The Persistence of Demand

Here’s the very short version of this very long post:

Hysteresis means that a change in GDP today has effects on GDP many years in the future. In principle, this could be because it affects either future aggregate demand or potential output. These two cases aren’t distinguished clearly in the literature, but they have very different implications. The fact that the Great Recession was followed by a period of low inflation, slow wage growth and low interest rates, rather than the opposite, suggests that the persistent-demand form of hysteresis is more important than potential-output hysteresis. The experience of the Great Recession is consistent with perhaps 20 percent of a shock to demand in this period carrying over to demand in future periods. This value in turn lets us estimate how much additional spending would be needed to permanently return GDP to the pre-2007 trend: 50-60 percent of GDP, or $10-12 trillion, spread out over a number of years.

 

Supply Hysteresis and Demand Hysteresis

The last few years have seen renewed interest in hysteresis – the idea that shifts in demand can have persistent effects on GDP, well beyond the period of the “shock” itself. But it seems to me that the discussion of hysteresis doesn’t distinguish clearly between two quite different forms it could take.

On the one hand, demand could have persistent effects on output because demand influences supply – this seems to be what people usually have in mind. But on the other hand, demand itself might be persistent. In time-series terms, in this second story aggregate spending behaves like a random walk with drift. If we just look at the behavior of GDP, the two stories are equivalent. But in other ways they are quite different.

Let’s say we have a period in which total spending in the economy is sharply reduced for whatever reason. Following this, output is lower than we think it otherwise would have been. Is this because (a) the economy’s productive potential was permanently reduced by the period of reduced spending? Or is it (b) because the level of spending in the economy was permanently reduced? I will call the first case supply hysteresis and the second demand hysteresis.

It might seem like a semantic distinction, but it’s not. The critical thing to remember is that what matters for much of macroeconomic policy is not the absolute level of output but the output gap — the difference between actual and potential output. If current output is above potential, then we expect to see rising inflation. (Depending on how “potential” is understood, this is more or less definitional.) We also expect to see rising wages and asset prices, shrinking inventories, longer delivery times, and other signs of an economy pushing against supply constraints. If current output is below potential, we expect the opposite — lower inflation or deflation, slower wage growth, markets in general that favor buyers over sellers. So while lower aggregate supply and lower aggregate demand may both translate into lower GDP, in other respects their effects are quite different. As you can see in my scribbles above, the two forms of hysteresis imply opposite output gaps in the period following a deep recession. 

Imagine a hypothetical case where there is large fall in public spending for a few years, after which spending returns to its old level. For purposes of this thought experiment, assume there is no change in monetary policy – we’re at the ZLB the whole time, if you like. In the period after the depressed spending ends, will we have (1) lower unemployment and higher inflation than before, as the new income created during the period of high public spending leads to permanently higher demand. Or will we have (2) higher unemployment and lower inflation than if the spending had not occurred, because the period of high spending permanently raised labor force participation and productivity, while demand returns to its old level?

Supply hysteresis implies (1), that a temporary negative demand shock will lead to persistently higher inflation and lower unemployment (because the labor force will be smaller). Demand hysteresis implies (2), that a temporary negative demand shock will lead to permanently lower inflation and higher unemployment. Since the two forms of hysteresis make diametrically opposite predictions in this case, seems important to be clear which one we are imagining. Of course in the real world, could see a combination of both, but they are still logically distinct. 

Most people reading this have probably seen a versions of the picture below. On the eve of the pandemic, real per-capita GDP was about 15 percent below where you’d expect it to be based on the pre-2007 trend. (Or based on pre-2007 forecasts, which largely followed the trend.) Let’s say we agree that the deviation is in some large part due to the financial crisis: Are we imagining that output has persistently fallen short of potential, or that potential has fallen below trend? Or again, it might be a combination of both.

In the first case, we would expect monetary policy to be generally looser in the period after a negative demand shock, in the second case tighter. In the first case we’d expect lower inflation in period after shock, in the second case higher.

It seems to me that most of the literature on hysteresis does not really distinguish these cases. This recent IMF paper by Antonio Fatas and coauthors, for example, defines hysteresis as a persistent effect of demand shocks on GDP. This could be either of the two cases. In the text of the paper,  they generally assume hysteresis means an effect of demand on supply, and not a persistence of demand itself, but they don’t explicitly spell this out or make an argument for why the latter is not important.

