An Inflation Expectations Experiment

Last semester, my senior thesis advisee Alex Rodrigue conducted a survey-based information experiment via Amazon Mechanical Turk. We have coauthored a working paper detailing the experiment and results titled "Household Informedness and Long-Run Inflation Expectations: Experimental Evidence." I presented our research at my department seminar yesterday with the twin babies in tow, and my tweet about the experience is by far my most popular to date:

I gave a 90 minute econ seminar today with the twin babies, both awake. Held one and audience passed the other around. A first in economics?

— Carola Conces Binder (@cconces) September 28, 2017

Consumers' inflation expectations are very disperse; on household surveys, many people report long-run inflation expectations that are far from the Fed's 2% target. Are these people unaware of the target, or do they know it but remain unconvinced of its credibility? In another paper in the Journal of Macroeconomics, I provide some non-experimental evidence that public knowledge of the Fed and its objectives is quite limited. In this paper, we directly treat respondents with information about the target and about past inflation, in randomized order, and see how they revise their reported long-run inflation expectations. We also collect some information about their prior knowledge of the Fed and the target, their self-reported understanding of inflation, and their numeracy and demographic characteristics. About a quarter of respondents knew the Fed's target and two-thirds could identify Yellen as Fed Chair from a list of three options.

As shown in the figure above, before receiving the treatments, very few respondents forecast 2% inflation over the long-run and only about a third even forecast in the 1-3% range. Over half report a multiple-of-5% forecast, which, as I argue in a recent paper in the Journal of Monetary Economics, is a likely sign of high uncertainty. When presented with a graph of the past 15 years of inflation, or with the FOMC statement announcing the 2% target, the average respondent revises their forecast around 2 percentage points closer to the target. Uncertainty also declines.

The results are consistent with imperfect information models because the information treatments are publicly available, yet respondents still revise their expectations after the treatments. Low informedness is part of the reason why expectations are far from the target. The results are also consistent with Bayesian updating, in the sense that high prior uncertainty is associated with larger revisions. But equally noteworthy is the fact that even after receiving both treatments, expectations are still quite heterogeneous and many still substantially depart from the target. So people seem to interpret the information in different ways and view it as imperfectly credible.

We look at how treatment effects vary by respondent characteristic. One interesting result is that, after receiving both treatments, the discrepancy between mean male and female inflation expectations (which has been noted in many studies) nearly disappears (see figure below).

There is more in the paper about how treatment effects vary with other characteristics, including respondents' opinion of government policy and their prior knowledge. We also look at whether expectations can be "un-anchored from below" with the graph treatment.

Consumer Forecast Revisions: Is Information Really so Sticky?

My paper "Consumer Forecast Revisions: Is Information Really so Sticky?" was just accepted for publication in Economics Letters. This is a short paper that I believe makes an important point. 

Sticky information models are one way of modeling imperfect information. In these models, only a fraction (λ) of agents update their information sets each period. If λ is low, information is quite sticky, and that can have important implications for macroeconomic dynamics. There have been several empirical approaches to estimating λ. With micro-level survey data, a non-parametric and time-varying estimate of λ can be obtained by calculating the fraction of respondents who revise their forecasts (say, for inflation) at each survey date. Estimates from the Michigan Survey of Consumers (MSC) imply that consumers update their information about inflation approximately once every 8 months.

Here are two issues that I point out with these estimates:

I show that several issues with estimates of information stickiness based on consumer survey microdata lead to substantial underestimation of the frequency with which consumers update their expectations. The first issue stems from data frequency. The rotating panel of Michigan Survey of Consumer (MSC) respondents take the survey twice with a six-month gap. A consumer may have the same forecast at months t and t+ 6 but different forecasts in between. The second issue is that responses are reported to the nearest integer. A consumer may update her information, but if the update results in a sufficiently small revisions, it will appear that she has not updated her information. 

To quantify how these issues matter, I use data from the New York Fed Survey of Consumer Expectations, which is available monthly and not rounded to the nearest integer. I compute updating frequency with this data. It is very high-- at least 5 revisions in 8 months, as opposed to the 1 revision per 8 months found in previous literature.

Then I transform the data so that it is like the MSC data. First I round the responses to the nearest integer. This makes the updating frequency estimates decrease a little. Then I look at it at the six-month frequency instead of monthly. This makes the updating frequency estimates decrease a lot, and I find similar estimates to the previous literature-- updates about every 8 months.

So low-frequency data, and, to a lesser extent, rounded responses, result in large underestimates of revision frequency (or equivalently, overestimates of information stickiness). And if information is not really so sticky, then sticky information models may not be as good at explaining aggregate dynamics. Other classes of imperfect information models, or sticky information models combined with other classes of models, might be better.

Read the ungated version here. I will post a link to the official version when it is published.