Efma_London_2016-Strands-008.jpg

FinTech Pulse

← Blog Home

Welcome to the Machine Learning Era of Banking

ON THE AI FUTURE OF FINANCIAL SERVICES

At Strands we're stacked full of FinTech experts of all stripes and colours, but did you know we have two in-house gurus on artificial intelligence and machine learning of our very own?

In the lead-up to the D-RAFT AI & Robotics Demo Day, we sat down with each of them to pick their brains about the future of AI and ML with a special banking twist. Thanks to the folks at Prowly for developing these incisive questions. 

Before we get started with the interview, some introductions are in order...

marc_circle.jpeg

Strands Co-Founder Dr. Marc Torrens is also our Chief Innovation Officer. His passion lies in overcoming new business challenges with state-of-the-art tech in the areas of Artificial Intelligence, Data Visualization, Usability and Product Design. 

Marc holds a Ph.D. in Artificial Intelligence from the Ecole polytechnique fédérale de Lausanne and an MSc in Computer Science from UPC. He is also a faculty member of the Business Analytics and Big Data Master Programme at the IE Business School in Madrid.

 

pau_circle.jpeg

 

 Pau Velando is the General Manager of Strands where he leads and coordinates business efforts worldwide. He is passionate about machine learning, artificial intelligence and cognitive computing.

Pau holds a degree in Industrial Engineering and a master in Machine Learning and natural language processing. He also holds an MBA from IESE Business School in Barcelona. 

 

THE INTERVIEW

How is AI different today from, say, 20 years ago? 

MARC:  From an academic point of view, there is nothing new or revolutionary that makes AI algorithms better nowadays than 20 years ago. The AI revolution we see today comes from a higher quality and quantity of data that is available in many industries. The newly available data has driven the development of amazing “Big Data" platforms to process and compute large sets of heterogeneous data more efficiently than ever before.
 
So the big difference is that now we can actually see the classic algorithms of 20 years ago in action, for practical use in industry. The algorithms themselves are basically the same, it's their applicability that has dramatically changed with exciting results.
 
PAU: As Marc said, the main difference is the quality and availability of data. Unlike 20 years ago, now we have access to huge volumes of data so we can train the systems properly -  that means show the machines millions of examples from which they can learn. 20 years ago, machines couldn't be trained properly due to data scarcity.
 
Is Big Data Still a Thing? 
 
PAU : Of course - it's the foundation of having accurate AI systems. After all, you can't have AI without Big Data! Both are definitely here to stay, not mere buzzwords. But like all buzzwords, "Big Data" IS misused and thrown around incorrectly. Big Data can be defined by the 3 V's: Variety, Volume, and Velocity (capability of using it quickly, readiness) - so if it has all 3, it's the real deal.
 
MARC: Well, I think Big Data has never actually been a formal computer science concept, but more a new way to efficiently compute with large and heterogeneous data sets. As such, the Big Data approach is here to stay in the sense that is a practice that will gain increasingly wider adoption. Internet of Things (IoT) will make Big Data essential for almost every industry, and not just a practice for visionaries. 
 
D-RAFT-AI-Robotics-Strands-4.jpg
 
What might the future hold for Big Data Intelligence, the post-convergence of big data and AI? 
 
PAU: Truly expert systems. Today we call systems that compile human-made rules "expert", but if a human doesn't know the rules, the system isn't expert enough. Now imagine a virtually infinite volume of data, with which we can train a system to become a real expert that doesn't need a human typing the rules. The machine starts creating the rule! This is the future of AI: self-learning systems.  
 
MARC: Big Data Intelligence is about analyzing data in platforms to extract new knowledge that can be used for various purposes such as taking better company decisions, allocating resources more efficiently, understanding customers better, market research, fraud detection, or even designing new products and services. This can be achieved with different computational approaches from simple statistics to sophisticated AI techniques, which is where the future seems to be heading. 
 
What might these technologies – AI, Big Data & machine learning – do?
 
PAU: Two very exciting things! The first is discovery, or the uncovering of patterns unobserveable to humans. From proteins in illnesses to global financial markets... things we cannot observe or conceive. Machines will do it faster and more accurately. And it's not a question of if, but when.
 
The second is the simulation of human reasoning: not that machines will substitute, but rather mimic human reasoning. Especially in natural language - the system will actually understand the context on an unprecedented level. Think about Siri. Isn't amazing that you can have a dialogue with Siri in any language?
 
MARC: It's important to approach Big Data less as a technology and more as a revolutionary paradigm that allows us to host and compute large heterogeneous data sets. Then, the question is how to extract new knowledge and insights from these "data lakes"? Advanced and descriptive statistic techniques can be used as a first step. Then, Machine Learning algorithms can be powerful in predicting certain behaviours based on past events.
 
