

A deposit created by the Bank of England buying a gilt could be spent on Australian bank debt, Brazilian government bonds, American technology shares or a Sydney office tower. The money had no nationality once it left the bond it came from, and nothing stopped it travelling.

Which is why the effect showed up in countries that never ran the programme. Australian house prices rose through a decade in which the Reserve Bank bought almost nothing. Emerging market borrowing costs compressed while the buying was being done in Washington, London, Frankfurt and Tokyo. The phrase used at the time was the global search for yield.

In 2013 Ben Bernanke said the Federal Reserve might slow its purchases, and currencies and bond markets sold off violently in Indonesia, India, Turkey, Brazil and South Africa. A statement about American policy repriced assets on four continents within days.

Hélène Rey's work on the global financial cycle makes the general case. American monetary conditions drive financial conditions worldwide, more or less regardless of whether a country floats its currency. A small open economy can set its own overnight rate. It cannot opt out of the global price of long money.

Why the groceries were spared

None of this showed up in the consumer price inflation figures, because the deposit was circulating among asset holders rather than spenders. Pension funds do not buy groceries. Insurance companies do not fill supermarket trolleys. The money stayed inside the financial economy, bidding up claims on future income, because that is where the people holding it operate.

In 2020 and 2021 governments posted cheques directly to households. Identical balance sheet mechanics, completely different recipients, and those recipients did buy groceries. The same broad policy produced asset inflation the first time and consumer inflation the second.

The measurement did the rest. Central banks target consumer price indices, and consumer price indices exclude assets. So a policy that raised asset prices enormously while consumer prices undershot registered as a success on the only gauge being read. The S&P 500 more than quadrupled between the start of 2009 and the end of 2020 while American consumer prices rose about 20%, and the second number set policy.

That invisibility is what let it run for a decade. Eight per cent on the CPI in 2012 would have stopped it. Eight per cent on asset prices did not.

Asset prices and borrowing costs are one thing

Bidding up the price of a bond is the same as bidding down its yield. Compressing credit spreads is the same as lowering the cost of corporate borrowing. The asset price inflation and the fall in borrowing costs were not two effects but one effect described from either end.

Borrowing costs fell everywhere. Government bonds, corporate bonds, mortgages. Companies borrowed cheaply and bought back their own shares, which lifted earnings per share by shrinking the number of shares rather than by earning more. Buybacks by S&P 500 companies ran to roughly five and a quarter trillion dollars across the decade to 2019, while non-financial corporate debt grew from about six and a half trillion to nearly eleven.

Decade-by-decade decompositions of share market returns split them into dividends, earnings growth and changes in valuation. For the 2010s they show almost all of it coming from earnings growth rather than re-rating, which looks like evidence that cheap money had nothing to do with it.

But a buyback lifts earnings per share without lifting earnings. Shrink the denominator and the ratio rises whatever profits do. Five trillion dollars of share retirement, funded substantially by debt issued at suppressed rates, sits in the earnings column of that decomposition rather than the valuation column.

I would not claim that accounts for the whole of it. It does mean the decomposition cannot be read as showing cheap money was absent, because one of the ways cheap money worked is recorded in the column said to prove it did nothing.

Neutral for what

The neutral rate of interest is the rate at which policy is neither stimulating nor restraining the economy. It cannot be observed. It has to be estimated, and the standard method runs a statistical filter over the history of interest rates, output and consumer price inflation to infer the trend underneath. Asset prices are not among the inputs.

Which builds the answer into the question. The rate the models report is the rate that is neutral for consumer prices, because consumer prices are what they are calibrated against. A rate can sit exactly at neutral for groceries and be wildly stimulative for assets, and the estimate has no way to register the second condition, having never looked at it.

Wicksell, who gave us the idea, meant something broader. His natural rate was the one that equated saving with investment in real capital, meaning new productive capacity rather than a change of ownership of what already exists.

