Technologies such as robotic process automation (RPA), natural language processing (NLP), and machine learning (ML) are revolutionizing the way businesses are run, be it any industry. Combined with the power of IoT and analytics, these technologies are can enable banks to improve operational cost efficiency and accuracy. Here we discuss how these technologies can automate some of the manual aspects of operational loss activities.
Banks typically have tedious processes for the management of their internal and external loss data. Picture this: sourcing data from various divisions of a bank and manually mapping incidents with various business lines, event types, risks, and so on, sometimes using end user computing (EUC) by way of spreadsheets, is complex to say the least. Moreover, there is a huge probability of human error as the amount of loss data increases.
Machine learning can be used in these areas, freeing up human resources to take on more productive work. But for that, banks need to ensure that all the relevant scenarios for automation are covered and then train the system for automation of manual tasks contained within the overall process.
Data preparation: To start with, we need a thorough loss database, which has significant number of incidents with all possible scenarios covered (that are required for the machine to get trained on). This is essential, as the lack of enough data will yield poor results, leading to under-utilization of technology. Data preparation will also result in higher accuracy in data mapping. The datasets needed will be part of the overall loss data sourced by the bank and then customized to cover all required scenarios.
’Noise’ cleaning: It is here that all the text clean-up activities like case conversion, removal of punctuations, special characters, and numbers are performed. Common English words, bank names, currencies, countries, and so on, are also removed. The benefit of this process is that the machine learning model will be restricted to focus only on specific keywords, for the mapping exercise to be as successful as possible. Word stemming is also performed in this stage to avoid confusion with similar words. For example, ’indexing’ may be stemmed to ‘index’ to achieve better results. This is done using NLP techniques and is directly applicable to incident description. This can replace the traditional method of managing loss data, where an individual has to manually read the incident description to perform this process.
Model creation: In this stage, the number of times a pertinent word appears in the incident description, is looked for. This may be a common English word, which is either removed from the model to avoid ambiguity while performing the mapping activity or may be of high importance, so that it can be used for mapping an incident to a risk, business line, event type, and so on. A machine is, therefore, trained in a way that it can identify which word to include and/or exclude for mapping purposes. Based on frequently appearing words in all the incidents, a word cloud is generated to focus on those sets of words and treat them accordingly.
Model calibration: The model is calibrated and validated until a significant level of accuracy in data mappings is achieved. The level of accuracy may vary from bank to bank, however the rule of thumb is – higher the accuracy, better is the calibration level. Again, these data mappings relate to business lines, event types, risks, and/or any other parameter as defined by the bank.
Model validation and deployment: Once the machine achieves a significant level of accuracy, it is fully validated against the dataset that is used for tuning various parameters of the model (that was selected for training the machine) and finally deployed.
Machine learning can provide better outcomes to manual aspects of the operational loss process
The output of the aforementioned activities using automation technologies can then be used to draw actionable insights that are further used for faster decision-making purposes – mostly financial and/or investment related decisions.
There are a lot of regular activities that organizations have to constantly deal with. For example, traditionally, they have been conditioned to work with limited resources, extract quality reports at a particular frequency, manage key stakeholders, investor relations, and so on. In the process, they are unable to ‘test the waters’ in newer areas such as automation, which is furthered hampered by the risk-averse mindset that prohibits them from venturing into unknown areas. Organizations, therefore, need to break the shackles of the ‘tried and tested’, and be open to embracing risks so as to adapt and transform continuously. This will also enable them to respond to constant challenges and threats, while delivering new business capabilities and models in an agile manner.