It is clear that the original use of the term hysteresis was understood strictly as what I am calling supply hysteresis. (So perhaps it would be better to reserve the word for that, and make ups new name for the other thing.) If you read the early literature on hysteresis, like these widely-cited Laurence Ball papers, the focus was on the European experience of the 1980s and 1990s; hysteresis is described as a change in the NAIRU, not as an effect on employment itself. The mechanism is supposed to be a specific labor-market phenomenon: the long term unemployed are no longer really available for work, even if they are counted in the statistics. In other words, sustained unemployment effectively shrinks the labor force, which means that in the absence of policy actions to reduce demand, the period following a deep recession will see faster wage growth and higher inflation than we would have expected.

(This specific form of supply hysteresis implies a persistent rise in unemployment following a downturn, just as demand hysteresis does. The other distinctions above still apply, and other forms of supply hysteresis would not have this implication.) 

Set aside for now whether supply-hysteresis was a reasonable description of Europe in the 1980s and 1990s. Certainly it was a welcome alternative to the then-dominant view that Europe needed high unemployment because of over-protective labor market institutions. But whether or not thinking of hysteresis in terms of the NAIRU made sense in that context, it does not make sense for either Europe or the US (or Japan) in the past decade. Everything we’ve seen has been consistent with a negative output gap — with actual output below potential — with a depressed level of demand, not of supply. Wage growth has been unexpectedly weak, not strong; inflation has been below target; and central banks have been making extra efforts to boost spending rather than to rein it in.

Assuming we think that all this is at least partly the result of the 2007-2009 financial crisis — and thinking that is pretty much the price of entry to this conversation — that suggests we should be thinking primarily about demand-hysteresis rather than supply-hysteresis. We should be asking not, or not only, how much and how durably the Great Recession reduced the country’s productive potential, but how how durably it reduced the flow of money through the economy. 

It’s weird, once you think about it, how unexplored this possibility is in the literature. It seems to be taken for granted that if demand shocks have a lasting effect on GDP, that must be because they affect aggregate supply. I suspect one reason for this is the assumption — which profoundly shapes modern macroeconomics — that the level of spending in the economy is directly under the control of the central bank. As Peter Dorman observes, it’s a very odd feature of modern macroeconomic modeling that the central bank is inside the model — the reaction of the monetary authorities to, say, rising inflation is treated as a basic fact about the economy, like the degree to which investment responds to changes in the interest rate, rather than as a policy choice. In an intermediate macroeconomics textbook like Carlin and Soskice (a good one as far as they go), students are taught to think about the path of unemployment and inflation as coming out of a “central bank preference function,” which is taken as a fundamental parameter of the economy. Obviously there is no place for demand hysteresis in this framework. To the extent that we think of the actual path of spending in the economy as being chosen by the central bank as part of some kind of optimizing process, past spending in itself will have no effect on current spending. 

Be that as it may, it seems hard to deny that in real economies, the level of spending today is strongly influenced by the level of spending in the recent past. This is the whole reason we see booms and depressions as discrete events rather than just random fluctuations, and why they’re described with metaphors of positive-feedback process like “stall speed” or “pump-priming.”¹

How Persistent Is Demand?

Let’s say demand is at least somewhat persistent. That brings us to the next question: How persistent? If we were to get extra spending of 1 percent of GDP in one year, how much higher would we now expect demand to be several years later?

We can formalize this question if we write a simple model like:

Z_t = Z*_t + X_t

Z*_t = (1+g) Z*_t-1 + a(Z_t-1 – Z*_t-1)

Here Z is total spending or demand, Z* is the trend, what we might think of as normal or expected demand, g is the normal growth rate, and X is the influence of transitory influences outside of normal growth.

With a = 0, then, we have the familiar story where demand is a trend plus random fluctuations. If we see periods of above- and below-trend demand, that’s because the X influences are themselves extended over time. If a boom year is followed by another boom year, in this story, that’s because whatever forces generated it in the first year are still operating, not because the initial boom itself was persistent. 

Alternatively, with a = 1, demand shocks are permanent. Anything that increases spending this year, should be expected to lead to just as much additional spending next year, the year after that, and so on.

Or, of course, a can have any intermediate value. 

Think back to 2015, in the debate over the first Sanders’ campaign’s spending plans that was an important starting point for current discussions of hysteresis. The basic mistake Jerry Friedman was accused of making was assuming that changes in demand were persistent — that is, if the multiplier was, say 1.5, that an increase in spending of $500 billion would raise output by $750 billion not only in that year and but in all subsequent years. As his critics correctly pointed out, that is not how conventional multipliers work. In terms of my equations above, he was setting a=1, while the conventional models have a=0. 

