Thursday, Nov 11, 2021 • 21min

AI: The Future Is Now

Play Episode

Artificial Intelligence used to be the stuff of fantasy movies and comics. Now it’s a part of every industry and everyday life. Our use of it is only going to increase in the future as AI becomes more and more intertwined with our day-to-day activities and industries recognise it's boundless potential. However, so far the finance sector has lagged in terms of uptake compared to other industries. Harish Sundaresh joins to discuss the adoption of AI in the finance sector, drawing on comparisons with the healthcare and transport sectors to argue for its benefits while warning of its pitfalls. Harish is leading the development and management of quantitative trading strategies in the multi-asset space at Loomis Sayles and Director of the firm's Factor-Based Investment Group. He also holds an MSc in applied mathematics (computational engineering) from MIT. Our sources for the show: FT resources, BBC, Reuters, Forbes, IDC, Electronic Design, CAR Magazine. This content is paid for by advertisers and is produced in partnership with the Financial Times' Commercial Department. Hosted on Acast. See acast.com/privacy https://acast.com/privacy for more information.

Talking about

Speakers

(2)

Harish Sundaresh

Tom Parker

Transcript

Verified

Tom Parker

00:06

Hello. I'm

Tom Parker

and welcome to "The Next Five" podcast brought to you by the

Partner Studio. In this series we ask industry experts about how their worlds will change over the next five years and the impact it will have on our day to day.

00:24

This week we're focusing on the future of

and I'll be speaking with Harish Sundaresh, Vice President of

Loomis Sayles

about how

is used in the finance sector, the pros and cons of its use and the outlook for the next five years.

Harish Sundaresh

00:39

One of the data sets, for example, we use is to track crude oil shipments, wherein we can actually get real-time location of these ships and what these ships contain.

00:50

But now imagine that data falling into the wrong hands, such as the pirates at sea, then they will be able to track down these containers in no time because they will be able to find out where the ships are going, where they are currently and what these ships contain.

Tom Parker

01:05

In 1951 the first working

program was written for the

Ferranti Mark 1

computer at

the University Of Manchester

in the UK. The program played a simple game of checkers at an amateur level but was a milestone achievement for its day.

01:24

In 1956, arguably the first automated reasoning

program, called

Logic Theorist

, was demonstrated at the

Carnegie Institute of Technology

. A year later the term "

Artificial Intelligence

" was first coined. Soon computer scientists and science fiction writers predicted widespread uses for

in the future, but it's no longer a game or a movie genre.

01:49

now has real life purpose in every industry and its integration is only set to increase. A study produced by

IBM

in 2020 found that three in every four businesses worldwide are already exploring or implementing

into their functions.

02:06

The global market, including software, hardware and services, is expected to be worth $330 billion by the end of this year and $500 billion by 2024. But one sector that has lagged slightly behind in the uptake of

is the finance sector. Here to offer human insight into

Artificial Intelligence

in the financial world is Harish Sundaresh, vice president at

Loomis Sayles

Harish Sundaresh

02:33

It is true that the adoption of

in finance has lagged other industries by quite a bit. In fact, I believe that you know, we are a few years away from an industry wide adoption of

and I think there are a few reasons for the same. The first and most important reason, I believe, is any

Machine Learning

or deep learning algorithm requires large amounts of data.

02:57

So while data sets in medicine and advertising and etcetera can be quite large. Financial data on the other hand, aside from some high frequency financial data sets, are actually quite small. This makes prediction quite difficult, so it's not for the lack of trying, but unfortunately a lot of people have tried but failed to generate alpha in the markets using

03:20

The second reason I think, is that the problem we are trying to solve is quite different. Financial data is notoriously non-stationary making time series prediction actually quite difficult. Whereas there are other areas such as online advertising, cancer detection, you know,

self driving

cars which is the next big thing, robotic vacuums, these are much more stationary problems.

03:43

And here I quickly want to define what stationary is. It essentially means that, you know, there are no dramatic regime changes over time and if there are, like in the case of financial data, that makes prediction quite difficult.

