Why do yield curve inversions predict recessions (and why this time it might not)?

 

The brief inversion of the 2s-10s US yield curve led to a rather typical explosion of commentary discussing whether we are looking at a recession in the US in the next 6-18 months. This commentary is mostly based on the empirical evidence that yield curve inversions are the most reliable predictor or US recessions.[1] This raises the question ‘Why do yield curve inversions predict recessions?’.

Start with establishing what do yield curve inversions predict almost for sure: yield curve inversions are market-based predictions of future decrease in short term interest rates. Since yields of different maturity are equal to expected value of short term interest rates over the length of given maturity, then decreasing yield curve has to imply that market expects short term interest rates to be lower at some point in time in the future.[2]

With this established, the question then becomes why would an expected future decrease in interest rates be a predictor of a future recession. At a first glance, it seems obvious: recessions are associated with a decrease in interest rates. In modern economies this is because central banks decrease nominal interest rates to fight recessions; but even outside of modern economies, this should be the case, as recessions decrease demand for investment/consumption, and hence decrease the natural rate of interest.

However, this answer provides only a partial justification for why do yield curve inversions predict recessions. While expected future recessions should imply downward sloping yield curves, this does not mean that downward sloping yield curves necessarily imply recessions: necessary condition is not the same as a sufficient condition.[3] In other words, a future decline in interest rates does not necessarily mean a full recession, defined as outright decline in economic output. This is especially if one considers nominal interest rates in current environment: for example central banks in CEE have increased policy rates to extremely high levels to fight a bout of inflation, causing inversions in yield curves. However, local market yield curves are just predicting that interest rates decrease towards their long-run equilibrium values, and not below those values, consistent with central bank’s expectations that policy rates will be gradually normalized, reaching the long-run equilibrium levels from above. Therefore, an inversion is not signaling a recession.

The example of CEE countries holds a more general lesson: expected decrease in short term interest rates does not equal a recession, as short term interest rates can decrease outside of recessions. For example, central banks might orchestrate a yield curve inversions if they intentionally temporarily increase short term interest rates above their long-term equilibrium values. Or more generally, natural interest rates might decrease in absence of outright recession, in case of what is often called a growth recession: a temporary decline in the growth rate of the economy without an outright decline in output. A prime example was the 2019 inversion in yield curve, which was prompted by the economic slowdown caused by Trump’s trade war. While we will never know for sure, it is quite likely that if it was not for the pandemic, the US economy would experience just period of slower growth without an outright decline in output.

Turning to today, while there are good reasons to worry about recession in US, there is a good chance why yield curve inversion might not a be signal of one: the Fed is explicit that it wants to increase short term interest rates above the long-term equilibrium values. This alone should imply a yield curve inversion, but one intentionally engineered by the central bank. Of course, this policy might actually cause recession, which will then just add to the mythology surrounding yield curve inversions.

 

---

[1] That said, theorical considerations also do play a role: one can construct models in which yield curve inversions cause, rather than just predict, recessions. I will ignore this aspect here.

[2] Strictly speaking, one could construct decreasing yield curve also by postulating term premium that decreases with maturity. While not impossible, especially in current environment with term premium compressed by regulatory requirements and large balance sheets of central banks, this seems to be rather an edge case. Moreover, argument carries less weight when one focuses on changes in yield curve slope, rather than just the arbitrary metric of negative-sloping yield curve: even if long-term yields are compressed due to compressed or even negative term premium, it should still hold that decrease in the slope of yield curve implies that short term rates are either expected to increase less, or to outright decrease. Similarly, the argument does not apply to the near-term yield curve spread, since there term premium plays limited role.

[3] Moreover, downward sloping yield curve is clearly not even a necessary condition: while one would generally expect downward sloping yield curve ahead of predictable recession, this of course does not apply to unpredictable recessions, and most recessions are thought to be unpredictable. Case in point was the pandemic recession, which was clearly unpredictable 6-12 months ahead of time. From this perspective, it actually seems odd that yield curve inversions are such a reliable predictor of recessions: it is as if something that is thought to be mostly unpredictable is being reliably predicted.