Cheap money can lift the price of existing assets a long way without lifting the rate at which new capital is built. Nothing in the national accounts is violated by that. Every purchase of an existing asset has a seller on the other side and the identities hold. What changes is the margin, where the reward sits between building something and buying something already built. Weak business investment alongside record asset prices is what that looks like from outside, and it describes the 2010s exactly.

Claudio Borio and William White at the Bank for International Settlements made this argument for most of two decades. Stable consumer prices alongside a credit and asset boom, they said, was the blind spot rather than the achievement. It was treated as a curiosity.

The obvious objection is that these models are built on short-term interest rates rather than long bond yields, and they already allow for the extra return investors demand when lending long. Pushing down the price of long money is not the same as moving the number the models report.

The difficulty sits elsewhere.

What the model could not see

The estimate rests on two relationships.

The first is the IS curve. It says that when the real interest rate sits above the neutral rate, output falls below what the economy could produce. The name is a fossil, left over from a 1937 diagram interpreting Keynes, where the letters stood for investment and saving. It tells you nothing useful about how the curve is used now, so do not spend an afternoon on it as I once did. Read it as the equation running from interest rates to output.

The LM curve that used to sit alongside it has been dropped, replaced by a rule describing what the central bank does. So the model now has three equations, two describing the economy and one describing the bank.

The second is the Phillips curve. It says that when output falls below what the economy could produce, inflation slows.

Together they let the filter work backwards. Watch how output and inflation behave, and infer the rate that would have left both undisturbed.

Both went quiet through the 2010s, for different reasons, and the method needs both.

The IS curve runs from the real interest rate straight to the output gap. But transmission in that decade ran through asset values. Cheap money lifted the price of houses, shares and commercial property, and the spending that followed came from people who felt wealthier or could borrow against something now worth more. The equation has no term for any of that. So the direct effect it measures comes out small, not because interest rates had stopped working but because the road they travelled is not on the map.

I have tried to measure how little is left. Running the standard model on Australian data in its canonical form, the interest rate coefficient comes out near 0.05, meaning a one percentage point change in the real rate moves output by a twentieth of a percentage point. That is the equation the whole estimate hangs from.

The Phillips curve failed twice, and the two failures are different in kind. 

The first is simultaneity. A bank targeting inflation forecasts the shocks that would move it and offsets them before they arrive, and where one arrives unforecast it moves to cut it short. But it does not treat all shocks alike. Demand shocks it offsets, because doing so steadies prices and output together. Supply shocks it looks through, because anchored expectations let it. So the variation that would identify the curve is removed, and the variation that corrupts it is kept. Supply shocks push prices up while pushing output down, which is the opposite of the relation being estimated. An estimator handed a sample where that is most of what remains will report close to nothing. And the anchor does its own damage, because expectations that do not move cannot carry slack into prices. A bank doing its job well leaves behind data arguing that the job cannot be done.

The second follows from the first. Because inflation was held near target, the economy sat on a narrow stretch of the curve for fifteen years. The relationship between spare capacity and inflation is not a straight line. It bends, and the bend sits out at ranges the data almost never visited. Across the stretch that was visited, the curve is close to flat. An estimator needs variation in the response to identify anything. Across that stretch there was little, and most of what there was, was noise, so the slope came back near zero with a comfortable standard error around it.

The model did not report that it had nothing to work with. It reported that the relationship was weak, which is a claim about the whole curve drawn from the one section visible. In 2021 the steep part turned out to have been there the entire time.

With both equations that quiet, the model cannot do what it was built to do. On the same Australian data it fails to identify the neutral rate at all. The latent variable collapses toward zero, which is what an estimator does when the data will not tell it what it has been asked to find. The working is here, and anyone who wants to tell me I did it wrong is welcome to.

The method that produces the number central banks quote does not produce a number at all on the data of a medium-sized open economy with good statistics and a long series.

The estimate followed the policy

When the two equations describing the economy go quiet, the only one still carrying information is the one describing the central bank. The filter leans on the history of the rates actually set because by then there is nothing else in the model to lean on. And that history had been pushed down and held there for a decade.

