A stock price manipulation detecting model with ensemble learning

Abstract

Stock price manipulation has a serious negative impact on the normal market trading environment and order, and among them, trade-based manipulation is particularly common and hidden, which makes it difficult to be detected by regulatory agencies. A variety of algorithms have been applied onto solving this problem, among which RNN-based models perform well due to their adeptness at processing time-series data. Several works have been conducted to integrate RNNs into ensemble models by treating RNNs as traditional classifiers, but the training process and the sub-model connection of traditional ensemble models could drop potentially informative features learned by RNNs. In this paper, we propose a model utilizing the stacking generalization technique that combines predictions of multiple submodules to give the final result while keeping the structure of neural networks. To validate the performance of our model, we collected public data from China Securities Regulatory Commission and created a dataset based on these data. The results show that our model performs better than the baseline in multiple metrics.