TL;DR: The nominal neutral rate of interest is around 4.6 per cent. Probably.
The cash rate is at 4.10 per cent. Half the commentary calls this restrictive, the other half calls it appropriate, and a vocal minority calls it inadequate. Almost everyone making these claims is implicitly comparing the current rate to neutral. Almost no one explains where their estimate of neutral comes from.
This post is an attempt to be honest about what we can and cannot say about neutral, using work I have done.
The short version. The post-GFC decade was unusual on multiple dimensions, including unusually low equilibrium real rates. The post-COVID period has seen a partial return to more familiar dynamics. Multiple methodologies, including a careful estimation I have built and tested, place the current cash rate within the range of plausible neutral estimates. The level is not extreme by historical standards, even though intuitions formed during the 2010s might suggest otherwise.
The longer version requires some setup.
The IS curve intercept
The natural rate of interest, r-star, is the textbook intercept of the IS curve at potential output. It is the real rate at which the economy settles at sustainable capacity, neither overheating nor running below. Add the inflation target and you get the nominal neutral cash rate.
The concept goes back to Knut Wicksell, writing in the late 1890s, who distinguished between the market rate of interest set by financial conditions and a natural rate determined by the real productivity of capital. When the market rate sits below the natural rate, desired investment exceeds desired savings, demand runs above capacity, and inflation rises. When the market rate sits above the natural rate, the reverse happens. The natural rate is the rate that clears the savings-investment market at full employment.
The picture above sketches the geometry. Output (GDP) sits on the horizontal axis, the real interest rate on the vertical. The IS curve slopes down because lower real rates support more output through investment and consumption decisions. R-star, the natural rate, sits where the curve crosses potential output (the speed limit of the economy). When the actual real rate sits above r-star, the economy is being held below capacity and inflation pressures ease. When it sits below r-star, the economy is being pushed above capacity and inflation pressures build. The whole curve shifts over time as the economy's productive capacity, demographics, and financial structure change. Identifying where r-star currently sits is the empirical problem this post addresses.
The supply-side reading of Wicksell takes the natural rate to be anchored by the real return on capital. In a competitive economy at full employment, the marginal product of capital determines the equilibrium real return that savers can earn and investors are willing to pay. That return cannot drift far from the underlying productivity of the capital stock without setting up arbitrage that pulls it back. Over long horizons, the real return on capital is bounded by the rate at which the productive capacity of the economy grows. Faster trend growth supports a higher equilibrium real rate. Slower trend growth, whether from demographic transition, productivity slowdown, or capital deepening, drags it down.
This is why trend growth appears as the structural anchor in the work that follows. It is not a free-standing empirical regularity. It is a proxy for the supply-side determination of the natural rate that Wicksell pointed to. The decline in Australian trend growth from around 3.6 per cent in the mid-1990s to around 2.4 per cent now is, on this reading, the visible signature of underlying changes in the productive capacity of the economy that should be pulling r-star down at roughly the same pace.
The market reading complements this rather than competing with it. Real bond yields measure where actual real rates settle when financial markets clear. Strip out the term and risk premia and what remains is the market's collective best guess about average future real short rates, which over long horizons should converge to the natural rate. The market reading is contaminated by everything that affects bond yields beyond the natural rate, including risk premia, liquidity preferences, and central bank actions. But it is direct in a way the structural reading is not, because it incorporates real-time information about how rates are actually clearing markets.
Wicksell himself doubted whether the natural rate could be measured precisely. The modern empirical apparatus has not fundamentally changed that conclusion, even with computational tools he could not have imagined. What it has done is make the limits of measurement visible, which is what the rest of this post is about.
The intercept is not directly observable. We infer it from how the economy has behaved at different real interest rates over time, while accounting for everything else that affects output and inflation. That is the empirical problem r-star estimation tries to solve, and it has proven remarkably hard.
The post-GFC decade was unusual
For most of the inflation-targeting era from 1993, the relationship between unemployment and inflation in Australia behaved roughly as the textbook predicted. From around 2010 to 2019, unemployment fluctuated between 5 and 6 per cent and inflation barely budged. The Phillips curve appeared to have flattened. The same flattening happened in the US, the UK, and the euro area.
