A Note on the Growing Evidence of Aggregate Stock Return Predictability

J. Bradford DeLong (2008), "A Note on the Growing Evidence of Aggregate Stock Return Predictability"

Get your own at Scribd or explore others:

http://www.scribd.com/doc/8264126/null

Comments

[GaryD]

Possibly a better explanation of stock price movements:
http://www.youtube.com/watch?v=RAXdie_gifI

[MattYoung]

HT to Gary, great sequence. good movie. Wish we had more financial comedies, they should be a genre.

It sounds generational, the predictability, as the movie sequence suggests.

[Jim V]

Some of us will take this as reassurance that our 401k's will recover, if we can just hang on for another 20 years, although I was hoping for five years myself. However, I expect a deeper look at what the statistics are saying, mixed with some of Taleb's "Black Swan" pessimism, would cure that, or at least make us wonder whether we are part of the "aggregate" or not.

[john]

Given the moving averages, I'd guess there was a staggering amount of autocorrelation in the residuals, which of course will bias your standard errors and the resulting t-statistics.

[How stupid do you think I am when I calculate standard errors?]

[nathan]

It looks smart to have Shiller in references ... most people would probably benefit from reading what he has to say.

Shiller had interesting comments in NYT recently. The stock market is only one component of the economic story.

[john]

My apologies, Brad, I should have considered the source - or done the analysis myself to check before commenting.

[Alex Tolley]

The problem I see with this is that the second graph is not using "out of sample data". because the the data sets are overlapping, your claim that it is not "junk science" is very much weaker than if the 2 data sets were disjoint.

[john]

How is r calculated from the Shiller spreadsheet? I can guess, but I guess wrong.

Robustness questions I'd like to answer, for personal satisfaction:

How sensitive is the significance to using, say, 15 or 25 year forward looking returns, and 7 or 13 year moving averages of earnings?

Does the coefficient of interest seem to change over time?

No doubt others will occur to me once I get the r calculation worked out.

[Mats]

Returns of risky securities should be very predictable if risk is highly mean reverting. And yes - just look simultaneously at the stock index (e.g. SPX) and volatility (e.g. VIX) charts!

[Sammy]

This scribd thing always makes my browser crash. Where is the pdf version?

[Gorobei]

BDL, cas you know, orrecting t-stats, etc, in the face of moving averages is notoriously tricky. What method did you use?

Actually, when a quant comes up with some like this, I usually just ask them to generate a bunch of synthetic data-sets by shuffling the annual log returns, and look at where the real results lie within the synthetic results. If the result still looks significant, we then try correcting for serial correlation in vol and shuffle away again. If still good after that, we start getting interested.

Not that I don't believe quants bearing multi-decade models to explain 130 years of data or anything.

[john]

The autocorrelation of the residuals being what it is - over 0.99 for a one-period lag when I ran the model - you'd be much better off taking first differences and running the regression. Then you get significance over the smaller subset too, or at least I did, and the autocorrelation is much, much lower. But I still doubt the variable definitions are meaningful, even though I'm sure they are conventional.

When I was asked to estimate the stock price of a proposed spinoff from a Fortune 20 company some 10 years ago, we did a lot of work - since the proposed spinoff was of over $20B current dollar value - and it did not appear that earnings of more than about 4 years old were relevant in stock valuation. Furthermore, IIRC about half the weight was on the most recent year, and there was a floor at slightly above $0 earnings (negative earnings are treated as small positive earnings.) All of this was very clear from modeling hundreds of companies over 20+ years, and we were pretty sophisticated about it. I simply can't believe in a variable definition that implies that, e.g., HP's earnings in 1998 are as relevant to its current stock price as its earnings in 2007 - and aggregating to the level of the S&P 500 won't change this, since if five years ago is irrelevant to all the stocks, it's irrelevant at the aggregate level too.

[Jrossi]

The thinking behind the efficient markets hypothesis has always puzzled me. The markets knows best. Well, the market is the aggregate buying and selling behavior of human beings. How many fully informed and rational human beings do I know? My gynecologist father-in-law, who gets investing tips from his golf buddies? No, not him. The gals in the office at work? Probably not them. The neurologists in my old roommate's medical group? Not them. The millions pumping money into stocks every month in their 401Ks, no matter the price. Not them either. And of course we've heard the stories about how rational investors will always screw these lunkheads and, through arbitrage or something, send prices back to their fundamental values. Yeah, right. Go tell it to LTCM Nobelists. Of course there are rational markets participants out there. They're called value investors.

[Ljean]

http://www.youtube.com/watch?v=7jO3Bm9Si64