We are

Career at

News at


  • Card Schemes
    Card schemes enable simplified and guaranteed exchange of money between merchants, customers and their banks, by operating international networks and setting uniform standards. More specifically, they define rules for the routing of payment authorizations and settlement requests in point-of-sale and e-commerce transactions between merchant acquirers and card issuers, as well as ATM withdrawals or purchases with cashback transactions.
  • Gradient boosting vs. deep learning. Possibilities of using artificial intelligence in banking
    Artificial intelligence is growing in importance and is one of the most discussed technological topics today. The article explains and discusses two approaches and their viability for the utilization of AI in banking use cases: Deep learning and gradient boosting. While artificial intelligence and the deep learning model generate substantial media attention, gradient boosting is not as well-known to the public. Deep learning is based on complex artificial neural networks, which process data rapidly via a layered network structure until a decision regarding a previously defined problem is made in the last layer. This enables the solution of complex problems but can lead to insufficient transparency and traceability in terms of the decision-making process, as one large decision tree is being followed. The German regulatory authority BaFin already stated that in terms of traceability no algorithms will be accepted, that is no longer comprehensible due to their complexity. In this regard, the authors argue that gradient boosting might be a viable alternative, as this model bases its decisions on several smaller decision trees, which are easier to interpret and understand. In addition, gradient boosting requires less data than deep learning to become functional. Statements about the dominance of either of the models are not possible, as both of them provide good results in terms of classifying high-dimensional data. However, considering regulatory requirements, such as transparency in the financial industry, gradient boosting might be a preferable approach to artificial intelligence.
  • Artificial intelligence in banking: Deep learning vs. gradient boosting
    Customer behaviours in retail banking have changed in recent years. Since 2014, customer willingness to switch bank accounts has doubled. Banks have various response options to sustain customer loyalty. However, those potentially migrating customers need to be identified beforehand. This can be done via AI-driven analysis of master data and transaction data. In 2016, Spanish bank Banco Santander published a public tender for this customer „churn case“, which was analysed by more than 5,000 data analysis teams worldwide. COREai participated in the tender and tested the two AI models gradient boosting and deep learning, comparing their results on various dimensions. The article elaborates on COREai‘s analysis strategy. In terms of the results, gradient boosting surpasses deep learning regarding its predictive accuracy and performance. The comparison was quantified by using a misclassification rate, which represents the proportion of incorrectly predicted data points and customers. During testing the deep learning method misprognosed 23% of the data, while gradient boosting only came to wrong results in 3.3% of the cases. Additionally, gradient boosting outperformed deep learning with shorter learning and prediction times. Also in terms of interpretability of the results, gradient boosing excels over deep learning, due to its easier interpretable functional structure using individual decision trees instead of one complex layered neural network, as is the case with deep learning. This potentially enables more effective drawing of conclusions regarding the reasons of a customer’s account termination. Taking those findings into account, the authors argue that gradient boosting is a more promising solution for use cases, as the one defined by Banco Santander. Nonetheless, no general statements about the dominance of either of the models can be made, as it is dependent of the use case’s data patterns. For example, deep learning remains unchallenged in terms of image or speech recognition.