Enhancing Abnormal-Behavior-Based Stock Trend Prediction Algorithm with Cost-Sensitive Learning Using Genetic Algorithms

Stock market trend forecasting is an important and popular topic in academia or industry, and many studies have been proposed to handle this problem. Because stock market transactions are dynamic and nonlinear systems, it is difficult to predict through a single feature. Therefore, in the previous work, an algorithm was proposed to build prediction model by using abnormal behaviors extracted from the given data as features and transfer learning. However, when using transfer learning to generate more training instances, it may cause negative transfer. Besides, hyperparameter for the previous model should be tune to find a more effective prediction model. To solve the two problems, we propose an enhanced stock trend prediction model in this paper. For the first one, we modify the original fitness function by combining the concept of the cost-sensitive learning. As a result, the proposed approach can have the ability to avoid instances that could lead to negative transfer. For the second problem, we utilize the genetic algorithms for searching the suitable hyperparameter to construct the prediction model. At last, experiments were made to show the effectiveness of the proposed approach.