Does inflation cause decrease in aggregate demand?

With inflation occupying the headlines, one can often read arguments that high inflation will cause a decrease in the aggregate demand and potentially a recession. What does basic macroeconomics have to say about this argument? The simple answer is that such reasoning is faulty, but the more complicated answer is that in current situation it might hold a grain of truth.

Start with the basic AS-AD model, which has the total output on the horizontal axis and the price level on the vertical one.[1] In this space, the aggregate demand is a downward sloping curve, with higher price level resulting in lower total output demanded. Meanwhile, the aggregate supply is upward sloping curve, capturing that higher price level means higher total output supplied.

On the first (very brief) look it might seem that inflation can indeed cause decrease in aggregate demand, but one should quickly realize that this reasoning is faulty: while the total output demanded will decrease when price level increases, the curve capturing the relationship between output demanded and price level is the same. In other words, this is a macroeconomic analogy of the mistake made by first year students of microeconomics, who struggle to distinguish between decrease in quantity demanded caused by higher price, and decrease in demand caused by some other factor. Since only the change in (aggregate) demand or supply can change the equilibrium price (level), then arguing that higher price (level) leads to lower (aggregate) demand is confusing causes with consequences: the price (level) is an equilibrium outcome of the model, and not a driving force of changes in the equilibrium. As one of my favorite bloggers says, an economist should never argue from a price change!

At this point one could feel that this is just semantics – economists often say that higher price means lower demand, knowing full well that it is really quantity demanded what they have in mind. They do this because they speak to general population that did not go through basis economic course and hence would not understand the terminology of “change in quantity demanded” vs. “change in demand”.

However, the problem is that in many situations the writers are not using this as a shorthand terminology, which then leads to problematic lines of reasoning. In case of inflation and aggregate demand, the problematic reasoning goes like this: we are currently seeing high inflation, which will cause a decrease in aggregate demand, and hence we are in danger of economy going into recession.

In this reasoning, the high inflation is what is driving the economy into recession. This reasoning is problematic because high inflation is a response of the model to some shock, not source of shock. To see this consider a scenario in which current inflation is caused by large increase in aggregate demand. In such situation, the AS-AD model tells us that output should increase and so should price level, both of which we observed. Is there a reason to worry about decreasing aggregate demand and recession? No, because we started with increase in aggregate demand!

What about the situation when the higher prices are cause by decrease in aggregate supply, say because of an supply bottlenecks? Here, we will indeed observe decrease in output – a recession – and inflation. However, it is not inflation causing the decline in output, it is the supply shock and corresponding decrease in supply. Worrying about inflation is double-counting: there is no further decrease in output because of the high inflation, all of the decrease is due to the supply shock.

The faulty reasoning can go even one step further: if inflation will cause decrease in aggregate demand, then, some argue, this is all good, because lower aggregate demand will lead to lower prices/inflation, which is what we need. Of course, one then can wonder whether lower prices will not cause higher aggregate demand, which will cause higher inflation… and so on and so on.[2]^,[3]

Overall, we are left with the conclusion that in basic macro saying “inflation will cause decrease in aggregate demand” is a nonsense. Does it mean the argument belongs to garbage bin? Or could we find support for it if we leave the world of basic macro? Every now and then I catch myself being annoyed with some argument which is a bad economics, but that captures some more complicated story that is valid, and I think this is the case here.

To see this, consider abandoning the notion of homogenous output, and considering heterogenous outputs. Large part of the European inflation right now is driven by jump in energy costs, reflecting the combination of diminished supply and revived (global) demand. While the diminished supply in AS-AD world would amount to concurrent decrease in aggregate supply and increase in aggregate demand, resulting in inflation, the AS-AD world is maybe not very suitable for the present situation. Instead, we could think about things in terms economy with two sectors, energy sector and all the rest, and assume that energy is supplied by foreign country. In such situation, increase in price of energy will lead to a decrease in aggregate demand in the rest of the economy, and hence inflation will indeed cause something which is akin to decrease in aggregate demand.