One economist describes the result as the central bank staring at its own shadow. The Federal Reserve's own committee revised its view of the long-run rate from 4.25% in 2012 to 2.5% by 2019, then reversed after rates rose. That is officials moving with the policy they were setting. The models did the same thing for a reason you can point at, because the history of the policy rate is one of their inputs.

The pandemic made it obvious, though not in the way you would expect. The New York Fed suspended publication of its estimates in late 2020 because the shock broke the model. When it resumed nearly three years later, the filter reported that essentially nothing had changed. The American neutral rate came back at 0.8% for the end of 2022 against 0.9% for the end of 2019. The authors said their estimates sat within a few tenths of a percentage point of where they had been, and that they found no evidence the era of historically low natural rates had ended.

Consider what the model was asked to absorb. The deepest contraction since the war, the largest fiscal transfer in peacetime, the largest inflation in forty years, five percentage points on the policy rate and a violent move up in market measures of the real rate. The answer came back unchanged to a rounding error.

That is not damning on its own. A genuinely structural quantity should sit still through shocks that are genuinely temporary, and reporting stability is one of the things the model is built to do. The question is where the stability came from.

The model that reported no change was not the model that had been suspended. The 2023 version does two new things. It scales up the assumed variance of the shocks for 2020, 2021 and 2022, which lets the filter downweight the pandemic quarters rather than discard them. And it adds a COVID term to potential output, driven by the Oxford index of government restrictions, so that lost output is attributed to lockdowns rather than to the trend.

Both repairs are sensible. Without the first, a handful of extreme quarters would dominate everything. Without the second, the filter would read closed shops as a collapse in the economy's capacity. Anyone working with this data in 2020 made choices of the same kind, because the alternative was to understand nothing at all.

But follow the chain. In this model the neutral rate is built on trend growth. Protect trend growth from the pandemic and you have largely protected the neutral rate from it. The finding that the neutral rate was unchanged is not independent of the decision to send the pandemic somewhere else. It follows from it.

Adjusting a model to fit the data is how the work gets done. It is also the easiest way to confirm what you already believed.

The uncertainty bands swallow the question anyway. Recent New York Fed research puts investor uncertainty about the neutral rate at around plus or minus 1.7 percentage points, when policy decisions arrive in quarter-point increments.

The global market compounds it. If the price of long money is set globally, national estimates of the neutral rate inherit that price whether or not they intend to. Australia's estimate is not driven by Australian thrift and Australian investment opportunities alone. It is driven by what capital can earn anywhere, adjusted for currency risk.

So every country's estimate reads a global price, and that global price was being held down by the largest central banks. A model run on Australian data, using Australian output and Australian inflation, will faithfully report a lower neutral rate that was produced substantially in Washington and Tokyo. It is not measuring an Australian structural shift. It is measuring an imported price, and it cannot tell the difference.

The explanation kept changing

Ben Bernanke's original explanation in 2005 was a global savings glut. Excess savings from Asia and the oil exporters were flowing into American assets and holding long rates down. It was a good argument for its moment. China's current account surplus was heading toward 10% of its economy, oil exporters were recycling their earnings, and foreign official buying of American government debt was large and visible.

Then the ingredients disappeared. China's surplus fell to somewhere between 1.5 and 2.5% for most of the 2010s, and briefly to near zero in 2018 and 2019. Global imbalances narrowed sharply after 2008. Oil exporters stopped recycling for the rest of the decade after prices crashed in 2014. And foreign central banks stopped adding. Their holdings of American government debt have been roughly flat in dollar terms since 2013, while the debt itself has tripled.

The cause shrank and the effect intensified.

So the explanation was replaced. First secular stagnation, a chronic shortage of demand. Then demographics, an ageing world saving more. Then a shortage of safe assets rather than a surplus of savings.

Each is arguable, and a succession of explanations does not prove any of them wrong. What it does show is that as each structural story weakened, another took its place while the estimated neutral rate barely moved. The estimate was more stable than the case for it.

