Being Wrong in an Interesting Way

My friend Mark Hulbert once had a philosophy professor at Oxford, who distinguished two ways of being wrong: “You can be just plain wrong, or you can be wrong in an interesting way.” In the latter case, Mark explained, correcting the wrong reveals a lot about the underlying truth.

The willingness to learn from our own errors, and those of others, is how a great deal of learning comes about. That’s certainly true biologically, where most of our skilled movements rely on feedback and progressive error-correction. It’s also true of invention and research. As Thomas Edison said, “I have not failed. I’ve just found ten thousand ways that won’t work. Many of life’s failures are people who did not realize how close they were to success when they gave up.”

In the past few decades, I’ve tested hundreds of propositions about the financial markets, many advanced by other market analysts, against historical data in order to test their validity. A couple of years ago, for example, several analysts argued that the ratio of U.S. market capitalization to GDP was no longer a valid measure of valuation, because of the substantial increase in foreign revenues of U.S. corporations. Others argued that U.S. GDP should be replaced by global GDP in the denominator. Examining decades of historical data, however, one finds that while foreign revenues of U.S. corporations are greater than in the past, the change has been fairly smooth over the past several decades, and the relative impact is far smaller than investors seem to imagine (one can’t simply reduce current valuations by one-third and leave all prior valuations unchanged). Meanwhile, replacing U.S. GDP by global GDP produces an apples-to-oranges measure that, not surprisingly, weakens its relationship with subsequent market returns.

Still, both proposals were wrong in an interesting way. The foreign revenues of U.S. corporations do matter, and by explicitly estimating them, we developed a valuation measure (MarketCap/GVA) that is even better correlated with actual subsequent S&P 500 total returns, and is more reliable than any other valuation measure we’ve examined in market cycles across history (including MarketCap/GDP, Tobin’s Q, Shiller’s CAPE, price/forward operating earnings, the Fed Model, and numerous others).

As Mark observed, correcting the error "reveals a lot about the underlying truth."


If we’re going to discuss anyone’s errors, it’s best to begin with my own. Even those who are familiar with this narrative are encouraged to review the brief discussion below, not least because this one includes an illuminating chart.

The investment discipline I advocate is decidedly focused on the complete market cycle. Though I’ve typically been frustrated during late-stage bubble advances, I’ve also shifted to a constructive or leveraged investment outlook after every bear market in more than three decades as a professional investor. Despite periodic frustrations, our value-conscious, historically-informed discipline put us ahead in every complete market cycle through 2009.

In late-2008, during a market collapse that we fully anticipated, and with the S&P 500 down more than 40% from its highs, I shifted to a constructive market outlook. As I noted at the time, the prior overvaluation of the market had been erased, and prospects for future long-term returns had substantially improved. Unfortunately, the behavior of the market and the economy (particularly employment losses) became entirely “out-of-sample” from the standpoint of the post-war data on which our market return/risk classification methods were based, and in 2009, I insisted on stress-testing them against Depression-era data. During the Depression, valuation levels similar to those of 2009 were followed by a further loss of two-thirds of the market’s value, and our existing measures of “early improvement in market action” would have been repeatedly whipsawed. Given that potential, I insisted on boosting the robustness of our methods to Depression-era data, post-war data, and also “holdout” validation data. At the time, I discussed this effort as our “two data sets problem,” and in the midst of that ambiguity, we missed a substantial rebound that both our pre-1999 methods and our current methods could have captured.