How Well Do the REY/RTV Models Catch Turning Points?
April 28, 2010 • Posted in Fundamental Valuation
A reader asked: “How far back have you tested the REY and RTV models. How did they perform during key market turning points, such as January 2000, October 2002-March 2003, October 2007-August 2008 and March 2009. Specifically, do you have the 3-month, 6-month and 12-month S&P 500 Index forecasts from August 1, 2008 and March 1, 2009 for the RTV, REY and REY-A Models, similar to what you publish daily?”
Per the descriptions in “Real Earnings Yield (REY) Model Details” and “Reversion-to-Value (RTV) Model Details”, the model regressions extend back to March 1989. The limiting data are the actual S&P 500 operating earnings publicly available from Standard & Poor’s and used in the “Aggregate Operating Earnings Forecast” that provides inputs to both models.
The daily charts on the home page summarizing current indications from the the models are ephemera, but the models can recreate past projections. A sample starting point of March 1989 does not reasonably support recreations around the 2000 and 2002-2003 extremes in the U.S. stock market using data known at those times. However, stepping the models back to conditions at the ends of July 2008 and February 2009 (and ignoring the fact that model specifications have changed a little since then) results in the following projections:
The scatter plots for the models convey a broader sense than these two observations of the uncertainties associated with using linear best fits to make extrapolations. For example, the following scatter plot relates the forward real earnings yield to the six-month future return for the S&P 500 Index (REY Model) as of the end of March 2010. The R-squared statistic of 0.14 indicates that the forward real earnings yield explains 14% of the variation in the 6-month future return. The six-month projection for the index return comes from: (1) finding the forward real earnings yield on the horizontal axis; (2) moving vertically to intersect with the best-fit line for the scatter; and, (3) moving horizontally left to locate the expected return on the vertical axis. The scatter suggests this approach may work better for high than low real earnings yields.
The REY Model has greater predictive power (less dispersion from the best-fit line and higher R-squared) than the RTV Model. However, quite a few market commentators use RTV-like reasoning (market price-earnings ratio) to assess market attractiveness.
Both models tend to lag sharp market turns, during which actual operating earnings also tend to make sharp turns. See the backtest at “Aggregate Operating Earnings Forecast” for a visualization of the lag in the earnings forecast that goes into both models. The earnings forecast assumes both short-term momentum and longer-term reversion (at a “typical” pace) to a long-run growth rate.
There is a similar lag in the “Inflation Rate Forecast” input to the REY Model. The lags in the earnings and inflation forecasts are common to mechanical forecasting models that regress and extrapolate historical data.
Other sources of error may affect the models, but probably not as systematically as these lags. See the ends of both model descriptions for discussions. Samples for the models are fairly small for statistical inference.