Ernest Chan (Predictnow.ai) - "How to Use Machine Learning for Optimization"
Вставка
- Опубліковано 28 лют 2023
- Abstract: Conditional Portfolio Optimization is a portfolio optimization technique that adapts to market regimes via machine learning. Traditional portfolio optimization methods take summary statistics of historical constituent returns as input and produce a portfolio that was optimal in the past, but may not be optimal going forward. Machine learning can condition the optimization on a large number of market features and propose a portfolio that is currently optimal. We call this Conditional Portfolio Optimization (CPO). Applications on portfolios in vastly different markets suggest that CPO can outperform traditional optimization methods under varying market regimes.
Speaker Bio: Ernest Chan (Ernie) is the founder and CEO of Predictnow.ai, a machine learning SaaS. He started his career as a machine learning researcher at IBM's T.J. Watson Research Center's Human Language Technologies group, which produced some of the best-known quant fund managers. He later joined Morgan Stanley's Data Mining and Artificial Intelligence group. He is the founder and non-executive chairman of QTS Capital Management, a quantitative CPO/CTA. He received his Ph.D. in physics from Cornell University and his B.Sc. in physics from the University of Toronto.
After finishing the video, the thing that immediately come to my mind is to apply this method on the parameter selection of a single strategy. Like doctor said in the beginning, walk forward rolling window method doesn't consider the current market infomation into the decision, but only the past performance of the strategy/portfolio itself, which might also be suffered from a lot of noise overfit issue or the plateaus area is hard to identify......etc
I wonder if I change to use this kind of machine learning way to dynamically change the parameter while backtesting, maybe those strategies that I already throwed away could revive.
Gread content!
man he did leave a lot to the imagination but this idea is so insane with AI actually having a neaural'ish network. I wonder if their approach is bruteforcing different set of conditions across asset classes till they find out the conditions that actually affect the markets and in what ratio, and then improve the current model. Coz thats fckin epic
So ml model takes in
... market regime features + trading strategy parameter(if any)+ allocation weights
... spits out sharpe ratio
i could easily see this overfit and not having predictive power.
If this worked, we would optimize parameter for single strategy. Rebalance parameter each month. But this is no different than walk forward optimization...
Also... regimes can change before the weights are rebalanced.
w
silllllllllly
Hghh
that's a terrible presentation. He did not give any definition of regime, nor did he describe a method how to measure regimes. A lot of talking though
I'm not very familiar with Quant Finance, but I thought the implicitly defined regimes makes sense. Sorta like HMM but with kinda very large number of hidden states
I initially felt the same way, however I believe the key area to pay attention to is 20:37 where he talks about the features used. If you dig a little deeper into time-series features this explains how the regimes are modeled IMO.