Although artificial intelligence (AI) has been around since the 1950s, we are currently riding the peak of the Gartner Hype Cycle. Separating the reality from the hype has therefore become a challenge. There is no question that machines can automate more human tasks than ever before, but with this promise, there is also a gap between the achievable and the expectation. There is, in other words, a sense that not much is actually being done by many individual organisations. This suggests that organisations want to use AI and machine/deep learning, but are perhaps not clear on how to do so.
This may be particularly true of the financial sector. Traditional banks and insurance companies are being pressed hard by the rise of new fintechs, which are breaking into profitable parts of the market such as payment provision. Coupled with deregulation of certain sectors and increased regulation of others, such as data protection, banks are up against the wall in more senses than one.
Even without the pressure from fintechs, traditional banking practices have their challenges. Consider automated teller machines, or ATMs, otherwise known as cash machines. Many of us will have experienced the situation where all the ATMs in a certain location have run out of money at the same time. It is annoying for customers to find empty ATMs, but it is also annoying for banks to miss out on the custom, or to have the reverse situation: cash tied up in ATMs that is not being accessed.
Fortunately, from a practical perspective machine learning can and does help. The current focus around AI is generally on automation of manual processes, with the rising prominence of conversational platforms (AKA chatbots), which use cognitive computing power such as natural language processing. There are, however, an increasing number of “smart machines,” or machines into which AI capabilities have been embedded to enable them to perform additional tasks. These machines include ATMs that can forecast demand more accurately, and assess the need to replenish.It is annoying to find empty ATMs - bring on the smart forecasting cash machines! #AI #AIBanking Click To Tweet
The rise of smart cash machines
These smart ATMs can benefit from the use of a particular type of deep learning method called recurrent neural networks (RNNs). RNNs are specifically designed to handle sequential data, such as speech, text or – importantly in this case – time series. RNNs are called recurrent because they perform the same task for every element of a sequence, like information about withdrawals from cash machines. The output for each element depends on the computations of its preceding elements.
They are very good at forecasting, especially when demand follows observable patterns. They are able to translate previous events into good forecasts of future demand. They do, however, need a lot of data to perform well: More data leads to better performance and increased levels of accuracy.
The demand for cash from a cash machine, for example, will depend on its location, local events, and the time of day or week, to name a few. A machine in or near a student union is likely to be heavily used on a Saturday night, and may be empty on Sunday if it cannot be refilled until Monday. A machine in a shopping street is more likely to be used steadily over the course of the week, with perhaps a peak in demand on Saturday morning. Holiday periods may also affect demand, and will certainly affect replenishment. RNNs use short- and long-term variations in demand to improve the accuracy of their forecasting, and predict the requirements for refilling the machines.
How Will AI Transform Banking? Read this report to understand more about the challenge and opportunity that AI presents.
Smoothing out the peaks and troughs
It may not sound hard to predict that demand for cash will be high on a Saturday night in an area that is full of entertainment venues, or that bank holidays may mean extra cash has to be loaded into cash machines on Friday because the machine will not be refilled on Monday. But flattening out the finer peaks and troughs in demand and supply is more difficult.
Banks want to make their cash work for them, which means getting it into the right place for their customers. Machine learning, and specifically RNNs, offer a way to do that by forecasting demand for cash from particular machines, and ensuring that it is there, ready. This also benefits the banks because it means their customers are happier: They have access to cash, and they are less likely to switch to another provider. In a world where customers are increasingly demanding, this matters.