Application of Machine Learning in Transaction Business of Investment Bank -Abnormal Transaction Recognition (Part 2)

Important global systemic banks that are extremely active in the financial market, such as Nordea, have tens of thousands of transactions settled every day. It is quite not cost-effective to control each transaction manually by the controller, which may also bring manual errors (such as different controllers giving different judgments on the same transaction). In particular, this process requires control personnel to have extremely rich trading and market experience and many controllers are former traders of the bank. Arranging these employees on such a time-consuming, laborious and repetitive work is not only a great waste of human resources, but also not conducive to the bank for retaining these employees.

In the bank database, there are tens of millions of complete records of abnormal data. We can use these data to train and fit a deep neural network to predict transaction control.

We finally implemented a deep neural network in which each transaction and the real-time price data related to it are treated as a sample, and the final output of this model is a Boolean variable, that is, whether the transaction is an abnormal transaction.

At the same time, we integrate this model into the existing business process. When a new transaction is reached, the deep neural network first predicts whether the transaction is abnormal, and then the results are reviewed by experienced controllers to decide whether to adopt the machine’s suggestion. All the final results will be saved in the training sample library of the network and will be included in the training sample when the model is recalibrated next time (the existing model is recalibrated every two hours).