The safe asset argument is not really a rival. If the problem was a shortage of safe assets, then every government bond a central bank removed from private hands made that shortage worse. It is a channel through which quantitative easing worked, not an alternative to it.

Which is why rules stopped helping

Central banks are often criticised for abandoning the Taylor rule, which sets the policy rate as a function of inflation and the output gap. Nobody was following it. And the rule takes the neutral rate as an input, so feed it an estimate derived from the policy rate you already set and it recommends roughly the policy you already ran. It appears vindicated because the argument is circular.

There is a second version of the same trap, and it is the Phillips curve failure described earlier. The neutral rate estimate is corrupted because the policy rate is one of its inputs. The Phillips curve is corrupted because the bank was offsetting the shocks that would have revealed it.

Two of the framework's central quantities, corrupted by the same mechanism. Policy changes the data from which policymakers then try to work out what would have happened without the policy.

Estimation needs variation, and stabilisation exists to remove variation. The better a central bank gets at holding the economy steady, the less the data can tell anyone about the system underneath. That is a cost of success rather than an argument against it, and it is in the same family as the Lucas critique, though the mechanism differs. Lucas is about a change of regime altering the relationships you measured under the old one. This is about a bank that is doing its job well enough to erase the evidence it needs to keep doing it.

None of which excuses the narrowness. Losing identification is the price of stabilising an economy and nobody chose it. Calibrating the whole apparatus against consumer prices alone was chosen, and it was chosen repeatedly, over two decades of people at the Bank for International Settlements saying it was a mistake.

Where we are now

The QE programme has been unwinding. Balance sheet reduction ran from 2022 until the Federal Reserve ended it in December 2025, and the American central bank spent much of 2026 buying short-term government debt again before its new chairman stopped that in August.

Something new arrived on the other side. The artificial intelligence build-out has created a large demand for real investment capital, with capital spending by the largest American hyperscalers running near seven hundred billion dollars this year, and global AI-related bond issuance forecast by Morgan Stanley to reach about five hundred and seventy billion, more than double last year.

If any of the recent rise in yields reflects a genuinely higher neutral rate rather than the removal of a suppression, this is the most plausible candidate. Capital having somewhere productive to go is the oldest reason for interest rates to rise. Whether it is large enough to shift the underlying rate, as against merely showing up in market prices, is subject to exactly the measurement problem this piece has been describing, and will not be settled for some years.

The American ten-year government bond yields about 4.7% and the Australian one about 5.0%. American yields ran above that level for most of the years between 1995 and 2001. The unweighted average across the nations in the chart above sits near 3.6%, which is roughly where it sat in 2005.

The strange period was not this one. It was the decade when money was nearly free, and we have spent that decade telling ourselves the strangeness was structural.

What this does not prove

None of this says quantitative easing was only an asset price policy. It lowered borrowing costs, and lower borrowing costs support spending and employment through the ordinary channels. Those channels worked. The problem is that the estimation method cannot separate a structural fall in the neutral rate from a long policy intervention operating through the same channels, because both leave the same footprint in the data.

I think QE was worth doing. Against the counterfactual of a far deeper recession, the benefits look larger than the costs, and the costs were real. A deep recession would have done most harm to those with the least education, the lowest incomes and the least wealth. That is a judgement rather than a measurement. Nobody gets to run the other world and check.

The failure was not the decision. It was the candour. The distributional arithmetic was known, and the Bank of England published it in 2012, but it sat in the technical literature rather than forming part of the case put to the public. A policy that lifts the wealth of asset holders in order to prevent something worse is defensible. It is more defensible when the people who own no assets are told that is the trade.

The neutral rate may well have fallen too. Demographics are real, and an older world with fewer workers does behave differently. The safe asset shortage is real. Investment demand genuinely was weak for years after 2008.

Both things happened. Structural forces pushed the neutral rate down, and the policy response pushed the measured estimate down further, and the method used to estimate it cannot separate the two because the policy rate is an input to the calculation.

