Recent Posts

Saturday, July 2, 2016

John Taylor Highlights The Uselessness Of Taylor Rules

The working paper "Finding the Equilibrium Real Interest Rate in a Fog of Policy Deviations" by John B. Taylor and Volker Wieland (April 2016) tells us a lot about Taylor Rules (link to primer). Although presumably not his intention, it tells us that the concept is fatally flawed. Furthermore, since a Taylor Rule (or another reaction function for the central bank) is at the core of Dynamic Stochastic General Equilibrium (DSGE) model, it provides another reason to believe that this class of models is inherently useless.

For central bank watchers, it underlines the reality that despite the pretensions of being a highly mathematical science, central banks are making up their interest rate policy as they go along. This is probably not a surprise to experienced market participants, but it could be a shock to fresh graduates who believe the propaganda pumped out by mainstream economics faculties.

Relation To The Analysis In My Book

Chapter 2 of Interest Rate Cycles is titled "Central Bank View of Policy Rates," and leads up to the Taylor Rule, which ties together most of the theoretical concepts I discuss (inflation targeting, inflation expectations, the output gap, and the natural rate of interest). I also discuss why Taylor Rules are somewhat unworkable in practice -- a determined economist can always find a rule to match their predetermined view on interest rates. That is, a hawk will always be able to find a rule that suggests the policy rate should be higher than it is now (John Taylor being a prominent example).

I have heard the argument in a different context that central banking has advanced way beyond this, and thinking has been revolutionised. Since most of my writing aims to be at an "introductory" level (although aimed at people with some knowledge of business cycle economics), I would argue that we need to start with the basics. Furthermore, the more "advanced" ideas just involve adding epicycles to existing DSGE models, and the core criticisms raised about Taylor Rules are just as applicable to things like "optimal control rules" (link to article explaining why I chortle when I write about them).

Any reaction function is going to be hostage to model parameters, particularly "sophisticated" ones like optimal control rules. (Their instability is why engineers abandoned optimal control rules. Simpler control laws -- which resemble the Taylor rule -- are more robust to model specification error.)

If The Natural Rate Moves Around At Random, Policy Is Discretionary

Taylor and Wieland (page 11):
In any case this is a controversial and debatable issue, deserving a lot of research. If one can adjust the intercept term (that is, RR* [the equilibrium real natural rate - BR]) in a policy rule in a purely discretionary way, then it is not a rule at all any more. It’s purely discretion. Sharp changes in the equilibrium interest rate need to be treated very carefully.
Why are estimation methods for the natural rate going to miss? The means of estimating them omit many factors (as highlighted by Taylor and Wieland), but particularly the operation of the automatic stabilisers of fiscal policy (which they ignore). I discuss this theoretical point in an earlier article.  To summarise, since the automatic stabilisers and monetary policy are correlated, while fiscal policy is banished from DSGE models, estimation techniques that  assume the DSGE model dynamics are correct will attribute too much importance to interest rate policy. This means that the estimated natural rate always ends up being close to the actual policy rate -- no matter what the policy rate is.

From John Taylor's perspective, discretionary policy is a deal-breaker. Like many advocates of free markets, his research programme is built around the belief that government policy needs to be hemmed in by policy rules. Although some free market advocates want to revert back to the mythical rules-based Eden of a Gold Standard, Taylor's preference is for rule-based interest rate policy. (Milton Friedman's belief in another rule-based policy -- targeting money supply growth -- led to Monetarism jumping the shark in the early 1980s.)

Even if one thinks that rules-based policy is a disastrous dead end (I do), the problem for DSGE macro is that it is built entirely around the concept. The solution to "inter-temporal optimisations" requires that the path of the policy rate be set by a rule of some sort, otherwise there is no way of finding a solution. One of the selling points of DSGE macro was that it provided such rules, even though it provides almost no useful information for how to set the policy rate at an individual meeting. However, if policy is entirely discretionary, why exactly are we forced to do all this pseudo-mathematics?

Yelping about the zero lower bound, or optimal control rules, does nothing about this critique. Since we have no idea what the "true" natural rate is, we have no idea whether any of that theory is applicable.

Article Misses The Point Of Monetary Policy

The article argues that the natural rate of interest has not really moved, rather that estimation procedures are missing two variables. (Neither of them are the correct variable, which is fiscal policy.)
  1. The variable "x* could represent a variety of influences on real GDP from regulations that negatively affect investment to tax policy that negatively affects consumption." (Incidentally, John Taylor has a new book out explaining how regulatory policy is strangling economic growth.)
  2. "d* is a possible deviation from the policy implied by the [Taylor] rule".
The effect of the first variable x* can be easily dealt with. The article states "With equation (1’), if one finds that y-y* is lower than the prediction P(y-y*), then the implication is not necessarily that the estimate of r* is too high and must be lowered. Now there is the possibility that x* is too low and must be raised. [emphasis mine]"

This is not how monetary policy works. Within the reaction function for the policy rate, there is no output corresponding to "regulatory policy." In fact, such changes are outside the mandate of the Federal Reserve (unless they believe that zealous enforcement of banking regulations by the Fed is what has caused disappointing growth, which is utterly implausible).

