By Andreas Deppeler1
1Adjunct Associate Professor, National University of Singapore
Artificial Intelligence (AI) is everywhere in financial services. Insurance marketing and underwriting models are based on behavioural micro-segmentation; call centre agents rely on natural language processing and voice transcription; banking compliance departments are using machine learning to reduce the number of “false positives” in anti-money laundering transaction monitoring. Nevertheless, financial services firms still seem to be slow in adopting AI.
There are three main reasons for this.
It is true that “the machine learning race is really a data race”.1 Even though financial services firms are traditionally data-rich, many struggle with outdated legacy architecture, organizational silos and poor data quality. Executives are now realizing that data is a strategic asset that needs to flow freely throughout the organization. Building a data-centric organization therefore means going beyond compliance with regulatory obligations like BCBS 239 or creating a chief data officer role: it means generating new business value by unlocking data.
As a discipline, Machine Learning is still in its infancy. We speak about “Data Science”, but heuristic and iterative methods of algorithm selection and tuning resemble alchemy more than science.2 In the past 12–24 months, as companies’ unrestrained data collection practices have become better known, the public mood and intellectual discourse turned more sceptical and cautious.
Some augur a dystopian future in which corporations use AI prediction engines to steer human consumers towards “guaranteed commercial outcomes”.3 Others bring to light the failure of data protection law to protect data subjects from potentially discriminatory “inferential analytics”4 drawn from AI and big data. In recent months, financial regulators have started to publish guidelines on governance, accountability and risk management of AI for decision-making.5 Many expect high-risk AI model development to be subject to stringent regulatory and professional licensing standards in the future.
Although every major university has created data science and business analytics curriculums in recent years, industry demand for graduates still outstrips supply. Furthermore, the top AI talent seems to be drawn towards technology firms rather than traditional banks or insurance companies.
What can financial services firms do to accelerate their AI journeys? How can they build explainable models without relying on an army of data scientists, consultants or external vendors? How can they train models across larger data sets while keeping all data private and secure? Answers may be found in two recent AI innovations (automated machine learning and federated learning) that are presented in this chapter. This non-technical introduction will hopefully encourage readers to dive deeper into these technologies and adapt them to their own corporate environments.
The machine learning model development process can be broken down into four steps:
Depending on data infrastructure and quality, data preparation can take anywhere from a few days to several weeks. Feature engineering typically requires close collaboration and exploration between data scientists and domain experts. Algorithm selection and tuning is a time-consuming, subjective, repetitive and manual process – an art rather than science.
Automated machine learning promises to do all of this in a fraction of the time. The user drags and drops the training data files, selects the target variable (“What do I want to predict?”), defines the model category (e.g. regression or classification) and specifies runtime parameters. The platform autonomously performs data preparation and feature engineering, then iterates through a library of common models, tunes and trains each of them and ranks them by accuracy and speed. The user deploys a highly ranked model to production with the click of a button. The entire process can be completed in a few hours. Commercially available automated machine learning platforms offer modules for explainability and create standard documentation for model validation and model risk management.7
Automated machine learning increases the productivity of data scientists by reducing the time spent on mundane tasks of model development. Some firms are even deploying automated machine learning outside of data analytics functions, encouraging business users to experiment with AI. Over the next few years, off-the-shelf automated machine learning will likely establish itself as an efficient tool for developing low-risk models (e.g. marketing or prospecting), but data professionals will still be needed to develop and maintain high-risk models (e.g. underwriting or trading).
Standard machine learning requires centralizing the training data on one machine or in a data centre. In federated learning, a shared global model is trained across many participating clients that keep their training data in local environments. Google coined the term “federated learning” in a paper8 in February 2016, then wrote about it in their AI blog9 in April 2017 and made available an open-source software called TensorFlow Federated10 in March 2019. Another popular open-source library for encrypted, privacy-preserving deep learning is PySyft,11 which is maintained by the OpenMined community.12
In federated learning, locally computed model training results (but no training data) are encrypted with a private key and sent to a central coordinating server. The server combines encrypted results from thousands of local models and only decrypts the average update, which is then used to improve a shared model. Once the shared model is trained and tested, all local models are updated.
Federated learning has been used for text and image prediction on phones and tablets, where training data are privacy sensitive. Soon, most smartphones13 and many IoT devices will be equipped with AI chipsets and connected through 5G. This will allow AI to move from the cloud and data centres to the “edge”. Federated learning allows those decentralized compute resources to train machine learning models in a privacy-preserving way. Potential applications include self-driving cars, industrial devices,14 smart buildings15 and medical diagnostics.16
In financial services, federated learning has been proposed for small business lending, anti-money laundering transaction monitoring and fraud detection.17 In a federated system, a shared model would be trained with aggregated model updates from several participating banks. Since each participant will have a slightly different data model, a trusted third party would be needed to standardize the model inputs and formally own and maintain the shared model.
Due to the high cost of compliance and the limited amount of good training data, the potential for federated learning in financial services is vast. The regulators will play a key role in encouraging the formation of industry consortia and in creating “sandbox” environments for experimentation and adaptation.