IBM created a series of webinars called, "AI for Professionals" in this particular video, Prashant Bhuyan, Founder and CEO, takes us behind the scenes at AccreteAI. Through an innovative process using artificial intelligence and collective human intelligence, Accrete.AI helps finance professionals overcome information overload, remove bias and make better investment decisions. Get an insight into how we leverages IBM's technology in addition to proprietary expert networks, shallow learning and deep learning.
Prashant and his team demonstrate how two domain specific ai powered cognitive investment tools from Accrete uncover insights within real-time news stories, tweets, and social media to help fund managers. Decision-makers can identify market-moving mergers, spot acquisition rumors, and track changes in sentiment by key topics mentioned in earnings calls, with greater accuracy.
See the full webinar Alpha with AI
Hi, my name is Prashant Bhuyan, (CEO and Founder, Accrete.AI). I spent over a decade in the high frequency trading (HFT) space and at some point in time I recognized that the nature of volatility is changing in a much more idiosyncratic nature. And I attributed this to the explosion in unstructured data texts, tweets, video's, images and blogs. For example, an errant tweet could send markets reeling for no rhyme of reason. Investors were spending way too much time and cognitive bandwidth endlessly scouring the web looking to reconcile contradictory pieces of information, looking for signals buried under growing mountains of noise. In this pain I saw an opportunity, shifting focus away from HFT to High Speed Cognition and we started Accrete to build smart cognitive investment tools designed to help investors overcome biases, make better decisions, and generate alpha in the digital age.
We started off by training supervised learning systems using IBM that would help us try to process and contextualize this information. The first problem with supervised learning in general is that they are prone to bias, and they fail in terms of accuracy when the context of information changes. We needed a solution that would help us process huge streams of real-time unstructured information. We started to experiment with a purely unsupervised learning approaches and found that there was very rarely enough training data that could help us build systems that would be useful in solving actual real world problems. We came up with a hybrid approach, partnering with IBM, building upon their large computational infrastructure we were able to create an approach called MACI, Machine Augmented Collective Intelligence. This approach is an iterative, collaborative approach between human experts that don't know each other, we de-bias the human expertise, together with shallow and deep learning engines that allow us to create a "brain". This "brain" is continuously accreting knowledge, getting smarter with every passing moment, and powering a factory in which we are building domain specific bias-free and contextually adaptive investment tools designed to help investors combat information overload and generate alpha.
Why is Accrete more equipped to solve the problem of information overload?
Yurdaer, CTO of Accrete.AI explain that Accrete is building a hybrid intelligence engine which brings together the best aspects of human intelligence and machine intelligence. Any system that produces an output is a representation of the truth that suffers from bias, in particular this representation suffers from two components of bias; the first component of bias comes from machines and the reason from that is because you don't have context models complex enough to represent the truth in machines and for that you either need a lot of samples which captures the truth for the machine or we need another intelligence which would tell and tune the machine for the truth. In order to be able to eliminate the bias that was introduced by the machine we introduce human intelligence. But human intelligence is also bias, and this is the second component of bias in the system. In order to eliminate the human bias we use machines, so there really a constant feedback between machine and human intelligence to improve the models so we have a better representation of the truth. We do this is by bringing together subject matter experts, and capture the domain knowledge from these subject experts by using IBM. With IBM's assistance we ask subject matter experts independently identify the main specific entities and relations. We generate a ground truth by looking at the common annotation agreed by all subject matter experts. This creates an annotation engine which may have some bias introduced by the subject matter experts and is deployed to a natural language understanding (NLU) unit. The NLU does not have domain specific knowledge, so it has its own bias, and in order to eliminate that bias what we do is override the representation of the NLU with the domain specific annotations we received from the subject matter experts. This is one way of eliminating the bias in the machine intelligence. This iteration goes on and every time we iterate, add new concepts, correct the representations of the old concepts the machine will be trained and learn better. Every time the subject matter experts look at the truth produced by the machine they will learn their own short comings and biases and improve their own systems using IBM.
Would you say it is very similar to a child riding a bicycle with training wheels, and with every iteration the training wheels become less important?
