英语论文网

he time of the task consists of book value at the beginning of the period ( t−1 bv ),expected earnings of the current period ( t x , for the period staring at t-1 and ending at t), and thenormal capital charge rate for the period ( t r ). Abnormal earning is defined as the difference betweenanalyst earnings forecast (best knowledge of actual earnings) and the earnings number achieved under
growth of book value at a normal discount rate. Underlying this definition is the idea that analystearnings forecasts are essential signals of firm valuation (following Frankel and Lee 1998, Francis et
al. 2000, and Sougiannis and Yaekura 2001).Next, I demonstrate how to improve the implementation of the RIM. I describe the necessary
procedures starting with a naïve regression. Then, I point out the violations of this naïve regression,and seek improvement by addressing these violations. Specifically, for RIM regressions to produce
reliable results, t v must have a normal zero-mean distribution and meet the statistical regressionassumptions. However, the regression assumptions are often not met, due to strong serial correlation
in t v . Serial correlation arises when a variable is correlated with its own value from a different timelag, and is a notorious problem in financial and economic data. This problem can be addressed by
using regressions with time series errors to model the properties of t v . My diagnostics also show
conditional heteroscedasticity in t v , which can be addressed with GARCH modeling. My procedureto identify the time series properties of t v is as recommended by Tsay (2002) and Shumway and
Stoffer (2005). I show that, by jointly estimating the RIM regression and the time series models of t v ,forecast errors are substantially reduced.My demonstration is based on SP500 firms, using 22 years of data spanning 1982 – 2003 toestimate the prediction models, which I then use to predict stock prices in a separate period spanning2004 - 2005. The mean absolute percentage error obtained can be as low as 18.12% in one-year-aheadforecasts, and 29.42% in two-year-ahead forecasts. It is important to note that I use out-of-sampleforecasts, whereas many prior studies use in-sample forecasting, in other words, they do not separate
the estimation period from the forecast period. In-sample forecasts have artificially lower forecasterror than out-of-sample because hindsight information is incorporated. However, to be of practical
value, forecasts must be done beyond the estimation baseline.
For a brief review of prior results, prior valuation studies based on the RIM have focused
more on determining value relevance, i.e., the contemporaneous association between stock price and
accounting variables, not to forecast future prices. As will be noted in this paper, the harmful effect of
autocorrelation is not apparent in estimation or tests of association, therefore value-relevance studies
may not have to address this issue. However, when the RIM results are applied for forecasting, it
yields large errors, although the RIM is found to produce more accurate forecasts than alternatives
such as the dividend discount model and the free cash flow model (Penman and Sougiannis 1998,
Francis et al. 2000). Forecast errors are disturbingly large, and valuations tend to understate stock
price (See discussions of large forecast errors in Choi et al. 2006, Sougiannis and Yaekura 2001,
Frankel a