• Pau Velando, General Manager at Strands Finance


• 06.10.2016 11:45 am
undisclosed , Pau Velando is the General Manager of Strands Finance. The majority of his professional career was spent as a Big 4 management consultant between Barcelona and Tokyo. An industrial engineer by training, Pau holds a Master degree in Machine Learning and Natural Language Processing, as well as an MBA from IESE Business School. In 2009, he founded a company that develops software solutions to prevent fraud and money laundering. In addition to leading business development and global sales at Strands, Pau teaches Finance at the UAB School of Business and Economics.

Before you read this post, have a look at this video:

We all recognize Super Mario World from Nintendo. But this video is a peculiar one: Mario is not being controlled by a human player. This Mario is a machine playing against another machine: the one originally designed by Nintendo.

The interesting thing here is that Mario - a machine - has learned to play without being explicitlyprogrammed. Think about that for a second: he has actually learned to navigate the game, achieving the optimal score while avoiding threats and ultimately surviving. How can this be? 

MACHINE LEARNING MAGIC

The system in our example has been trained using a trial and error approach on sterioids: think thousands of iterations, where the machine plays the game over and over and over again. Over time, these types of algorithms - called genetic algorithms - learn which strategies produce the highest rewards and gradually eliminate the rest. 

What you have just witnessed is the result of a training process, where Mario was rewarded for every positive act and punished for every negative one - until he dies or reaches the goal. After a very large number of iterations, the system has learned what events offer a reward and which are harmful:

• The system has learned that moving Mario forward and jumping for coins and over holes leads to higher scores. The opposite leads to death.
• The system has also learned that in front of a mushroom it has to jump, otherwise death.
• The system has learned that inside the bricks are coins that can be accumulated in form of wealth, which is a reward and so on.

In this way, a neural network or another algorith was trained to help Mario navigate the course of life. The machine can then make autonomous decisions that lead to a better financial situation of Mario and prevented from risks, harms and death.

Now what on earth does this have to with the world of SMEs and finance?

SUPER MARIO STYLE: NURTURING PARTNERSHIPS BETWEEN SMES AND BANKS

Just like a machine can be trained to help Mario navigate the course of life, a machine can be trained to provide autonomous, precise and relevant financial advisory to a bank’s customers so that they can make better, more informed, decisions about their businesses.

The technology is there, for Mario and for banks. The key question here is the availability of data upon which the system can learn and infer the next best action. Every time you open some of the most popular applications like Amazon, Facebook or Netflix, the most sophisticated algorithms springs into the action of analysing every little bit of data. For example, every second there are around 20,000 people on Facebook with 300 million photo uploads every day, 540 comments every 60 seconds and 293,000 status updates. With this amount of data, deep-learning algorithms can analyse thousands of posts every second with “near human” level of accuracy and, thus giving birth to "near human" intelligent systems.

What about Artificial Intelligence in SME and banking industry? Will it take over the tech and data-driven borrowing and lending decision making? While it's true that banks do not have as much data as Amazon, Netflix or Facebook and AI has not become mainstream, the extent to which fintech companies are changing the processes is undeniable. The only one reason why it still has not become mainstream is the challenge of training the system due to banks having scarsity of data.

While banks are far from being able to deploy AI systems that provide autonomous, accurate and relevant advice to their SME customers (the same for retail customers), there’s a lot that can be done with the use of Machine Learning - ML - solutions being trained in not-so-big datasets.

Generally speaking, the first and most important starting point is a good segmentation strategy. Just like different games need different learning processes, different peer-groups need different training strategies.

Once a good segmentation strategy is in place, then work the classification and prediction algorithms to provide a clear understanding of the past, where is the SME coming from and its revolutionary context. Here’s certainly a lot to be done; banks can deliver a lot of value to their SME customers by training their systems on focused, highly contextualised spaces, for example on predicting delays in payments or estimating the likelihood of accepting the next offer for renewing the fleet of bikes.

3 TYPES OF ML SOLUTIONS

In the space of ML, there is a lot that can be done with the typical dataset that an average bank has. Most common ML solutions can be classified in one of these types:

1. Clustering: given a set of items - e.g. customers, branches, ATMs or products-, the system will propose groupings of “similar” items according to some sets of attributes, features and metrics. These type of algorithms are key in discovering peer groups and clusters since they unfold relationships otherwise invisible for the human expert (i.e. discovering new peer groups by segmenting customers based on spending habits + gross margin seasonalities + turnover profile)
2. Classification: given an unseen item, the system classifies it according to its best estimation (i.e. recommendations on actions to take: prescriptive analytics, recommendations on financial and non-financial products that the SME might need, personalized benchmarking in spending iike utilities)
3. Prediction: given a series of events and items, the system is able to make a prediction (either quantitative or qualitative) with a certain likelihood of happening (i.e. working capital forecasting, forecasting events, anticipating expenses and payments, defaults or overdrafts)

FINAL BOSS: TAKEAWAYS

Why did we see that a machine can successfully guide Mario through his life, avoiding death and earning wealth and value? The answer is that a series of broadly available series of learning algorithms have been properly trained for this particular game, given a rich-enough dataset that was available to train the system.

While the technology is there, it’s proven and widely accepted, the SME space in any bank cannot use AI-based solutions yet due to lack of broad-enough datasets and examples to train the systems. Moreover, the fact that this space is highly fragmented and big enough to act on its individuals only by exception (when some relevant event happens) makes it the perfect place for deploying autonomous, highly contextualised systems that make the interactions smarter and insightful.

Machine learning solutions set the scenario for an smart automation of the interactions between banks and SME´s. Yet the challenge remains on the data side rather than on the tech side. Banks need to learn how to choose the right datasets and the right contexts to train their systems properly. The real challenge (or as gamers call it, the Final Boss) for banks to beat will be to effectively use machine learning to help their SME clients survive, thrive and win their game.

This article originally appeared on blog.strands.com