Two kinds of balance

The trouble is not that the neutral rate is hard to measure. It is that the question has been asked too narrowly for the answer to mean what people take it to mean.

Start with the estimates we have. For the second quarter of 2024 the Richmond Fed put three published figures side by side. Lubik-Matthes gave 2.6%, Johannsen-Mertens gave 1.8%, Holston-Laubach-Williams gave 0.8%. Nearly two percentage points of disagreement about the same quarter of the same economy, from methods built by serious people working with the same data. Policy decisions are made in quarter-point steps.

Now add the variables the standard model leaves out. Juselius, Borio, Disyatat and Drehmann did exactly that at the Bank for International Settlements, building the same kind of filter with leverage and debt service included, so the rate it reports is the one at which borrowing is also in balance. Their estimate declines over thirty years like everyone else's. What differs is the verdict. Actual policy rates sat below their neutral rate for almost the whole period, while the conventional estimate sits below theirs and eventually turns negative.

Same data, same countries. One model says policy was roughly neutral for twenty years. The other says it was loose for twenty years. The difference is entirely in what each was allowed to look at.

One number is being asked to describe two kinds of balance. There is a rate at which consumer prices hold steady, and a rate at which borrowing stays within the capacity to service it, and for most of the postwar period they sat close enough together that nobody had to choose between them.

The second condition is not a target for asset prices, and nobody sensible wants share prices frozen or credit growing at zero. Credit rising with nominal income is healthy and leverage can rise for good reasons. The condition is about whether the debt being taken on can be carried, which is measurable without anyone having to rule on what a share ought to be worth.

The two had been drifting apart before the crisis, which is what Borio and White were warning about. QE turned the gap into a chasm. The money largely stopped reaching the people who buy goods, and weak investment demand, private deleveraging and fiscal restraint all pushed the same way. The standard neutral rate models can see only the first of the two conditions.

That is not a measurement problem. It is a specification problem, and it is worse, because a measurement problem announces itself with wide error bands while a specification problem returns a confident number to the wrong question.

The rate is defined against potential output, which is also unobservable and also estimated from the same data. There is no single neutral rate across the term structure, only a family of them, and the models report one figure. Wicksell's rate, the one that equates saving with investment in real capital, is a different object from the flexible-price rate of the modern models, and the two get used interchangeably. And if demand shapes supply over time, which is what the productivity record after 2008 suggests, then potential output is partly an outcome of policy, and the anchor moves when you lean on it.

I do not think those difficulties are fatal. I do think they mean the estimate cannot bear the weight that has been put on it.

What to watch instead

If the estimate answers only half of the question, what do you use for the other half?

Things you can see. Credit growth. Lending standards and the spread between what banks charge and what they pay. Debt service relative to income. The exchange rate. Asset prices and what people are borrowing to buy them. Where new investment is going, and whether it is going into building anything.

None of that is a number you can drop into a rule. All of it was flashing through the 2010s, and none of it entered the estimate that policy was being judged against.

Independent targets generally need independent instruments. One interest rate can serve both conditions while they line up, which is what happened for most of the postwar period and is why the problem stayed hidden. When they diverge, no improvement in estimating either one closes the gap, and it has to be closed by a second instrument. That is the case for macroprudential policy arrived at the long way round, from measurement rather than from ideology.

The Reserve Bank is closer to this than the Federal Reserve, less because its methods are better than because it declines to put a single figure at the centre of the argument. It runs several approaches, reports a range, and its officials say plainly that the range is wide. When I could not identify the rate on Australian data, I fell back on triangulating between trend growth and the indexed bond yield, which gives a region rather than a point. That is not a worse answer. It is an honest description of what the data supports.

The models were not wrong about the world. They were answering a narrower question than anyone reading them believed, and they were answering it with the policy rate as an input. That leaves judgement, exercised in public, with the evidence shown. It is a less satisfying place to end than a number. It is also the honest one.