Tom Parker

03:58

Let's hit the brakes there for a second. Stationary or non-stationary is also known as weak or strong

. Weak

performs a specific task, pre programmed by a human hand. This can be found in the "Siri" function on your phone and other

Chatbots

. Strong

performs multiple functions and teaches itself to solve problems on the fly without human involvement. It's otherwise known as

Machine Learning

04:23

The use of

autonomous vehicles

is a great example of the need for the evolution from weak to strong

. There are six levels on the

autonomous car

scale with no automation at level zero to level five, full automation, where no human interaction is required at any point. Tesla's autonomous cars are the most famous example of a level two.

04:46

In March this year,

Honda

released the first road legal, level three

self driving

production car. Currently only available in

Japan

, it can drive itself without human help in certain conditions, like on a motorway or when stuck in traffic, but that's not that impressive when you look at the claims made in recent years.

05:07

Uber

predicted most of its fleet would be fully

self driving

by 2020. At the moment we seem to be stuck still asking "how much milk do you like in your cup of tea?", i. e. how strong is weak?

Harish Sundaresh

05:20

I think it is somewhat fathomable to imagine a world in the next few years where

self driving

cars are operating along with human driven cars, and that's what makes it difficult, because if you assume that machines will make logical, rational, data-driven decisions every step of the way, they are going to be human beings who are also going to be sharing the road with these machines and humans, as you know, will always introduce noise or disturbances into the system.

05:49

While machines will do exactly what they're told and, you know, do it well, the human element can make it somewhat unpredictable, which is why it has been difficult, somewhat difficult, to go from weak

in autonomous cars where we are now, to a five because it is this external noise or disturbance that has been put into the system, you know, through this human element, and this same concept can be extended into the financial world.

06:17

So how can we use

Machine Learning

to look ahead? So if you think of a crisis as a noise or disturbance to the system, much like how humans are to just the

self driving

cars, the question really is how do you avoid this next crisis? And this is why we make the analogy with cars because it is important to forecast the future steps, right?

06:38

Because how do you react to these disturbances appropriately? As opposed to looking in the rear view mirror, much like what global macro data does. So as a portfolio manager, if I'm able to look ahead and I'm able to avoid crisis, or be able to predict the next one, it will allow me to take corrective action and, you know, I will be having some good years PNO-wise.

Tom Parker

07:04

The pandemic has changed, obviously everything. The way we live, work, travel, spend money. What data is out there to show this and what is the data taught us? Because, is there something that we can learn from that and predict how we're going to live in the future?

Harish Sundaresh

07:22

Yeah, I think the pandemic has changed quite a few things in my opinion and it has made

Machine Learning

quite prevalent and relevant.

, as you know, is now everywhere. It has and it will impact most industries. One of my most favorite applications of

post pandemic is actually in the prediction of community spreading of infections.

07:42

There are firms that are actually sending to your homes free thermometers which will automatically upload temperature readings to the cloud every single time you use it. This data is then analyzed, using

Machine Learning

algorithms, which can make accurate and fast prediction, especially of cluster spreading of infections, in certain areas.

08:03

So imagine if a city or a county got a ten day head start to

Covid

before it actually rises exponentially, this, while it will not stop or prevent the pandemic, it will allow these counties to react much faster as they will really know what to expect. As a result I think this can halt community spreading of such infections, so this I think is quite critical.

Tom Parker

08:27

In November last year,

Google's DeepMind

program,

AlphaFold

, solved one of biology's biggest challenges. It accurately predicted the 3D shape of proteins from their DNA sequence. How a protein works and the function it performs inside cells is based on its 3D structure.

08:46

The technique could help discover hundreds of millions of proteins not yet modeled. Something that scientists are calling a "game changer" and one that will transform medical and drug research forever.

08:58

While still in its developmental stage,

AlphaFold

is already helping to tackle the

COVID-19

pandemic by dramatically reducing the time it takes to map the proteins in the

Coronavirus

. A process that human led experiments would take months to achieve.

Harish Sundaresh

09:15

Yeah, so

is being used in drug development and disease prediction. So the faster the use of

has actually allowed the drugs or will allow drugs to come onto the market much faster. Imagine, in earlier times, researchers would have to read millions of journal articles to make a connection between a drug and a disease, but with techniques such as natural language processing these connections can be made in a matter of minutes.

