Point forecasts are evil.

Economists are asked to make point forecasts, and they oblige. But it’s a dumb thing to do, and they know it. Practitioners, who should know better, rely on these point forecasts far more than they should. Because, in economics and especially in markets, there are enormous error bars around any reasonable point forecast, and those error bars are larger the shorter-term the forecast is (if there is any mean-reversion at all). I can no more forecast tomorrow’s change in stock market prices than I can forecast whether I will draw a red card from a deck of cards that you hand me. I can make a reasonable 5-year or 10-year forecast, at least on a compounded annualized basis, but in the short term the noise simply swamps the signal.^[1]

Point forecasts are especially humorous when it comes to the various year-end navel-gazing forecasts of stock market returns that we see. These forecasts almost never have fair error bars around the estimate…because, if they did, there would be no real point in publishing them. I will illustrate that – and in the meantime, please realize that this implies the forecast pieces are, for the most part, designed to be marketing pieces and not really science or research. So every sell-side firm will forecast stock market rallies every year without fail. Some buy side firms (Hoisington springs to mind) will predict poor returns, and that usually means they are specializing in something other than stocks. A few respectable firms (GMO, e.g.) will be careful to make only long-term forecasts, over periods of time in which their analysis actually has some reasonable predictive power, and even then they’ll tend to couch their analysis in terms of risks. These are good firms.

So let’s look at why point forecasts of equity returns are useless. The table below shows Enduring’s year-end 10-year forecast for the compounded real return on the S&P 500, based on a model that is similar to what GMO and others use (incorporating current valuation levels and an assumption about how those valuations mean-revert).^[2] That’s in the green column labeled “10y model point forecast.” To that forecast, I subtract (to the left) and add (to the right) one standard deviation, based on the year-end spot VIX index for the forecast date.^[3] Those columns are pink. Then, to the right of those columns, I present the actual subsequent real total return of the S&P 500 that year, using core CPI to deflate the nominal return; the column the farthest to the right is the “Z-score” and tells how many a priori standard deviations the actual return differed from the “point forecast.” If the volatility estimate is a good one, then roughly 68% of all of the observations should be between -1 and +1 in Z score. And hello, how about that? 14 of the 20 observations fall in the [-1,1] range.