The flat Phillips curve was one symptom of a broader strangeness. Real interest rates fell across the developed world to levels structural models had not predicted. Asset prices ran above what conventional valuation models suggested. Bond yields compressed in ways that strained standard open economy macro. No fully agreed-upon explanation has emerged. Secular stagnation in the Summers framing offers one unified account. Demographic and savings-glut stories offer others. None has settled the question.What this meant for r-star estimates is that they fell sharply. Models that tried to identify the equilibrium real rate consistent with full employment kept producing lower numbers, partly because trend productivity slowed, partly because demographic shifts raised desired saving relative to investment, and partly because of factors that none of the standard explanations fully captured. By the late 2010s, mainstream r-star estimates for advanced economies were near zero or even negative, and policy commentary built up around the assumption that this was the new normal.
The post-COVID period has been different
Inflation surged to levels not seen in decades. Central banks tightened sharply. The flat Phillips curve reasserted itself with much more recognisable dynamics. Unemployment fell, inflation responded, central banks acted, inflation began to ease. The post-2010 anomalies started looking less like permanent features of the new economy and more like consequences of an unusual period.
The NAIRU estimation I have built tells a related story. Australian NAIRU drifted down through the 2010s as the flat Phillips curve allowed lower unemployment without inflation pressure. It has risen modestly post-COVID to around 4.84 per cent. The labour market has been operating below NAIRU for the past few years, consistent with the inflation pressure we saw and with the policy response that followed.
If r-star fell to unusually low levels during the post-GFC regime, it should be drifting back to more normal levels as the regime ends.
The estimation problem
The standard approach in the international literature is the Holston-Laubach-Williams model, a state-space framework that tries to identify r-star jointly with potential output and trend growth. It is the workhorse of modern r-star estimation and what most central banks use when they cite an r-star figure.
When I implemented HLW for Australian data in its canonical form, the model failed to identify r-star. The latent variable that is supposed to capture independent movement in the equilibrium real rate collapsed toward zero. R-star mostly tracked long-run trend growth, with no useful independent variation.
The flat line with very wide credible bands is identification failure rather than a confident finding of stability. The model is reporting that it cannot tell the latent r-star apart from trend growth, and the bands reflect the lack of information rather than genuine uncertainty about a well-identified object.
This matches what other researchers have found for similar small open economies. McCririck and Rees in the RBA Bulletin in September 2017 found r-star somewhere in the 0.5 to 1.5 per cent real range using model averaging, explicitly because no single specification produced confident estimates. Luci Ellis at the RBA in October 2022 reported nine model estimates ranging from negative 0.5 to positive 2 per cent, in a speech titled "The Neutral Rate: The Pole-star Casts Faint Light." Both pieces are worth reading in full and reach similar conclusions to the work here, though through different methodological routes.
The obvious fix is to bring in more identifying information. The most commonly suggested source is the inflation-linked bond yield, which directly measures expected future real interest rates. When I added the indexed bond as an observation of r-star, the model identification improved but the resulting estimates became implausible. R-star ended up tracking the bond yield minus a constant term premium, swinging from above 4 per cent in the 1990s to a trough below negative 1 per cent in 2022.
The negative trough is a giveaway that something is wrong. Real term premia compressed substantially during the post-GFC period. The constant term premium assumption pushed all of that compression into r-star itself rather than into the premium. R-star becomes a relabelling of the bond yield, with the IS and Phillips curves doing little identifying work.
A third approach works better. Instead of treating the bond yield as another observation that has to be reconciled with a freely moving latent r-star, build it into the definition of r-star directly. Let r-star be a weighted blend of two anchors, trend growth on one side and the bond yield minus a term premium on the other, with the data choosing the weight. Two observable anchors, one scalar weight, no unidentified random walks.
The model converges and the estimates are credible. R-star comes in around 4 per cent in the mid-1990s, falls through the post-GFC period to a trough around 0.5 per cent in 2022, and has recovered to around 2.1 per cent at the end of 2025. The 90 per cent credible band runs from roughly 1.4 to 2.7 per cent at the latest point, wider through the middle of the sample. The trajectory is robust across reasonable specification choices and the level is consistent with the McCririck-Rees range, the Ellis range, and the most recent IMF Article IV estimate.
