The Forbes FinTech50 is prestigious list. Assembled by the editorial staff based on their research, stories and conversations, the list is considered one of the most elite in the space – up there with the RiskTech100 from Chartis and a handful of other carefully curated lists. Needless to say, we were delighted to make the cut this year, joining some groundbreaking companies like Ripple, Betterment, Stripe and others.
Our inclusion is a function of the fact that we have posted real results with demanding customers. Forbes references our work with Citi on CCAR and HSBC on AML, but there is lots of other work that is not public. Work at a major Canadian bank identifying potential customer churn with exceptional accuracy. Our work with a major UK bank on predicting complaints for the purpose of building better products and evaluating risk (see Well Fargo). Our work with another UK bank on asset and liability management. Our work with a major Nordic bank determining the areas of weakness in their mortgage models. Our work with several banks and ratings agencies on predicting changes to credit ratings. The list goes on.
The connective tissue here is that these problems are highly complex, exceptionally valuable and, for lack of a better term, chronic. That is, they remain unsolved despite the payoff for solving them. This is not to suggest that we have unlocked the magic money tree – we haven’t (and wouldn’t write about if we had), what it does suggest is that our approach is different – and is able to produce results where traditional techniques, even traditional machine leaning techniques fail. This is a function of our capability set and our process. We have written about it extensively, but the TLDR version goes like this:
- To be successful in the face of modern financial data, you have to be adept at unsupervised learning. That is, you need to be able to learn from data in the absence of questions or hypotheses. The reason you need this capability is that the size of the data sets (rows and columns) creates so many potential questions that can outnumber the atoms in our galaxy. You need unsupervised learning to provide a principled starting point for inquiry.
- While prediction is the holy grail, the truth is that it is a well understood problem. If you want to improve prediction (supervised learning), there are options, but the best is unsupervised learning. By having an unbiased representation of the data, clean groups, one ends up with better prediction. Not enough people understand this point. This is probably a function of the fact that it is not in the best interest of the “prediction only” crowd to discuss it.
- While supervised and unsupervised learning capabilities are critical, they have little utility in the real world if they are not explainable. By explainable we don’t mean “what algorithm was used and where”, we mean “what did the algorithm actually do” in ways that humans can understand. We call this justification. Without it, AI is destined to solve academic problems. With it, AI can move into production against some of the most valuable use cases out there.
- AI needs a vehicle for action. If the output of our AI systems is a .pdf or PowerPoint document, we haven’t achieved much. AI needs a user interface. If it is an application that is used broadly in the organization and with frequency – then it has profound implications. AI needs those vehicles. We think applications are a key option, but programatic interfaces between systems works too.
- Finally, AI needs to continue to learn. To do that it needs to be positioned in the workflow, seeing data. If an AI application is not getting smarter it is getting stupider. It can only do that if it continues to see data. We see this all the time with our clients and refer to it as “model drift.”
The key is to be able to execute ALL of these framework components. Not one or two or even three – but all five. Our ability to do so is what differentiates us. It is what allows us to work on such compelling and difficult problems. It is what allows to make these lists.
Having said that, it is not just the framework, the underlying technology or the ability to justify the results. There is an amazing group of people powering our success. Domain experts with slick data skills, superb coders with a passion for finance and sellers with the intellectual capacity to translate the complex into in the compelling. While we work hard at this everyday because we love the challenge – the recognition is nice. Really nice. Thanks Forbes!
The Forbes FinTech50 is prestigious list. Assembled by the editorial staff based on their research, stories and conversations, the list is considered one of the most elite in the space – up there with the RiskTech100 from Chartis and a handful of other carefully curated lists. Needless to say, we were delighted to make the cut this year, joining some groundbreaking companies like Ripple, Betterment, Stripe and others.
Our inclusion is a function of the fact that we have posted real results with demanding customers. Forbes references our work with Citi on CCAR and HSBC on AML, but there is lots of other work that is not public. Work at a major Canadian bank identifying potential customer churn with exceptional accuracy. Our work with a major UK bank on predicting complaints for the purpose of building better products and evaluating risk (see Well Fargo). Our work with another UK bank on asset and liability management. Our work with a major Nordic bank determining the areas of weakness in their mortgage models. Our work with several banks and ratings agencies on predicting changes to credit ratings. The list goes on.
The connective tissue here is that these problems are highly complex, exceptionally valuable and, for lack of a better term, chronic. That is, they remain unsolved despite the payoff for solving them. This is not to suggest that we have unlocked the magic money tree – we haven’t (and wouldn’t write about if we had), what it does suggest is that our approach is different – and is able to produce results where traditional techniques, even traditional machine leaning techniques fail. This is a function of our capability set and our process. We have written about it extensively, but the TLDR version goes like this:
- To be successful in the face of modern financial data, you have to be adept at unsupervised learning. That is, you need to be able to learn from data in the absence of questions or hypotheses. The reason you need this capability is that the size of the data sets (rows and columns) creates so many potential questions that can outnumber the atoms in our galaxy. You need unsupervised learning to provide a principled starting point for inquiry.
- While prediction is the holy grail, the truth is that it is a well understood problem. If you want to improve prediction (supervised learning), there are options, but the best is unsupervised learning. By having an unbiased representation of the data, clean groups, one ends up with better prediction. Not enough people understand this point. This is probably a function of the fact that it is not in the best interest of the “prediction only” crowd to discuss it.
- While supervised and unsupervised learning capabilities are critical, they have little utility in the real world if they are not explainable. By explainable we don’t mean “what algorithm was used and where”, we mean “what did the algorithm actually do” in ways that humans can understand. We call this justification. Without it, AI is destined to solve academic problems. With it, AI can move into production against some of the most valuable use cases out there.
- AI needs a vehicle for action. If the output of our AI systems is a .pdf or PowerPoint document, we haven’t achieved much. AI needs a user interface. If it is an application that is used broadly in the organization and with frequency – then it has profound implications. AI needs those vehicles. We think applications are a key option, but programatic interfaces between systems works too.
- Finally, AI needs to continue to learn. To do that it needs to be positioned in the workflow, seeing data. If an AI application is not getting smarter it is getting stupider. It can only do that if it continues to see data. We see this all the time with our clients and refer to it as “model drift.”
The key is to be able to execute ALL of these framework components. Not one or two or even three – but all five. Our ability to do so is what differentiates us. It is what allows us to work on such compelling and difficult problems. It is what allows to make these lists.
Having said that, it is not just the framework, the underlying technology or the ability to justify the results. There is an amazing group of people powering our success. Domain experts with slick data skills, superb coders with a passion for finance and sellers with the intellectual capacity to translate the complex into in the compelling. While we work hard at this everyday because we love the challenge – the recognition is nice. Really nice. Thanks Forbes!
