This paper investigates the usefulness of analytical models of cost behavior, which are typified by CVP analysis, when applied to earnings forecasts. For this purpose, we empirically examine the three models :（ 1） the basic “CVP model” which disaggregates costs into fixed and variable, （2） the “difference model” which relies on first-order differenced costs and sales revenue, and（ 3） the “asymmetric model” which considers cost stickiness. As a result, it appears that the “difference model” forecasts more accurately and the “asymmetric model” forecasts more unbiasedly than the “CVP model” does.