The technical detail, including specifications I tried and discarded, is in the model notes.
A finding about Australian transmission
One result is worth highlighting because it is robust across all three specifications. The IS curve coefficient on the real interest rate gap, the parameter that captures how strongly rate changes affect output through the standard textbook channel, sits at around negative 0.05 in every specification I tried. That is small. The rate channel through the IS curve is genuinely weak in Australian data, regardless of how r-star is identified.
This is consistent with existing work on Australian monetary transmission, including research that emphasises housing, exchange rates, and cash flow channels rather than the standard interest-rate-to-investment story. The implication for the modelling is direct. If the IS curve has limited leverage on output through real rates, then no r-star estimation methodology can extract a confidently identified r-star from Australian macro data alone. This is a feature of the Australian economy, not a limitation of any particular specification. The blend approach is the response. It uses external observable anchors to do the identification work that the IS curve cannot do on its own.
The honest description of what the blend produces is not "the IS curve intercept estimated directly" but "a constrained estimate of r-star that uses observable anchors where the IS curve cannot identify it on its own." The conceptual target remains the IS intercept that Wicksell pointed to. The empirical strategy is to discipline that target with structural and market signals that are independently observable.
What the data says about the anchors
The most informative single result from the blend specification is the posterior on the weight parameter. The weight on trend growth versus the bond anchor could in principle be anywhere between zero and one. The data has a posterior centred at 0.51 with a 90 per cent credible interval running from 0.18 to 0.86.
The data does not strongly discriminate between structural and market anchoring. Any r-star estimate that relies exclusively on one anchor or the other is imposing more structure than the data supports. This applies to estimates based purely on trend growth, which lean entirely on the structural reading. It applies equally to estimates that read r-star directly off bond yields, which lean entirely on the market reading. Both approaches have empirical content. Neither dominates. The honest answer is a blend, with the weighting itself uncertain.The decomposition
Trend growth, the structural anchor, has fallen smoothly from around 3.6 per cent in the mid-1990s to 2.4 per cent now. This is the supply-side story Wicksell pointed to, with demographic transition, slowing productivity, and capital deepening all reducing the real return on capital that the natural rate must equilibrate to. The bond anchor, derived from real bond yields minus a term premium, has been much more volatile. The term premium of 0.65 percentage points is estimated jointly with the rest of the model rather than calibrated. Term premia obviously vary over time, and I tried specifications that let the premium move with the slope of the yield curve, but those specifications did not cohere. The static estimate is the disciplined fallback. The bond anchor inherits some sensitivity to that simplification, particularly in periods of rapid term premium movement, but the substantive r-star path is robust to the choice. It hit nearly 5 per cent in the mid-1990s, fell into negative territory through 2020 and 2021 to a trough around negative 1.3 per cent, and has snapped back to around 1.7 per cent. R-star sits between the two, taking level discipline from the structural anchor and dynamics from the market anchor.
This is r-star estimated honestly. Not as a single confident number, and not as a direct read of the IS curve intercept, but as a constrained blend of two observable signals with explicit uncertainty about the weighting. The current value sits at around 2.1 per cent real, which translates to a nominal neutral cash rate of around 4.6 per cent assuming the 2.5 per cent inflation target.
Methodologies converging
Worth pausing on this before turning to policy. The current r-star estimate of around 2.1 per cent real is close to where the structural anchor would put it on its own, close to where the market anchor would put it on its own, and close to where a separate deterministic Cobb-Douglas calculation from my NAIRU model's trend potential growth would put it. The HLW estimate of trend growth here is modestly higher than the NAIRU model's Cobb-Douglas calculation, by around 20 basis points, which is the kind of small disagreement you expect when different models estimate the same latent object using different identifying information. The methodologies are not reading off the same instrument and small gaps are normal. What is striking is how close they end up.
During the post-GFC decade these methodologies pointed at very different numbers. Bond-anchored estimates ran much lower than structural ones. The data could not pick between them, and any choice of approach was contested. Call this period the Great Divergence. The scepticism about r-star that built up during it was an honest response to the empirical situation. Confident estimates were not available, and any methodology you chose carried identifying assumptions that another methodology would dispute.