Exactly, it's rationale is to not only helping the subject matter experts to learn how to "ride this bicycle" in this case make annotations using IBM but at the same time, by learning how to ride a bicycle better the subject matter experts are helping the engine move faster and if it has any value in making a decision in the financial domain. How about a real world example, "Tesla is going to the moon" versus "Tesla is on fire" a stock trader may interpret "Tesla is on fire" as a stock that is rising rapidly in price. How is the machine suppose to know whether "Tesla is going to the moon" means the same thing as "Tesla is on fire" or does it actually mean Tesla is building a rocket ship to go to the moon? This is the fundamental question! The context is not only built by only the words that are visible in the sentence, but the whole document; the keywords, entities and relations that are extracted from the document help to build the context of that particular sentence. So we are not only looking into the context of a sentence but within the context of the whole document we are trying to understand what that particular sentence means.
Let us show you how Accrete is applying these technologies in the real world to help investment managers make better investment decisions in today's financial markets.
Andrés, Head of Product at Accrete.AI starts off by stating that, we have taken this notion that creativity and independent thinking which can otherwise be referred to as reasoning from first principles or the scientific method really lies at the root of all alpha generation. We have built these products in such a way that we deliver exactly that concept at scale for our customers. Rumor Hound, which is probably one of our most exciting products, can be thought of as your most prolific M&A analysts digitized, uploaded to the web and let loose scouring for every rumor, document, soundbite, whisper, or anything related to M&A. Putting all that information in one place and understanding what's real, what's not, what's credible and what's actionable. The problem is rumors can emanate from any source, there are thousands of sources out there on the web and elsewhere and we basically ingest all of that, categorize it by recency and credibility and then deliver it to our investors in an easy to use dashboard that they can make sense of.
This is what our dashboard looks like and just recently we had a really interesting development less than 24 hours ago with Lionsgate ($LGF.A). Our system picked it up before some of the major wire houses and we can actually see what the rumor looks like right here that our system picked up, this can be text from within a much larger document that the system extracts or just something as short as a tweet. Correspondingly, we saw very tradable action in the price action of Lionsgate. This rumor was picked up overnight, the stock gapped up and continued going upward for the remainder of that day and even into today as well. We've seen this now over the last week roughly where the system has already picked up a ton of actionable rumors and we expect it to continue to do so. Our customers are definitely very excited.
Accrete.AI also has another product called Topic Deltas, and this product basically gives you the ability to listen to all earnings calls both current and historical and understand everything that's been mentioned in there by category, every single topic, remember every nuance mentioned in the language and compare this across time and against market expectations leading up to that call. Really what this product is doing is very similar to Rumor Hound by actually listening to and transcribing all this disparate information and remember the nuances in human language that can be difficult to pick up sometimes and this is challenging enough on a stock-by-stock basis let alone the entire market. We simplify this process so that our clients can do a lot more with the same analyst team that they're already using. What we see on the screen right now is the output of a completed earning call would look like, so this is AMD ($AMD) back in Q4 2016 and each concept is categorized over here on the left, the individual snippets that were listened to and transcribed as it relates to that concept are in the middle columns, and the right most columns are scored by sentiment. The left column is the previous earnings call and the right column is the current earnings call, allowing users to easily compare that sentiment from call-to-call and then this is done graphically down at the bottom as well. For this particular earnings call we actually saw some pretty pronounced price movement both at the Q4 earnings call which happened here on the chart. The overall consensus was pretty positive from call-to-call leading up to that and correspondingly a large price move. Interestingly, in Q1 2017 we had the exact opposite effect where the topic sentiments went quite considerably down over that period of time and the stock price also reacted similarly. This gives you a little flavor of what Accrete.AI is up to, Rumor Hound and Topic Deltas are just two of multiple products that we already have commercially available and we're working on a bunch of other stuff that's really cool.
As you can see our products are helping solve real world problems, we're helping analysts, portfolio managers, risk managers, and hedge funds spend less time searching for information and focus all their cognitive bandwidth on making better decisions, overcome bias and ultimately generate alpha in this digital era of markets. We think that as these systems become more complex and as this explosion in unstructured information continues to grow it's going to be harder for human beings to actually trust a 'black box'. One of our driving principles at Accrete is to transcend the black box and develop a trusting relationship between our products outputs and our customers. We do this by exposing pieces of the reasoning behind the outputs with an eye towards the future, and in the future I think humans and machines are going to work more collaboratively and it's going to be a symbiosis between human intelligence and machine intelligence. Accrete is thrilled about our partnership with IBM and really excited about embarking on this journey together because we both have the same point of view that machines and humans will be working together and not against each other. Accrete is really excited about the fact that we will be launching all our products on IBM's cloud for financial services.