fession at large has not adopted any such methods on a substantial scale

for applied work.

As far as I recall, no one had ever mentioned Bayesian methods at

MIT, though I found Ed Leamer, an assistant professor at Harvard, who

was deeply involved in using Bayesian foundations to adapt the scienti¬c

method to economics [Leamer (1978)].

Campbell: Who taught econometrics?

Shiller: Franklin Fisher had written a book on the identi¬cation prob-

lem in econometrics [Fisher (1966)]. We went through that whole book,

an elegant treatise, but perhaps too much on that topic. The econometrics

course I had with Edwin Kuh doesn™t stand out in my memory, but I

can say that he impressed me about the importance of regression dia-

gnostics and of isolating in¬‚uential observations, practices that not enough

people implement even today [Belsley, Kuh, and Welsch (1980)]. I learned

the essential lesson to be skeptical of econometric results. I remember

when Leonall Anderson and Jerry Jordan came to MIT in 1968 to pre-

sent the “St. Louis Model” of the U.S. economy [Anderson and Jordan

(1968)]. The results were impressive, but not really received well at MIT.

Later, our skepticism was borne out. Ben Friedman reestimated the same

model in 1985 and found that the new data provided by the mere passage

of time had destroyed their results [Friedman (1977)]. There are lots

of ways econometric analyses can go wrong.

I then started reading time-series analysis on my own. I never took a

course in that.

Campbell: Did you read Box and Jenkins?

Shiller: I certainly did, but I wanted to combine it with Bayesian

methods. Arnold Zellner at the University of Chicago somehow dis-

covered me; he invited me starting as a graduate student to a series of

Bayesian econometrics conferences. It was at one of these conferences in

1972 that I ¬rst met Sandy Grossman, then only 19 years old, and

236 John Y. Campbell

already a dazzling intellect. I was fortunate to have the opportunity to

work with him later on several papers.

I tend to attribute my interest in Bayesian statistics and time-series

analysis to my physical-science orientation, which had been with me since

childhood. I have long admired scientists. I thought that the Bayesian

methods would help adapt the scienti¬c method to economics, help us to

base our analysis on what we do know, and let the data speak for what

we do not know. The kind of science that appealed to me was the kind

that was based on careful observation followed by induction that allowed

you to discover a general principle. It was that discovery process that

excited me. Bayesian econometrics appealed to me then as a good

approach since it didn™t impose some arbitrary model. In fact, the prior

was supposed to come from some previous analysis; your prior was your

earlier posterior.

I also thought that science is, at its core, really intuitive. Charles

Darwin didn™t follow a research program that was outlined for him. He

was trying to think how this whole thing works and observed everything

he could. Leamer referred to “Sherlock Holmes inference,” in response

to that ¬ctional detective™s attention to all the details, but I would prefer

to call the ideal “Charles Darwin inference.”

Campbell: You started to mention the Lucas Critique.

Shiller: When I ¬rst read Lucas™s paper in 1975, I thought that there

was nothing new in it. The idea of rational expectations was already

prominent at MIT, through Modigliani and Sutch.

Campbell: But they didn™t actually cause you to change your mind

about econometric modeling. Their papers assume a ¬xed structure.

Shiller: If you were to take Franco aside then, and ask him, “isn™t

there a risk that if policy changes, the expectations structure might change,”

he would say, “obviously.” But, Lucas presented this in a very forceful

way. . . . Lucas is a great writer.

Campbell: Another idea that was ¬‚oating around at the time was the

ef¬cient-markets hypothesis. Did you come across that in graduate school?

Shiller: Well, that was already well established.

Campbell: I am just wondering if that was a big part of the discussion

in graduate school at the time?

Shiller: My dissertation was about the expectations theory of the term

structure, which was an ef¬cient-markets model. We talked a lot about

ef¬cient markets.

Campbell: Did you at the time already have seeds of the critiques that

you later mounted so effectively?

Shiller: Well, as I just said, it didn™t seem to me that ordinary people

were estimating autoregressions as was represented in those models. There

An Interview with Robert J. Shiller 237

were already seeds of my later views of excess volatility in my mind. I

noticed that when I estimated autoregressions, if I constrained the sum

of coef¬cients to be one in the short-rate autoregression”that is, to have

a unit root”I could come close to explaining the volatility of long rates.

It bothered me that the difference between this sum and one wasn™t well

estimated, in other words, there seemed to be great uncertainty about

whether there was a unit root. As you know, this has turned out to be a

very contentious issue.

Campbell: This was before Dickey“Fuller and any of the other unit

root literature in econometrics [Dickey (1975), Fuller (1976)].

Shiller: The issues of unit roots were very much bothering me then. I

thought that maybe there was excess volatility. In the case of the term

structure, if there is not a unit root in the short rate process, then there

would appear to be excess volatility in long rates. That unit-root/excess-

volatility issue is not in my dissertation, but I was wrestling with that as

I wrote it.

Campbell: Let™s move forward then for now. You left graduate school.

Your ¬rst job was at Minnesota. What happened there?

Shiller: I had some wonderful colleagues there, such as Tom Sargent

and Chris Sims. But, for me that was the slowest period of my life

in terms of academics. I didn™t publish for several years. I felt that I had

to get on with my personal life. The biggest thing then was that I met

my future wife Ginny. Now we have been happily married for almost

27 years.

Campbell: Then you picked up the theme of excess volatility after a

few years.

Shiller: I had written about a rational expectations model of the term

structure, but, after thinking about it intuitively, wondering what is causing

the big movements in long rates, I cast about for other interpretations.

Campbell: The ¬rst paper was on long-term interest rates [Shiller

(1979)].

Shiller: It seemed tangible and real to me that the long rates were not

moving only for rational reasons.

Campbell: How then did you carry the analysis to the stock market?

Shiller: That was a very simple transition. As you know, the expectations

theory of the term structure is a present-value model, and the ef¬cient-

markets theory of the stock market is also a present-value model. I

thought that the stock market might be an even better example of excess

volatility. Another advantage to the stock market was that one could get

a lot of data. I found the Cowles data, and created from it time series of

price, dividends, and earnings back to 1871. That was what I needed,

since the present-value relation extends over so many years, as you know.

238 John Y. Campbell

Campbell: As I learned from you! That is an interesting point, that

you were doing work on historical ¬nancial data, very early on. Also,

Jeremy Siegel has become known for that. I wonder if the two of you

discussed that.

Shiller: Well, we did. Using only a short recent sample period seems

scienti¬c to many people, because they think that the best data, which

are collected with greatest accuracy, should always be used. So, people

thought that you should rely not on long historical time spans, but

rather on high frequency of sampling. You can get daily data more

recently, while if you sought long historical time series the best you

could get further back was monthly, or annual. So, people wanted to

stay with these recent data, and perhaps they thought that doing so

was being very “scienti¬c.” But, I had a different concept of what “sci-

enti¬c” means, and I thought, from my own reading in science, that

scientists have to look at discrepant data, at things that are not so well

measured.

Campbell: You also were aware that with the long span of the present-

value relation, the testing required a long sample.

Shiller: Well, that seemed very intuitive to me. My student Pierre

Perron and I wrote a paper [Shiller and Peron (1985)] presenting a

Monte Carlo study on power of tests as frequency of observation goes to

in¬nity, holding the sample length, measured in years, ¬xed. In the cases

we studied, power does not appear to go to in¬nity as the number of

observations does. Later, Pierre teamed up with Peter Phillips and devel-

oped a real theory con¬rming this [Perron and Phillips (1988)]. Also, I

should mention the work that my student Andrea Beltratti and I did