The current convergence is itself informative. It says the structural and market readings of Wicksell, which diverged sharply in the unusual conditions of the 2010s, have largely re-aligned in the post-COVID period. That is what you would expect if the post-GFC regime really has substantially passed. The implicit invitation here is to take the position that r-star is useful in the abstract, hard to estimate during certain periods, and currently in a more tractable empirical state than it was during the Great Divergence. Whether the tractability holds is an open question. Convergence today does not guarantee convergence tomorrow. But the conditions that made strong scepticism the right response during the post-GFC decade have eased, and the empirical case for treating r-star as a usable directional anchor is stronger now than it was five years ago.
What this says about current policy
The cash rate is 4.10 per cent. The current real cash rate, calculated using inflation expectations rather than realised inflation, is around 0.9 per cent. The model's estimate of r-star is 2.1 per cent, with a 90 per cent credible band running from 1.4 to 2.7 per cent at the latest point.
The real cash rate sits below the lower bound of the r-star credible interval. That points toward policy being on the accommodative side of neutral rather than restrictive. The reading is directional rather than precise. Using realised inflation instead of inflation expectations would give a less accommodative-looking real cash rate, because realised inflation has been above expectations. And the snapshot does not capture the cumulative effect of past tightening still working through the economy through transmission lags. A balanced reading is that current policy is somewhere between mildly accommodative and roughly neutral, with the lagged effects of past tightening still dragging on activity.
The historical comparison points the same way and is consistent with the model. A nominal cash rate of 4.10 per cent is not unusual by the standards of the inflation-targeting era. Most of the period between 1993 and 2008 saw cash rates between 4 and 7 per cent. That historical range sits comfortably inside the model-consistent neutral range of around 4 to 5 per cent nominal. Both the historical record and the formal estimation point at the same place, which is that current rates are unremarkable. The post-GFC decade was the outlier. Anyone whose intuition about normal rates was formed between 2010 and 2020 is working from an unrepresentative sample.
The honest position
R-star is useful as a concept. The IS curve intercept is something policy needs to be able to think about, and the Wicksellian natural rate, anchored in the supply-side determination of the real return on capital, has a long intellectual lineage that produces real insight. Abandoning the concept entirely would leave a hole that the alternatives do not fill.
R-star is also notoriously hard to estimate. The standard methodology fails identification on Australian data. Adding obvious sources of identifying information produces implausible estimates. The principled middle ground, the blend approach, gives credible estimates with credible bands wide enough to span most of the policy-relevant range. Anyone claiming a precise r-star number is overclaiming what the data supports.
Macroeconomics is full of latent objects we cannot directly measure but that organise how we think about the economy. R-star, u-star, y-star, pi-star, the natural rate of interest, NAIRU, potential output, underlying inflation. None of them is observable. All of them are estimated with substantial uncertainty. All of them are essential to coherent policy thinking. The scepticism some have about r-star applies, in different degrees, to all of them. If you reject one on the grounds of identification difficulty, consistency requires rejecting most of macroeconomics.
The work is still worth doing. Even with wide credible bands, the model rules out a lot. It rules out current rates being extremely restrictive on any sensible reading. It rules out them being extremely loose. It tells you the post-GFC decade really was unusual, and the post-COVID period really has seen a return toward more familiar dynamics. It supports the historical comparison, which says current rates are unremarkable by the standards of the inflation-targeting era as a whole.
That is what the model can deliver. Not a precise calibration of policy to a target gap. A useful directional judgement supported by transparent uncertainty.
Stating the conclusion in nominal terms, this work suggests a neutral cash rate of around 4.6 per cent. That is the 2.1 per cent real estimate plus the 2.5 per cent inflation target. The credible band runs from roughly 3.9 to 5.2 per cent nominal. The current cash rate of 4.10 per cent sits inside that range, slightly below the central tendency.
The cash rate at 4.10 per cent is not extreme. It sits within the range of plausible neutral estimates, slightly below the central tendency of those estimates depending on how you measure the real rate, and well within the band of historical experience. Whatever is going wrong or right with the Australian economy at the moment, it is not because the cash rate is in unprecedented territory.
The headline claim that 4.10 per cent is not extreme survives this analysis, even after the wide uncertainty around r-star is properly acknowledged.
No comments:
Post a Comment