What about savings, not present in basic static macroeconomics, could those create link between inflation and aggregate demand? Inflation clearly leads to decrease in value of savings, so it might seem that inflation will cause decrease in savings, which will cause decrease in aggregate demand. However, this story is not fully satisfying from macroeconomic perspective: in general, higher prices also mean higher profits and higher wages, which should cancel the effect on aggregate demand. But it might be the case that the higher prices are indeed driving decrease in aggregate demand, if we either consider heterogeneity of households or some form of myopic behavior. In case of heterogenous households, it might be that the distribution of costs of higher prices and benefits of higher prices is such that it shifts money towards households who will spend less. In case of myopic households, it might be that higher households are spending for as long as they have money in their accounts, but once they draw down these balances, they stop spending.

So, there, we have it. In general, the argument “inflation will cause decrease in aggregate demand” rings hollow, but it might not be always the case if the inflation is either heterogenous in terms of its source, coming from decrease in supply in particular sector of the economy, or heterogenous in terms of distribution of impacts on households. I don’t think that’s what the pundits have in mind, though.

P.S.: Note the microeconomics equivalent of this blogpost is a blogpost about “Do higher car prices cause lower demand for cars?”. The answer is yes, if one means decrease in quantity demanded, but no, if one means that people will not want to buy cars anymore and that car makers should be worried: as soon as price of cars will go back down, quantity demanded will again increase.

 

---

[1] One could re-formulate the model so that we have (surprise) inflation on the horizontal axis instead of price level.

[2] The is again analogical to microeconomics, where you will see students in an endless loop of reasoning: higher price means lower demand, which means lower price, which means higher demand, which mean higher price, which means lower demand, …

[3] This last point points towards a way one could salvage the argument that inflation will lead to decrease in aggregate demand. Maybe, it could be an argument about disequilibrium and dynamics. Maybe, the decrease in aggregate supply was not matched by decrease in aggregate output demanded yet, and hence decrease in output is to be expected in due course. It does not take the form of shift in the curve, but rather movement from disequilibrium position to position back on the curve.

However, this does not solve the puzzle fully: the already observed higher prices mean that we have moved, so current inflation should not cause further decline in output demanded in future. Moreover, in economics it is typically the prices that adjust faster than quantities. After all, the decrease in supply means that we are not able to produce the same quantities as before, so it is hard to see how it is the prices that adjusted first and quantities will adjust later.

Reminder: Price setting is a coordination game

 Inflation is the main topic of macro discussion right now, with price increases exceeding expectations, even after accounting for the additional shocks. As usual, the media focus is on the situation in the U.S., but the phenomenon is pretty much global This means that other countries can offer us some clues on the nature of the phenomenon.

One such country is Czechia, which combines features that make it an interesting learning case. On one hand it is a country with firmly established inflation targeting, so that it is not plagued by the regular instability of inflation expectations that haunt many emerging markets. This means that its experience is relevant for central banks in developed economies, unlike experience in, say, Brazil or Turkey. On the other hand, it is a small open economy with significant focus on industrial production. This means that its experience is not only sufficiently different from US or euro zone, but that it is more exposed to current sources of inflationary pressures than these countries.

So what is the recent inflationary development in Czechia? Simply, inflation is running amok: prices have increased by around 1% in each of the last 4 months. That is,  increase of 4% in span 4 month. Or to put in yet another way, it is the increase central bank desires in a two-year period occurring in four months. While the sources are to some degree global, such large increase cannot be explained by these factors alone. And before one jumps into conclusion that this is result of domestic policy mismanagement, it is important to note that the government support for the economy was not abnormal – apart from decrease in income tax benefiting mostly the rich – and that the economy is actually doing quite badly relative to its peers.

