Transforming Financial Risk Management with AI and Data Engineering in the Modern Banking Sector

Abstract

Machine learning has become a more potent force after Google’s creation of a revolutionary new machine learning product, TensorFlow, in 2015. This product was free and open to all. Anyone could assemble a small group of smart, eager juniors, give them a few weeks of training, and by pooling calculations across many high-tech GPUs lets them train a smaller, simpler, unlabeled machine learning model on a few decades of data. Also in 2010, software development became an intensely competitive area for the banking sector. Important new online platforms spawned a whole new field of firms producing exploitable software.

The needs of finance and banking are often thought of as a combination of raw memory (storage), connection speed (bandwidth), and thought speed (the speed at which logical instructions can be executed on that memory subject to many constraints). Currently, there is a multitude of suicides by regulators, directors, and traders resulting from unavoidable miscalculations. Though too late to stop these jumps and errors, they should be reducible nevertheless.

At both the strategic and tactical levels, “optimal” decisions depend on minor details of the data processing, machine learning field, algorithms, past decisions, etc. These optimal processes can now be assembled into semi-autonomous teams that should learn optimally and robustly in real-time on as far back data as exists. Why this all works is due to four “empirical assumptions” that form a dichotomy separating most data analysis either out-of-context or in-context. Each synthetic, comparative “test” believed important can be broken into numerous possible “cooking experiments”, i.e., deconstructions of, say, the Filter Test, that could also be devised so as to continue learning for decades or centuries into the past. A bank “information platform” architecture is proposed to solve each singular “optimal decision” for multiple tests, adaptively improving on tests of performance, relevance, cost, etc. The most robust finds exit “jump” rates that make for the biggest profits and the least suicides. Each side of a trade at an exchange could then expect to succeed equilibrium betting millions of times on the price differences without ever losing.