We often receive questions relating to the ensemble method that guides our hedging strategy. Along with last week's comment (Notes on Risk Management ), the following section is intended to provide a broad overview.
One of the main approaches we use to estimate return and risk prospects is to group current market conditions among historical instances that are most similar. Each point in history is defined by various "features" based on a broad range of key factors, including valuations, trend-following indicators, market breadth, sentiment, credit spreads, economic factors, overbought/oversold measures, and so forth. In order to make the analysis less dependent on any particular historical period (e.g. postwar data, bubble-era data, Depression-era data), or any single set of indicators, we extend this analysis to a very large number of randomly selected sub-samples across history.
This sort of analysis is an example of an "ensemble method," which has several benefits, the two most important being on measures of "accuracy" and "robustness." It's easy to fit a model to past data, but those models often break down quickly in new data. So to evaluate accuracy, we estimate return and risk on data that the model has not "seen" previously, and find that the ensemble approach generally performs better than alternative methods. Equally important, the ensemble is robust to very large changes in the underlying economic environment, because randomizing over numerous sub-samples of history reduces the likelihood that the model is "over-fitted" to a particular economic environment.
I wish I could say that we anticipated the depth of the 2008-2009 credit crisis so completely that I developed these ensemble methods in advance, already confident in how they would have performed even in Depression-era data. Unfortunately, that's not the case, and shareholders are well aware of the challenges we went through in stress-testing our approach against other periods of credit crisis.
Though these weekly comments forewarned much of what actually occurred during the credit crisis, I certainly didn't anticipate what I still consider to be terrible policy mistakes - particularly the absolute unwillingness to restructure bad debt, in preference for kicking the can down the road with public funds. It was a far cry from how U.S. regulators had responded to the S&L crisis, and how other international banking crises had been successfully addressed (for example, in the early 1990's, the Swedish banking crisis was durably resolved by the government taking receivership of a large portion of the banking industry, wiping out existing shareholders, writing down bad assets, and then taking the banks public to recapitalize them under new owners).
For anyone who was responsible for investing the funds of others, the proper response to the 2008 crisis was to stress-test every method, though I'm not convinced that much of Wall Street has stress-tested anything at all. For us, stress-testing meant taking our models to Depression-era data, because it was clear that events of the time were largely "out of sample" from the standpoint of post-war data. At the time, we were basing our estimates of market risk and return on data since about 1950, which I had - incorrectly - believed was sufficient to capture "modern" market behavior.
While our existing hedging approach performed well in that Depression-era data overall, the occasional losses were far deeper than I was willing to risk for our shareholders. The result was what I called a "two-data sets" problem, which demanded that our hedging methods perform well, out-of-sample, and with tolerable drawdowns in data drawn from both post-war and Depression-era periods. We reached a satisfactory solution in 2010 with the introduction of our ensemble approach. The cost was that in hindsight, my decision to fully stress-test our methods shot us in the foot by missing a rebound in 2009 that we should not have missed, had our present approach been already in hand.
In real-time, our hedging approach has repeatedly demonstrated value over complete bull-bear market cycles, both adding returns and defending against severe market losses (exceeding 50% downturns twice in the last decade). We'll certainly have periods where we appear remarkably out-of-step with the prevailing trend of the market, particularly in overvalued, overbought, overbullish periods of speculation. But defending against losses in these periods is essential to risk management, despite the tendency of bulls to declare victory at halftime. I remain confident that our investment approach will continue to navigate the financial markets effectively over the course of complete market cycles.