A Hierarchical Futures Data Analysis Framework for Multiple Accounts Detection

In futures markets, participators are required to report their Multiple Accounts (MAs) to the market supervisors. However, a small portion of traders tent to use MA to extend their position limit or even manipulate the market prices. To detect these traders, a systematic approach have to be developed to automatically assess the behaviour similarity of each account pair. In this paper, we propose a Hierarchical Futures Data Analysis framework (HierFDA) with multiple data processing components and an unsupervised machine learning model to detect MAs. Since we aim to investigate the whole market with millions of accounts, HierFDA is designed to be memory-efficient and can be easily parallelized. It consists of two layers of components to filter out the pairs of accounts with different levels of similarities. As most of the market participators are 'regular' individuals and only a tiny portion of the accounts are MA, we apply a deep auto encoder for unsupervised anomaly detection to identify the accounts with extremely high trading similarities. By training with sufficient samples which are not previously identified as MA, the deep neural network can automatically balance the importance of the similarity features on the results. The offline test results show that HierFDA has superior robustness to the sample noise. And we also deploy HierFDA to the production environment of Dalian Commodity Exchange. The online results show that over 97% identified MA groups are approved by the expert and the rest are merged with previous manually discovered groups.