Predicting stock price movements: regressions versus economists

Abstract

The forecasting performance of the Livingston survey and traditional prediction models of stock prices is analysed. The survey forecasts look similar to those from a ‘too large’ prediction model: poor out-of-sample performance and too sensitive to recent and irrelevant information.

Notes

¹The data of the Livingston survey are from the Federal Reserve Bank of Philadelphia (http://www.phil.frb.org/). The monthly S&P 500 and S&P Industrials price indices are from S&P's Trade and Securities Statistics (1976), Compustat and DataStream. The dividends on the S&P 500 and the inflation rate are from Shiller's (2000) homepage (http://www.econ.yale.edu/%7Eshiller/data.htm). For a discussion of the historical S&P series, see Jones and Wilson (2002). The 3-month T-bill rate is from FRED II (http://www.research.stlouisfed.org/fred2/).

²Strictly speaking, we would like to have survey data on the expected capital gain ϵ _{t – 12} (P_t /P_{t – 6}), but the survey only contains information on expected index levels. Therefore, ϵ _{t – 12} P_t /ϵ _{t – 12} P_{t − 6} is used as an approximation.

³The ‘out-of-sample R ²’ is , where s is the first period with an out-of-sample forecast, r_t the actual value in t, the model forecast of the value in t and the recursively estimated average value.

⁴Other nonlinear restrictions, for instance, using the max. of the prediction and the negative of the recent dividend yield, give similar results.

⁵In contrast, Dokko and Edelstein (1989) found that the survey was unbiased in the subsample 1955 to 1985.

⁶ There is good theoretical and empirical evidence that combining (averaging) forecasts typically reduces the forecast error variance (see, for instance, Bates and Granger, 1969).

⁷Bacchetta et al. (2008) also found a negative correlation between survey expectations and the dividend yield (using the UBS/Gallup poll 1998 to 2003), but a positive correlation with an interest rate.