He didn’t spell this out, and I didn’t think of it that way at the time. I don’t think anyone did. But once you do, it seems to me that while Friedman was wrong in terms of the standard multiplier, he was not wrong about the economy — or at least, no more wrong than the critics. It seems to me that both sides were using unrealistically extreme values. Demand shocks aren’t entirely permanent, but they also aren’t entirely transitory.  A realistic model should have 0 < a < 1.

Demand Persistence and Fiscal Policy

There’s no point in refighting those old battles now. But the same question is very relevant for the future. Most obviously, if demand shocks are persistent to some significant degree, it becomes much more plausible that the economy has been well below potential for the past decade-plus. Which means there is correspondingly greater space for faster growth before we encounter supply constraints in the form of rising inflation. 

Both forms of hysteresis should make us less worried about inflation. If we are mainly dealing with supply hysteresis, then rapid growth might well lead to inflation, but it would be a transitory phenomenon as supply catches up to the new higher level of demand.  On the other hand, to the extent we are dealing with demand hysteresis, it will take much more growth before we even have to worry about inflation.

Of course, both forms of hysteresis may exist. In which case, both reason for worrying less about inflation would be valid. But we still need to be clear which we are talking about at any given moment.

A slightly trickier point is that the degree of demand persistence is critical for assessing how much spending it will take to get back to the pre-2007 trend. 

If the failure to return to the pre-2007 is the lasting effect of the negative demand shock of the Great Recession, it follows that  sufficient spending should be able to reverse the damage and return GDP to its earlier trend. The obvious next question is, how much? The answer really depends on your preferred value for a. In the extreme (but traditional) case of a=0, each year we need enough spending to fill the entire gap, every year, forever. Given a gap of around 12 percent, if we assume a multiplier of 1.5 or so, that implies additional public spending of $1.6 trillion. In the opposite extreme case, where a=1, we just need enough total spending to fill the gap, spread out over however many years. In general, if we want to get close a permanent (as opposed to transitory) output gap of W, we need W/(a μ) total spending, where μ is the conventional multiplier.²

If you project forward the pre-2007 trend in real per-capita GDP to the end of 2019, you are going to get a number that is about 15% higher than the actual figure, implying an output gap on the order of $3.5 trillion. In the absence of demand persistence, that’s the gap that would need to be filled each year. But with persistent demand, a period of elevated public spending would gradually pull private spending up to the old trend, after which it would remain there without further stimulus.

What Does the Great Recession Tell Us about Demand Persistence?

At this point, it might seem that we need to turn to time-series econometrics and try to estimate a value for a, using whatever methods we prefer for such things. And I think that would be a great exercise! 

But it seems to me we can actually put some fairly tight limits on a without any econometrics, simple by looking back to the Great Recession. Keep in mind, once we pick an output gap for a starting year, then given the actual path of GDP, each possible value of a implies a corresponding sequence of shocks X_t. (“Shock” here just means anything that causes a deviation of demand from its trend, that is not influenced by demand in the previous period.) In other words, whatever belief we may hold about the persistence of demand, that implies a corresponding belief about the size and duration of the initial fall in demand during the recession. And since we know a fair amount about the causes of the recession, some of these sequences are going to be more plausible than others.

The following figures are an attempt to do this. I start by assuming that the output gap was zero in the fourth quarter of 2004. We can debate this, of course,, but there’s nothing heterodox about this assumption — the CBO says the same thing. Then I assume that in the absence of exogenous disturbances, real GDP per capita would have subsequently grown at 1.4 percent per year. This is the growth rate during the expansion between the Great Recession and the pandemic; it’s a bit slower than the pre-recession trend.³ I then take the gap between this trend and actual GDP in each subsequent quarter and divide it into the part predictable from the previous quarter’s gap, given an assumed value for a, and the part that represents a new disturbance in that period. So each possible value of a, implies a corresponding series of disturbances. Those are what are shown in the figures.

If you’re not used to this kind of reasoning, this is probably a bit confusing. So let me put it a different way. The points in the graphs above show where real GDP would have been relative to the long-term trend if there had been no Great Recession. For example, if you think a = 0, then GDP in 2015 would have been just the same in the absence of the recession, so the values there are just the actual deviation from trend. So you can think of the different figures here as showing the exogenous shocks that would be required under different assumptions about persistence, to explain the actual deviation from trend. They are answering this question: Given your beliefs about how persistent demand is, what must you think GDP would have been in subsequent years in a world where the Great Recession did not take place? (Or maybe better, where the fall in demand form the housing bubble was fully offset by stimulus.)

