4 ms·

The individual home flipper working out of their own money is going to only choose projects that look very favorable. The corporation is going to hire a manager

by gvb 4y ago

The individual home flipper working out of their own money is going to only choose projects that look very favorable. The corporation is going to hire a manager and that manager will probably be told to close a certain amount of deal flow and will feel pressured to close deals that aren't very profitable and not have such a financial incentive to keep the deals profitable. The corporations trained an AI on data that showed a strong upward trend in selling prices. The result is the AI predicted that most flips would be profitable and thus almost always said "buy." This worked until the housing market slowed down, the upward trend in selling prices slowed (and reversed in some cases), and the AI predictions of profitability became wrong.

PaulHoule 4y ago
There are a few reasons why A.I. projects commonly fail: ① Insufficient or poor quality training data (e.g. better an old algorithm on good data than a cutting edge algorithm on poor data) ② No calibration. Calibration seems to be the best kept secret in ML (people know need good training data but they are lazy or not brave enough to insist on it... I've made that mistake, but https://scikit-learn.org/stable/modules/calibration.html https://scikit-learn.org/stable/modules/calibration.html seems to be truly obscure) In the case of a trading strategy or other commercial action you would be calibrating on expected return or possibly something that balances risk and return like Sharpe ratio. ③ Non-stationarity. Distributions are changing all the time on their own, but in markets they get changed by your own actions and those of people following the same strategy of you. This book has a great study of a hedge fund strategy that burned out the way many algorithmic strategies do https://www.amazon.com/Hedge-Funds-Perspective-Financial-Engineering/dp/0691145989 https://www.amazon.com/Hedge-Funds-Perspective-Financial-Eng... I think it's key that the burnout happens because of greed. A given strategy can absorb a certain amount of capital. When more capital gets attracted to the strategy the returns inevitably go down, market participants can try to make up for this by increasing the leverage but you can see how that goes... So I would still blame too much greed and too much capital because it inevitable that markets fluctuate.

by gvb 4y ago

PaulHoule 4y ago

There are a few reasons why A.I. projects commonly fail: ① Insufficient or poor quality training data (e.g. better an old algorithm on good data than a cutting edge algorithm on poor data) ② No calibration. Calibration seems to be the best kept secret in ML (people know need good training data but they are lazy or not brave enough to insist on it... I've made that mistake, but https://scikit-learn.org/stable/modules/calibration.html https://scikit-learn.org/stable/modules/calibration.html seems to be truly obscure) In the case of a trading strategy or other commercial action you would be calibrating on expected return or possibly something that balances risk and return like Sharpe ratio. ③ Non-stationarity. Distributions are changing all the time on their own, but in markets they get changed by your own actions and those of people following the same strategy of you. This book has a great study of a hedge fund strategy that burned out the way many algorithmic strategies do https://www.amazon.com/Hedge-Funds-Perspective-Financial-Engineering/dp/0691145989 https://www.amazon.com/Hedge-Funds-Perspective-Financial-Eng... I think it's key that the burnout happens because of greed. A given strategy can absorb a certain amount of capital. When more capital gets attracted to the strategy the returns inevitably go down, market participants can try to make up for this by increasing the leverage but you can see how that goes... So I would still blame too much greed and too much capital because it inevitable that markets fluctuate.