What is Corrective AI?
Ernest Chan, founder, PredictNow.ai Inc.
When it comes to machine learning, the holy grail is Artificial General Intelligence; essentially cloning a silicon replica of a human being capable of making autonomous decisions. Unfortunately, (or fortunately depending on your point of view), this technology is still far off. For instance, despite the billions of R&D dollars spent on building fully autonomous vehicles, we are yet to find many of them driving around our neighborhoods. (A sobering assessment of Tesla Fully Self Driving capability, here). On the other hand, these days we can hardly buy a new car without a fully loaded set of assisted driving technologies.
The same is true in asset management. I have spent decades researching machine learning and applying it to fully autonomous trading systems at various financial institutions including IBM’s Watson Lab, Morgan Stanley’s AI group, and the hedge fund I founded. If this application had been a resounding success, I would be too busy sipping champagne on my superyacht to write about this! The reality is that few funds have found success in this endeavor. But, a few years ago, I learned that there was a much more pragmatic way to deploy machine learning to commercial problems that does not require decades of patience. It is already here.
PredictNow.ai believes that AI is much more effective in correcting erroneous decisions made by humans or other simple algorithms, than in making them from scratch. This is a concept that we call “Corrective AI”.
When we learned of this concept from the financial machine learning expert Dr. Marcos Lopez de Prado, (who described it using a more technical term “metalabelling”) we immediately tested it on our fund’s crisis alpha strategy in late 2019. To our amazement, it made 3 consequential, and most importantly, correct predictions on that strategy. From November 2019 to January 2020, Corrective AI told us that we shouldn’t expect any crisis in the market, and therefore shouldn’t run our crisis alpha strategy. Investors thought we were asleep during those three months and grew concerned. Starting in early February, 2020, the strategy suddenly flashed red and told us to expect a crisis and recommended trading at full leverage. The subsequent COVID-induced financial crisis led to about 80% gross return in the next few months. But Corrective AI was not done yet. In early November 2020, it once again told us that we shouldn’t expect any more crisis. A few days later, on November 9, Pfizer and BioNTech announced their vaccine.
After witnessing the shocking effectiveness and the general applicability of this approach, I decided to launch a new company called Predictnow.ai to commercialize it and take it beyond finance. (I’ve since hired a new CEO to take over management of our fund, which is in its fifth consecutive profitable year partly due to Corrective AI.) After all, asset management is the most challenging use case for AI, though it is projected to be a $2.7 billion market in 10 years. If Corrective AI worked for us in asset management with its ultra-low signal-to-noise and ultra-competitive arbitrage environment, surely it would work even better in other enterprise applications!
Corrective AI Across Industries
You may wonder what domain expertise we have outside of asset management to make experts in those domains listen to us. That was also what people thought when the legendary physicist Richard Feynman was asked to sit on the commission that investigated the Space Shuttle Challenger’s disaster. What did Feynman know that thousands of NASA’s aeronautical experts didn’t? Feynman wasn’t even an aeronautical engineer! But it turned out his unique and simple insight was right: the O-rings that sealed the rockets lost resilience in cold weather. I am no Richard Feynman (though I did graduate from the physics Ph.D. program where he was a professor), but similarly, our team is able to bring a fresh point of view to problems outside of our domain to surprise the experts.
One example of this includes our work with the principals of an oil exploration firm. Prior to our Corrective AI approach, they used a simple formula with just a small handful of variables to predict the productivity of an oil well. After studying their data we discovered that these data actually have multiple dimensions. Instead of using just a few variables, they could be incorporating 400,000 of them! This much larger data set could allow us to potentially make more accurate predictions of oil well productivity. Note that we are not asking the experts to replace their existing formula, which is well-known and well-trusted in the oil industry — we are only correcting it. AI for oil and gas exploration alone will be a $5.7B market in the next 10 years.