The first graph, with persistence = 0, is easiest to understand. If there is no carryover of demand shocks from one period to the next, then there must be some factor reducing demand in each later period by the full extent of the gap from trend. If we move on to, say, the persistence=0.1 figure, that is saying that, if you think 10 percent of a demand shock is normally carried over into future periods, that means that there was something happening in 2012 that would have depressed demand by 2 percent relative to the earlier trend, even if there had been no Great Recession. 

Because people are used to overcomplicated economics models, I want to stress again. What I am showing you here is what you definitionally believe, if you think that in the absence of the Great Recession, growth in the 2010s would have been at about the same rate it was, just from a higher base, and you think that whatever fraction of a change in spending in one year is carried over to the next year. There are no additional assumptions. I’m just showing what the logical corollary of those beliefs would be for the pattern of demand shocks,

Another important feature of these figures is how large the initial fall in demand is. Logically, if you think demand is very persistent, you must also think the initial shock was smaller. If most of the fall in spending in the first half of 2008, say, was carried over to the second half of 2008, then it takes little additional fall in spending in that period to match the observed path of GDP. Conversely, if you think that very little of a change in demand in one period carries over to the next one then the autonomous fall in demand in 2009 must have been larger.

The question now is, given what we know about the forces impacting demand a decade ago, which of these figures is most plausible? If there had been sufficient stimulus to completely eliminate the fall in demand in 2007-2009, how strong would the headwinds have been a few years late? 

Based on what we know about the Great Recession, I think demand persistence in the 0.15 – 0.25 range most plausible. This suggests that a reasonable baseline guess for total spending required to return to the pre-2007 would be around 50 percent of GDP, spread out over a number of years. With an output gap of 15 percent of GDP, a multiplier of 1.5, and demand persistence of 0.2, we have 15 / (1.5 * 0.2) = 50 percent of GDP. This is, obviously, a very rough guess, but if you put me on the spot and asked how much spending over ten years it would take to get GDP permanently back to the pre-2007 trend, $10-12 trillion would be my best guess.

How do we arrive at persistence in the 0.15 – 0.25 range?

On the lower end, we can ask: What are the factors that would have pushed down demand in the mid 2010s, even in the absence of the Great Recession Remember, if we use demand persistence of 0.1, that implies there were factors operating in 2014 that would have reduced demand by 2 percent of GDP, even if the recession had not taken place. What would those be?

I don’t think it makes sense to say housing — housing prices had basically recovered by then. State and local spending is a better candidate — it remained quite depressed and I think it’s hard to see this as a direct effect of the recession. Relative to trend, state and local investment was down about 1 percent of GDP in 2014, while the federal stimulus was basically over. On the other hand, unless we think that monetary policy is totally ineffective, we have to include the stimulative effect of a zero policy rate and QE in our demand shocks. This makes me think that by 2014, the gap between actual GDP and the earlier trend was probably almost all overhang from the recession. And this implies a persistence of at least 0.15. (If you look back at the figures, you’ll see that with persistence=0.15, the implied shock reaches zero in 2014.)

Meanwhile, on the high end, a persistence of 0.5 would mean that the demand shock maxed out at a bit over 3 percent of GDP, and was essentially over by the second half of 2009. This seems implausibly small and implausibly brief. Residential investment fell from 6.5 percent of GDP in 2004 to less than 2.5 percent by 2010. And that is leaving aside housing wealth-driven consumption. Meanwhile, the ARRA stimulus didn’t really come online until the second half of 2009. I don’t believe monetary policy is totally ineffective, but I do think it operates slowly, especially on loosening side. So I find it hard to believe that the autonomous fall in demand in early 2009 was much less than 5 percent of GDP. That implies a demand persistence of no more than 0.25.

Within the 0.15 to 0.25 range, probably the most important variable is your judgement of the effectiveness of monetary policy and the ARRA stimulus. If you think that one or both was very effective, you might think that by mid-2010, they were fully offsetting the fall in demand from the housing bust. This would be consistent with  persistence around 0.25. Conversely, if you’re doubtful about the effectiveness of monetary policy and the ARRA (too little direct spending), you should prefer a value of 0.2 or 0.15. 

In any case, it seems to me that the implied shocks with persistence in the 0.15 – 0.25 range look much more plausible than for values outside that range. I don’t believe that the underlying forces that reduced demand in the Great Recession had ceased to operate by the second half of 2009. I also don’t think that they were autonomously reducing demand by as much as 2 points still in early 2014. 