There are multiple applications in banking, such as fraud detection, customer clustering, attrition rate prediction, likelihood to acquire a new product, and so on. 
 
What’s your strategy for acquiring enough data to test a machine-learning product?
 
PAU:  This is a tough one because you could either simulate it - if you know the pattern that represents a peer group, to increase the volume, you could get "statistically similar" data to simulate it, just for testing purposes. The other thing - in our business, you could integrate from external sources, eg. external account aggregation: bring data from other sources in a strategical way.  
 
MARC:  It's true that there isn’t a simple answer to this question, it depends a lot on the domain and specific problem you want to solve. Ideally, you should get three sets of data: train set, cross set and test set. The train set is used to build the different models, the cross set is used to select the best model, and the test set is used to generalize and evaluate the prediction error of the selected model.
 
Determining the size of these sets should be done by iterating the process until the final prediction error is minimized. Of course, there is also a trade-off between error measurement and the size of the sets. The important thing is to use a methodological approach and avoid any type of bias when creating your data sets.
 
finovate_europe_2016_-_pau_stand.jpg
 
With machine learning, how do I know it works?
 
PAU: You should receive meaningful results or outputs. Meaningful means accurate, precise, predictable and repeatable... like when you repeat a scientific experiment. Machine learning is a scientific method. Like if your PFM predicts your account will go into overdraft in 3 days and it happens. Or if you get an offer to fly to Asia and you accept it. The likelihood that you would accept that offer represents the quality of the predictor built into that system.  
 
MARC: In general, you know ML works when testing your model on the test set of data, as a last step in an iterative process. This final evaluation gives you the error of the prediction with respect to the data in the test set. You need to reach a margin of error that is acceptable for the specific domain and problem you are solving. For example, in some domains, the most important evaluation could be to minimize the false positives of the prediction while in other problems the avarage error of the prediction could be the most important evaluation. In general, you know it works when the resulting error is acceptable for the specific domain and problem you are solving.
 
What about machine learning do you feel people just don’t understand?
 
PAU: People mistakenly see machine learning as a sort of magical black box. In goes some data and out comes a result, and nobody knows what happens in between. In reality there is a great deal of trial and error! Few people truly understand the amount of fine-tuning an algorithm takes - the magic doesn't happen right away. Fine-tuning depends on each use case, so you can't even generalize how long it might take to perfect an algorithm.
 
MARC: In my opinion, the concept as a generic approach to learn from data is pretty well understood by the industry. However,  too many efforts and resources often go directly building the right algorithmic model and not enough to looking at the data itself.
 
This is a mistake, since one of the most important aspects of any ML project is analyzing the data: understanding the features you need to use, the dependencies between them, cleaning up redundancies, removing outliers, and so on. It's like if you want to prepare a good meal, you had better focus more on the quality of the ingredients rather than the recipe itself! In other words, never forgot that famous saying in computer science: trash in, trash out!
 
Lots of entrepreneurs are scared of launching imperfect products. What would you say to that?
 
MARC: This is the most common mistake made by entrepreneurs across all disciplines. The lean approach has to be radically applied. Entrepreneurs should launch their products as soon as possible, and iterate very often to improve them. There are many examples where very successful products have been discovered almost by chance while testing other concepts. Entrepreneurs should be humble and very open to new ideas while developing the initial idea. Pivoting is a necessity. Most of the times, entrepreneurs are overly perfectionist, and whenever the product is finally ready in their eyes, it's already far too late for the market. I remember the days when even Google was still in beta version even after being in the NASDAQ! Innovation is always in beta, and never finishes!
 
PAU: In the ML space, it's very important to focus on niche apps and solutions. If somebody aims at a very broad application and a very broad set of solutions, it's probably going to fail. If entrepreneurs focus on solving a particular problem with a particular algorithm, they are more likely to succeed. Why? A narrowed variability of results. There are 12-15 algorithms available to use. Very difficult to chose - for solving what? Prediction or classification? It's important for entrepreneurs to choose a niche and chose an approach, and excel at that. You lose a lot of time by launching a million products imperfect. You can also get algorithms to compete - imagine 5 can be used for 1 problem. Put them to compete against each other, and see which produces the most accurate solution. Once you have decided on your algorithm, THEN you must perfect it before launching.
 
Discover how Strands technology uses AI

Topics: Machine Learning, Artificial Intelligence

Author: Nicole Harper, Content Manager on Sep 23, 2016

Find me on:

Subscribe to Email Updates

Posts by Topic

see all