Similarly, semiconductor manufacturers often use an expert system, handcrafted by process engineers and including many rules to predict the outcome of a manufacturing process. This system is like the old computer programs that played Go — they played well against amateurs, but not against champions. But then DeepMind’s AlphaGo machine learning program came along and defeated the best human players. Instead of using rules created by experts, AlphaGo learned from past moves to win future games all by itself. In an attempt to replicate this success in manufacturing, we are in discussions with the Head of Production Control and the Head of AI of a major semiconductor manufacturer about how Corrective AI can learn from a much larger set of predictors and make real-time adjustments to their manufacturing process. More generally, the AI market for manufacturing that includes semiconductors, automobiles, consumer goods, and aerospace and defense exceeds $1.6 billion today, with much room for growth.
In practically every enterprise, there will be opportunities for Corrective AI to improve their process. Our goal is to apply our patent-pending optimization technique delivered via a common SaaS platform, but with bespoke input features, to all of them.
Deep Tech, Early Success
Naturally, because of our domain expertise in finance, we have created our earliest product and found the earliest success in the asset management industry. We launched our SaaS platform for asset managers to correct their investment decisions, and have 30 paying customers using it to date. These subscribers range from professional traders to institutional fund managers, spanning both traditional and crypto asset classes. Some of them even signed profit sharing agreements with us.
New users to our platform are often surprised to find that we don’t use the much-hyped deep learning technique that, for example, AlphaGo used. Researchers have found that deep learning, while enormously successful in image and speech recognition and other natural language processing tasks, is not well-suited to most commercial data sets (including most financial data sets). Such data are often “heterogeneous”. (See for example, “Tabular Data: Deep Learning is Not All You Need” by Shwartz-Ziv and Armon, 2021.) Furthermore, deep learning tends to require a much larger data set than is available in commercial use cases, and often requires a much longer time to train than the Corrective AI technique we use. Another advantage of our technique is “explainability” — we have published multiple papers on how we can offer better intuition to the human operators of what input features are really important to the predictions. (See “The Best Way to Select Features?” by Man and Chan, 2020, and “Cluster-based Feature Selection” by Man and Chan, 2021.) Ultimately, we have gone beyond merely “correcting decisions’’ to “improving decisions’’ — i.e. optimizing the parameters of a commercial process via our patent-pending technique called Conditional Parameter Optimization.
In the process of serving asset managers, we learned another key success factor in commercializing AI. No matter how good your algorithm is, it is useless to the customer unless accompanied with pre-engineered input features, customized to each application domain. For asset managers, we created over 600 such features in traditional markets, and over 800 such features in crypto markets. It requires our deep domain expertise to create these features, but many of our customers find them to be our most compelling value proposition. The same phenomenon is happening in our proof-of-concept with our oil exploration partner. Recall that they thought they had just a handful of input features to predict oil well productivity, but one of our data scientists with geological engineering expertise showed them how we can create hundreds of thousands of them.
All of our research and product development successes were accomplished by a small team of approximately 7 engineers. We are still raising a pre-seed round of $1M, though 85% is already subscribed. Despite our early stage, we have no difficulty attracting amazing talent: among them a math Ph.D. from MIT and an optimization Ph.D. from UC Berkeley. To deploy Corrective AI in each new industry vertical, we will need to hire an engineer with specific domain expertise to create pre-engineered features. It sounds difficult especially amidst the reported shortage of AI talent, even major enterprises have difficulty competing with Big Tech for such talent. But because of the high visibility of our published research and educational materials (including highly acclaimed books released by a major publisher), engineers and data scientists from various verticals often show up (virtually) at our door. We are also able to attract tech and asset management leaders to become our early investors, such as the co-founders of DV Trading and the founder of Solace (you would not have heard of Solace unless you are the CEO of a major bank or a senior executive in the semiconductor industry). These investors haven’t just made passive investments in us, they continue to introduce C-level contacts and refer former board members and senior executives from their own organizations to assist us. Even the former CTO of Thomson Reuters has joined our board.
PredictNow.ai already uses machine learning to correct investment decisions for multiple users everyday. Replicating that success in another vertical requires a little imagination but no leap of faith.