So what is the lesson here? This is speculative, but one way to think about the situation is in terms of what we know about the incentives for price setting. Increasing prices is in its nature a coordination game: each firm does not want to raise its prices in order not to lose customers to its competitors. This is what lies behind sluggishness of price adjustment in economic models.

But what if everybody knows that everybody will be increasing prices because of spike in input prices such as commodity prices, energy prices, component prices etc.? Well, if price setters know that their competitors will be increasing prices now, and that they will need to increases prices relatively soon anyways, then they might as well decide to increase prices right away. After all, higher prices are less likely to be noticed in environment when everybody is increasing prices.

This means that the rapid price growth might be result of coordinated equilibrium: everybody increasing prices because they observed public signal that everybody will be increasing prices. And as such, some of the price increases we are seeing might be prices increases brought forward, causing inflation to spike more than one would expect. In other words, we might be seeing a year worth of price increases crammed into several month in which everybody expected everybody to hike prices.

This means two things. First, in the short term inflation is likely to surprise on the upside even more, with other price setters jumping on the inflationary bandwagon. Second, the inflation might drop relatively quickly: once the period of rapid price increases is over, further price increases might not be forthcoming, since they have been brought forward. In some sense, we will have reached the natural level of prices, just in different path, with rapid inflation followed by quick slowdown in inflation, rather than gradual rise we are used to. Of course, the second observation applies only if the inflation does not cause change in inflation expectations, in which case rapid inflation could start to feed on itself.

 

This pattern might even fit the historical experience: it seems that commodity-driven inflationary periods which did not result in shift in inflation expectations are often followed by long periods of subdued inflation. Most recent example is the example of inflationary spike in 2011, and long years of low inflation afterwards.

 

P.S.: Note that this is one example of why the Calvo pricing is so not up to the task: it does not allow firms to choose when to change their prices, only by how much, and hence does not allow for coordinate price increases.

P.P.S.: In terms of time series modelling, this would suggest that the price level, not just its part rate of change, is important. Or alternatively, that there are more regimes in inflation.

P.P.P.S.: There is a notion of attention-driven inflationary behavior in here: maybe individual price rises are less likely to be noticed when they are done in environment of many prices rising.

 

 

 

ARMA models, multiple components and recoveries of consumption series from COVID

Introduction – ARMA models and behavior of multiple components

Time series models of the typical ARMA family should be viewed as a tool to tell ourselves stories about how time series behave. They focus us on the question of primary importance - how will the shocks be propagated into future and hence influence future values – and allow us to answer it in simple terms. For example autoregressive model with high coefficient tells us that effect of shocks will be eliminated only gradually and hence propagated over long period of time. On the other hand moving average models of low order tells us that the effect of shocks will disappear quickly. Or finally, moving average model with negative coefficient tells us that the recent shocks will have opposite effects in near future.

This simple pallet of models is very flexible and allows us to capture most of the behavior of real world series. One particular place where it can be applied is in thinking about detailed consumption and production series influenced by the pandemic (think purchases of cars or box office sales). In principle, these series can be thought to be composed of at least 4 components, 3 relating to lockdowns (or more broadly social distancing) and one to the overall economic shock.

The three components related to lockdowns effect capturing the effect of lockdowns which prevent consumption/production; effect of pent-up demand after lockdowns are lifted (i.e. most people who very planning to purchase car will still do so sometime after lockdowns); and effect of substitution, as money saved on some regular expenditures due to lockdowns are spend on some other items due to impossibility of time-substitution (e.g. people will not catch up with all their regular cinema and restaurant visits they have missed). The other component is component capturing effects on confidence and/or expectations about current/future incomes, i.e. the usual component we see in recessions, albeit this time the bad news arrived in more time-compressed manner.

