This paper integrates competing earnings prediction models into a composite model that considers the joint predictive ability of Mean Absolute Percentage Forecast Errors - Composite Model vs. Univariate Benchmark Models Across Size (Market Value) Quintiles (Table present) current earnings and current price. The results show that current earnings play a key role in predicting future earnings when the variance of earnings relative to that of price is low. In contrast, the predictive power of current earnings is substantially lower for firms with a high variance ratio; current price assumes greater predictive significance for these firms. The composite model is superior overall to the univariate models tested, and is substantially so for the group of firms with a high variance ratio. This result indicates that the effectiveness of the composite model varies with the variance ratio, which appears to be the appropriate variable for conditioning earnings forecasts. In general, the above findings suggest that failure to condition forecasts may explain the weak results obtained by prior studies that used complex time-series or non-earnings-based models visa-vis a simple random walk.