Machine learning algorithms, such as the genetic algorithm, have often been applied to financial problems, but not enough is known about how to systematically incorporate financial knowledge into these generic learning algorithms. The general hypothesis of this paper is that semantic similarity among financial concepts can be exploited in a hybrid genetic algorithm. A Knowledge-guided Genetic Algorithm for Forecasting is introduced to predict the values of financial statement variables. The mutation operation is guided by domain knowledge to make small or large changes in an organism. The algorithm makes a bigger (or smaller) change in the organism when the variables being forecast have higher (or lower) variability. The specific hypothesis is that the use of problem-specific knowledge improves the prediction accuracy. The experimental results show that the use of domain knowledge improves the performance of the algorithm. The knowledge used in this experiment would reasonably be extended in various ways to be used by a refined genetic algorithm.