Tuesday, January 28, 2014

Natural Language Queries Are Different In Finance

Kensho's "Warren" wants to be a Siri-like AI for the finance sector.  It's cool that they're reaching for innovation.  I think finance can benefit from AIs provided they are deployed in relevant ways, and as long as their human operators know their limitations.

The sample question about energy shares, oil, and Middle East instability strikes me as a no-brainer.  Energy stocks benefit from oil price shocks of any kind.  Deep data mining of the history of energy prices and share price movements is within the reach of existing data subscription services, so I don't see how this "Warren" AI would add value.  I don't buy the argument that highly specialized quants need hours of coding to find a satisfactory solution.  That's a crutch hedge funds use to justify their enormous fees.  A good analyst can use data to make an intuitive leap from experience.  I do that all the time.  The quality of that experience matters because it processes judgments about human nature.  Natural language queries in finance must incorporate real-time effects of other humans acting on events.

The qualitative differences between the financial body of knowledge and the medical body of knowledge are extremely relevant to the utility of financial AIs.  IBM's Watson can give consistent answers to medical queries because it can reference a body of knowledge validated by centuries of scientific experimentation.  The available data points don't change after interacting with observers.  Financial data behaves differently.  Markets change because human decisions move them, and published interpretations of those changes move them further.  This complex feedback system is similar to Heisenberg's uncertainty principle, and George Soros' The Alchemy of Finance discusses it in detail.  My point is that a finance AI confined to interpreting historical data misses the real-time changes wrought by human interaction.

Mastering the effects of uncertainty, observer interactions, and black swans implies finance AIs need a different computational approach.  Quantum processing may be the correct solution.  If human observation of random events has quantum effects, a quantum computer driving an AI can deliver a more robust interpretation.  This is all pretty far-out stuff.  I look forward to seeing what the next generation of finance AIs can deliver.