You will have your own priors, of course. My fundamental point is that your priors on this stuff have wider implications. I have not seen anyone spell out the question of the persistence of demand in the way I have done here. But the idea is implicit in the way we talk about business cycles. Logically, a demand shortfall in any given period can be described as a mix of forces pulling down spending in that period, and the the ongoing effect of weak demand in earlier periods. And whatever opinion you have about the proportions of each, this can be quantified. What I am doing in this post, in other words, is not proposing a new theory, but trying to make explicit a theory that’s already present in these debates, but not normally spelled out.

Why Is Demand Persistent?

The history of real economies should be enough to convince us that demand can be persistent. Deep downturns — not only in the US after 2007, but in much of Europe, in Japan after 1990, and of course the Great Depression — show clearly that if the level of spending in an economy falls sharply for whatever reason, it is likely to remain low years later, even after the precipitating factor is removed. But why should economies behave this way?

I can think of a couple of reasons.

First, there’s the pure coordination story. Businesses pay wages to workers in order to carry out production. Production is carried out for sale. Sales are generated by spending. And spending depends on incomes, most of which are generated from production. This is the familiar reasoning of the multiplier, where it is used to show how an autonomous change in spending can lead to a larger (or smaller) change in output. The way the multiplier is taught, there is one unique level of output for each level of autonomous demand. But if we formalized the same intuition differently, we could imagine a system with multiple equilibria. Each would have a different level of income, expenditure and production, but in each one people would be making the “right” expenditure choices given their income. 

We can make this more concrete in two ways. First, balance sheets. One reason that there is a link from current income to current expenditure is that most economic units are financially constrained to some degree. Even if you knew your lifetime income with great precision, you wouldn’t be able to make your spending decisions on that basis because, in general, you can’t spend the money you will receive in the distant future today.

Now obviously there is some capacity to shift spending around in time, both through credit and through spending down liquid assets. The degree to which this is possible depends on the state of the balance sheet. To the extent a period of depressed demand leaves households and businesses with weaker balance sheets and tighter financial constraints, it will result in lower spending for an extended period. A version of this idea was put forward by Richard Koo as a “balance sheet recession,” in a rather boldly titled book. 

Finally there is expectations. There is not, after all, a true lifetime income out there for you to know. All you can do is extrapolate from the past, and from the experiences of other people like you. Businesses similarly must make decisions about how much investment to carry out based on extrapolation from the past – on what other basis could they do it?

A short period of unusually high or low demand may not move expectations much, but a sustained one almost certainly will. A business that has seen demand fall short of what they were counting on is going to make more conservative forecasts for the future. Again, how could they not? With the balance sheet channel, one could plausibly agree that demand shocks will be persistent but not permanent. But with expectations, once they have been adjusted, the resulting behavior will in general make them self-confirming, so there is no reason spending should ever return to its old path. 

This, to me, is the critical point. Mainstream economists and policy makers worry a great deal about inflation expectations, and whether they are becoming “unanchored.” But expectations of inflation are not the only ones that can slip their moorings. Households and businesses make decisions based on expectations of future income and sales, and if those expectations turn out to be wrong, they will be adjusted accordingly. And, as with inflation, the outcomes of which people form expectations themselves largely depend on expectations.

This was a point emphasized by Keynes: 

It is an essential characteristic of the boom that investments which will in fact yield, say, 2 per cent in conditions of full employment are made in the expectation of a yield of, say, 6 per cent, and are valued accordingly.

When the disillusion comes, this expectation is replaced by a contrary ‘error of pessimism’, with the result that the investments, which would in fact yield 2 per cent in conditions of full employment, are expected to yield less than nothing; and the resulting collapse of new investment then leads to a state of unemployment in which the investments, which would have yielded 2 per cent in conditions of full employment, in fact yield less than nothing. We reach a condition where there is a shortage of houses, but where nevertheless no one can afford to live in the houses that there are.

He continues the thought in terms that are very relevant today:

Thus the remedy for the boom is not a higher rate of interest but a lower rate of interest! For that may enable the so-called boom to last. The right remedy for the trade cycle is not to be found in abolishing booms and thus keeping us permanently in a semi-slump; but in abolishing slumps and thus keeping us permanently in a quasi-boom.

 

1. I use to think that the reason physical wells needed to be primed was to create a siphon effect. But according to David Macaulay, it’s because the leather valves used in old pumps don’t make a good seal unless they are wet.
2. Empirically disentangling the multiplier from persistent demand effects is a bit tricky. But we’ll skip over that for now.
3. Picking a different no-recession counterfactual growth rate would, obviously, change the picture, but not too radically unless it was a lot lower or higher.