These components behave very differently from each other. Moreover, not all series contain all components, and even if they do, these components might behave differently for different series (think of more persistent lockdowns/social distancing effects for most sensitive industries). This means that in contrast to normal recessions there will be much bigger heterogeneity across individual consumption series. Moreover, the first three are very unique to our current situation and their novelty together with their complexity means that we can easily get confused by temporary movements. It should be therefore worthwhile to consider the different options we might encounter over the next few months and what it means for behavior for given series. Below are few examples.

Example 1 – Spending on restaurants: Pure lockdown shock with no pent-up demand and substitution

The first example applies to series which is affected by the lockdowns, but does not benefit from the pent-up demand and substation effects.  An example could be spending in restaurants – restaurants in many countries were closed for most of April and to a lesser degree May and June, so that the spending in restaurants is substantially lower in these months. At the same time, people will return to restaurants as soon as their re-open, so that July should be back to normal level of spending (or close to that). The graph below translates this discussion into deviations from normal level of spending (left chart) and into growth rate of the series (right chart). The magnitudes are illustrative and not meant to be representative.

Turning to the way this is captured in the standard time series ARMA models, the profile in the chart is achieved by using MA(2) process with single (negative) shock in April. The moving average model extends the effect of shocks for the number of periods corresponding to the order of the process, in our case two extra periods.[1] Intuitively (but not necessarily), the coefficients on lags of shocks are lower than 1 so that the effect of the shock decreases over time. Of course, the length of the shock effect as well as its strength in each period can be easily changed with the order of the model or the size of the coefficients: moving average models give us complete flexibility over this, albeit at a cost of possible overparametrization.

An alternative to lengthening the effect of shock through increasing the order of moving average model is to use autoregressive model. AR models eliminate effects of shocks gradually, multiplicatively.[2] Therefore, in contrast to MA models, the effects of shocks persist for more periods than is the order model. In case of our lockdowns we can think of lasting lockdowns for most dangerous industries, say restaurants focused on international tourists (or similarly international air travel). Picture below captures such situation:

Of course we could combine moving average and autoregressive models to get ARMA model, which combines the profile of responses in first and second pictures:

Note that the different three options of lockdown effect profiles translate into slight, but important variation in growth rates we should observe. In first case large drop is followed by immediate and rapid growth for 3 months. In second case the growth is still immediate, but it is not rapid; instead the recovery is spread over many months. Finally, in the last case the growth is not even immediate but actually comes with some delay.

The previous pictures all assumed that there is no confidence shock associated with the pandemic. This was our collective assumption in January and February, but in March it became clear that we will not return to normal as soon as the lockdowns are over. Simply, the pandemic also caused global recession, which influences are current behavior through expectations about future economic situation, by for example lowering our expectations of future incomes (or even current incomes for the unlucky who lost their job already). This of course causes people to cut back on expenditures already now, and hence spending on restaurants would now be lower even if we would eliminate Covid-19 with a magic wand. To capture this, the next picture adds the confidence channel to the previous picture.[3] The key difference is that after lockdowns are over we do not return to the normal level of spending, but rather remain slightly below for the rest of the year. Correspondingly, the growth rates in May-July period are lower, but mild growth continues for the rest of the year.

Turning to econometric calibration, what we have added is AR(2) process for confidence. As before, we can see that the autoregressive model makes the effect last very long: the confidence component is negative throughout the whole year. Moreover, AR models of order higher than 1 with first coefficient bigger 1 also feature amplification: the peak effect occurs with delay. Therefore, the confidence component is largest in absolute value in June, two months after the initial shock. Finally, the sum of the coefficients is very close to 1, so that the effect of the shock is eliminated very slowly and even at the end of the year it is close to its maximum size.

Of course, things could be easily re-parametrized. We could ensure that the peak effect is either sooner or later, and we could ensure that the elimination is faster or slower. This could be done either within the context of AR(2) model, so would not require additional parameters, or by changing the order of the model. Examples of two re-parametrizations are below:

Therefore, the difference between autoregressive and moving average processes is in terms of trade-off between flexibility and over-parametrization: autoregressive process can give us some flexibility for given number of parameters, while moving average processes can give us infinite amount of flexibility at cost of many parameters.