09:39

So, you're going to be able to read millions of articles and make triangulations or connections between these articles and try and connect drug and disease. I think this is a very exciting application of

as it will transform the speed of drug development, you know, which typically takes ten, twenty years. Hopefully this will be much shorter in time frame.

10:01

Similarly, hospitals in Israel and United States have started to adopt

AI-based

predictive analytics. So the number of patient monitoring devices using data to train

models has actually risen from like about fifty thousand to about 3.1 million in 2021. This includes the use of

for just preventive healthcare solutions like the thermometer idea that I just talked about.

10:24

So, with more devices connected to

and essentially these predictive analytical models, the hospitals will save a significant amount of money but more importantly they will save a significant amount of lives. So overall I think a lot of changes will take place in the near future and it will happen really quickly and we will have to evolve with it, otherwise we will be very quickly left behind.

Tom Parker

10:47

What are the limitations of relying on

and what can still go wrong?

Harish Sundaresh

10:51

All good things will come at a cost. The biggest downside I can think of is data privacy. As you know,

requires a lot of data and a lot of the data is actually personal. So one of the data sets, for example, we use is to track crude oil shipments. Wherein we can actually get real-time location of these ships and what these ships contain.

11:12

But now imagine that data falling into the wrong hands, such as, you know, the pirates at sea, then they will be able to track down these containers in no time because they will be able to find out where the ships are going, where they are currently and what these ships contained and so that makes it a problem.

11:27

The next thing, for example, we have personal credit card data. This is a very important data set for us. Right now the data is an aggregated data set but as we speak, you know, this personal credit card data is also served in secure servers of many corporates, many places you shop online and they're being stolen. Everything about you, unfortunately, will be stored somewhere in the cloud and therefore data privacy and fraud detection solutions will become all the more important.

11:55

So, data or data privacy I think is a big downside to

Machine Learning

, that it requires data and firms are saving all things personal and private about you somewhere in their systems. But, having said that,

Machine Learning

can also be part of the solution. As more and more data is stored, data privacy solutions and fraud detection applications become extremely relevant and important.

Tom Parker

12:16

We've seen the real life applications of

in other industries. Tracking

Coronavirus

, predicting where it's going to crop up in the world, how it's going to help healthcare institutions, how it's going to drive your car. Those are real life applications that affect our day to day.

12:32

How will the better use of

in finance, apart from predicting crises so that you can manage the markets better and make better use of pension funds, how else do you think it will actually affect the investor on the streets day to day?

Harish Sundaresh

12:46

The biggest reason to use

is the information explosion that we have witnessed. 90% of the world's data has actually been created in the past two years. This information explosion has really made it difficult for humans to analyze and keep track of millions of datasets, right.

13:04

In the past I could look at three to five features and try and predict the direction of crude oil, whether it's going up, whether it's going down, but now there are hundreds of data sets. For example, the shipping data set I mentioned that tracks the movement of each and every crude oil container at sea.

13:20

As a portfolio manager now it's not enough for me to just track these three to five features. I need to be looking at these hundreds of data sets for me to make an educated opinion as to whether markets are going up or down. So this concept of dimensionality reduction, wherein we can use

Machine Learning

to analyze hundreds and thousands of features or datasets to come up with a simple "go long or short" answer is actually tremendously useful for portfolio managers.

13:47

Similarly, you know, another application is, we can map parking lots of big chain stores and predict how much traffic is going in and out of the stores. So think Walmart, think Costco. As a result, you can actually predict earnings for these companies.

14:01

All this requires, unfortunately, large scalable infrastructure that can not only store a large amount of information but also require the processing power to analyze the same. So, in finance I'd say the biggest game changers are the shipping data sets, the geographical traffic data sets, the credit card statements, analyzing company earning statements in a matter of seconds.

14:24

Whereas in the past, it used to take hours to analyze a single company post earnings. One dataset we'd like is company hiring information. So if you're able to track the career page of every single company in the

Russell 3000

and look at their hiring patterns to decide whether the company is growing or cutting back. I think that could be phenomenal when it comes to generating Alpha.