The Taylor rule just specifies the policy rate, and the natural rate of interest is a model-dependent variable which summarises the effect of interest rates on the economy based on the estimated economic structure. Since the Fed has no control over  x*, it has no choice but to set policy to counteract it. It cannot set rates based on the United States economy being at some utopian equilibrium, it has to take into account how the economy is actually operating,

The d* variable is an extremely puzzling addition to the Taylor Rule. It essentially says that if the policy rate deviates from the predicted rate, then (r* + d*) will move towards the (real) policy rate. Since these two variables are essentially inseparable, we end up with a term (r* + d*) that always drifts towards the real policy rate. It is equivalent to saying that the natural (real) rate of interest is the moving average of the (real) policy rate. This is a specification that has the implication that the natural rate of interest has essentially no predictive power. (No matter where the actual policy rate is set, the term  (r* + d*) will end up where it is, which means the policy rate setting has no long-term effect on the economy.) In which case, why do we even care where the central bank sets the interest rate?

Concluding Remarks

Mainstream economics is built upon unmeasured variables, and has drifted towards being entirely unfalsifiable. Until this reality is taken seriously, it will consist of pointless debates such as we see in this Taylor and Wieland paper.

(c) Brian Romanchuk 2016


  1. Very interesting analysis, Brian. Thanks.

    Yes, these 'free variables' are complete rubbish. They remind me of some of Keynes' quips in 'Professor Tinbergen's Method' about various of Tinbergen's chosen variables.

    Are you also familiar with the natural rate estimation procedure? It is a very strange process that builds on a number of other unobservables. It is based on the idea that there is a relationship between interest rates and an investment-savings (IS) curve. The whole estimation process is built up from there. But of course the IS curve is not empirically observable. So they seem to be building castles in the sky from what I can tell.

    1. I wanted to keep my book at an introductory level, so I did not go into the estimation procedures. There's a lot of them, and if I point out the problems with one, people will just complain that there is a "better" one. As a result, I did not write out a fuller critique, as it would require examining them one at a time; and frankly, they are spitting 'em out faster than I can read them.

      A fairly common technique is the use of a Kalman Filter. In practically every case I ran into, I had a hard time replicating the work, since they waved their hands around the exact procedure they used. Since I will likely end up using slightly different inputs, I have no idea whether I am replicating the work properly. As a result, I gave up and just took an example where the author helpfully provided the final output.

      Returning to the Kalman Filter, it comes out of 1960s Control Systems engineering (my Ph.D. area). I did not work much with them, but the core assumption is that you have a correct model of the system, and then you use directly measured variables to create best estimates of state variables. (Note that I say "measured" and not "observed," as "observed" has a different technical meaning than what common sense would suggest. If we can estimate a non-measured state variable -- like the natural rate of interest -- it is "observable." It is "unobservable" if there is no way of estimating it from measured outputs. Although that sounds stupid, it matters a lot for control systems engineering design. You have to ensure that you have sensors that can capture all the state variables that matter. For example, having a plane without an altimeter is a pretty bad idea...)

      The key problem is that *you assume that the model dynamics are correct*. Let that sink in, and then ask yourself how useful mainstream modelling is if you start to question the *assumption* that the real rate of interest drives the evolution of the economy. There is a reason that I have largely given up on mainstream macro, other than to call it non-falsifiable.

      As an aside, do you have access to the Blue Chip Consensus data? If you do, send me a message (via the blog form), and I would suggest an interesting experiment to look at (with regards to another area of interest rate macro).

    2. No access to Blue Chip Consensus, no. Why don't you do a post.

      On the Kalman Filter process, I'm not even sure if they were using measured variables. For example, they plug in Phillip's Curve estimations.

      a) These are estimations, not measures in a true sense.
      b) The Phillip's curve probably does not really exist at all, so the causal force underneath their estimation is questionable at best.

      I think in the future this stuff will really be looked back on with bemusement.

    3. The Kalman Filter needs to work on some observed data, but all the dynamics are baked into an estimation model. That is, if we assume that these are the dynamics, where is the model state most likely to be based on the observed data?

      Of course, they may have used a model to generate the input to the filter; it depends upon the researcher.


Note: Posts are manually moderated, with a varying delay. Some disappear.

The comment section here is largely dead. My Substack or Twitter are better places to have a conversation.

Given that this is largely a backup way to reach me, I am going to reject posts that annoy me. Please post lengthy essays elsewhere.