“I can calculate the movement of the stars, but not the madness of men.” - Isaac Newton In this introductory article, we explore three types of errors inherent in all financial models. We examine these errors using a simple probabilistic model that can be used for predicting the federal funds rate, an interest rate of seminal importance to the U.S. and the world economies. We have collaborated with Google's TensorFlow Probability (TFP) team to develop a fed funds predictive model for this article. It is presented here for illustrative purposes only. TFP is Google's latest, open source, probabilistic machine learning language that can help mitigate the trinity of errors in all financial models.
“One can predict the course of a comet more easily than one can predict the course of Citigroup's stock. The attractiveness, of course, is that you can make more money successfully predicting a stock than you can a comet.” -James Simons. (2008)
Great blog post. Two key points I’d like to comment on: “Equally importantly, the framework needs to update continually the model or its parameters — or both — based on materially new datasets” This is absolutely necessary “Such models will have to be trained using small datasets, since the underlying environment may have changed too quickly to collect a sizable amount of relevant data.” The “training” thing I disagree with 100%. You can’t train the models to conform to abnormal distributions. It can only be a game of handicapping with forward analysis and implementation. I truly believe RenTec has everyone sold on the idea (or mask) of implementing expert ML, AI or whatever sounds hip, but in reality, they’re just expert handicappers and implementors. Like this guy who cleaned up at the track for many years: https://www.bloomberg.com/news/features/2018-05-03/the-gambler-who-cracked-the-horse-racing-code
Clearly, you have not read the article sle. You definitely should. It exposes the dangerous vanity of quants who don't understand the severe limitations of their mathematical models. Firstly, I'm talking about TensorFlow Probability (TFP) and not TensorFlow. Big difference. They couldn't be further apart in their ontology and epistemology. Secondly, I'm a user of TFP and not a vendor. I did register as a vendor when I was selling something to this forum a year or so ago. Wonder what you would have thought if I had just posted this article with another username or had a friend post it. Thirdly, this is an educational article that has received glowing reviews. Here is one from the curator of 'the internet's most useful data science articles': "Excellent post. Short, correct, and somehow still contrarian even after 2008 clearly demonstrated the problems with much of “traditional” modeling done in finance. Finance textbooks still have not been rewritten". See post here: http://roundup.fishtownanalytics.co...cebook-learning-ds-on-a-budget-dsr-154-136138 If sharing an educational article that you are proud of is self promotion, I'm guilty as charged. You're welcome to report my post to Baron and have it deleted.
Thanks for your feedback digitalnomad. I'm happy you liked my article. You wrote: "The “training” thing I disagree with 100%. You can’t train the models to conform to abnormal distributions. It can only be a game of handicapping with forward analysis and implementation." I don't believe we are disagreeing here. TFP has a probabilistic framework and a system using TFP (or any other probabilistic programming language) can be trained to be an expert handicapper. Neither humans nor machines can consistently predict the future with high accuracy. Finance is not physics. However, by consistently working in a probabilistic framework, it's possible to mitigate ones risks. No silver bullet. Just an edge - probably.
I wasn’t aware of this, but still have my doubts. I do find the idea fascinating, and will further research the potential capabilites of TFP. I appreciate your input