Finally, all the charts above assumed that ¾ of the impact we observe in May is due to lockdowns and rest due to demand effects. Of course, for different consumption series different distribution will make sense: for other industries it will be mostly demand shock and bit of lockdown shock (spending on newspapers); for some it will be more equal; and for some there might be even a negative lockdown shock (think purchases of food in grocery stores).

Example 2 – Spending on cars: Lockdown shock with pent-up demand but no substitution

The second example applies to series which is affected by the lockdowns, but after lockdowns are lifted it benefits from the pent-up demand (but not from substitution effects).  An example could be spending in cars – car dealerships in many countries were closed for most of April, so that the spending on cars is substantially lower in that. At the same time, most people who were planning to buy car will do so soon after lockdowns anyway irrespective of the pandemic. Such situation is captured in next chart:

Here the crucial thing is to realize, that the pent-up demand gives us temporary small boost in the May (and even smaller in June), so that the rebound is faster than before. After June the profile is again determined solely by the confidence component and hence is identical as before. There is one even more important realization: since the pent-up demand gives us temporary boost in terms of level, there is actually decline in the spending in June, even though the June deviation from normal is smaller than before. This highlights the problematic nature of growth rates in coming months.

In terms of econometric calibration not much has changed: lockdown and confidence components are as in original case, while pent-up demand component is moving average model as the lockdown. The way things have been specified is using the same error as in lockdown equation and set current-period coefficient to zero, which is rather unusual. Nevertheless, this is for simplicity: one could write this down more elegantly, but it would blunt the message the pen-up demand is just partial reversal of lockdown shocks.

Example 3 – Spending on home-improvement: Lockdown shock with pent-up demand and substitution

The last example applies to series which is affected by the lockdowns, but after lockdowns are lifted it benefits from the pent-up demand as well as substitution effects.  An example could be spending on home improvement – this was impossible during the lockdown since hobby markets were closed for most of April, so that the spending on home improvement is substantially lower in that month. At the same time, most people who were planning to do such home improvement will do so soon after lockdowns anyway irrespective of the pandemic. And on top of that, since people saved money on some regular expenditures with which they do not plan to catch up – think no tourism, restaurant visits or cinema outings – they decide to use left-over money for home improvements. Such situation is captured in next chart:

 The picture resembles the previous case but the rebound is even stronger in May, so that the deviation from normal actually turns positive in this month. However, once the temporary positive effects dissipate the confidence effect takes over and the series goes back below normal. The growth rate in this situation is even more strangely looking than in previous case – after drop there is even larger jump, followed by large second drop before stabilization. Hence, growth rates do not tell useful story almost at all.

Econometrically, we have added another moving average process very similar to pent-up demand, so there is nothing new in that sense.

Conclusion – More caution than usual needed

Given the nature of the shock the behavior of individual consumption series will vary greatly. This divergence reflects the fact that in current situation – in contrast to normal recession – series can be thought of as composed of several differently-behaving components. This then can play havoc in how individual series will behave: not only will there be large swings in growth rates, there might be also multiple switches from positive to negative growth rate. This highlights how growth rates can be very misleading in coming months. While level will certainly be more informative, one can imagine reversals even here. So the overall message is that one has to be cautious when drawing conclusions from  individual consumption series in coming months.

---

[1] Of course, one could argue that what is really happening is three successive shocks, but this distinction is relatively slight in current situation. The notion here is that the effects of lockdowns always linger for some time, so that they can be described by some firm process.

[2] In the simplest case of AR(1) model, say with coefficient 0.9, 10% of the previous period effect is eliminated in each period.

[3] Note that the April impact is the same as before, just that part of it is demand shock. This is true throughout this document.