Predictability of convective rainfall in a numerically simulated storm is assessed in conjunction with nonlinearity imposed by systematic perturbations in water vapor using a three-dimensional cloud model (ARPS). Nonlinearity is explicitly quantified using the tangent linear approximation, and predictability is measured by the noise-to-signal ratio. It is found that the behavior of nonlinearity at an early period of error insertion exerts strong influence on error dynamics and predictability in the future. The rate of increase in nonlinearity is larger for perturbations with larger magnitude and positive direction. Although dynamics of individual storms are sensitive to specified perturbations, the predictability limits generally go beyond 140 min after perturbations are introduced. The major contribution to the total error in the domain-integrated accumulated rainfall fields comes from the phase error rather than the amplitude error, especially under the region of secondary storms. The location and general behavior of the main storm (supercell) is predictable in a longer timescale than suggested by Lorenz's theoretical analysis. On the basis of the results from this study, some implications for practical storm-scale predictability are discussed.