In my experience, the biggest problems lie in the fact that markets change quite rapidly; you can't make a model based on 2 years of past data, as the market has changed by then. I do believe more dynamic approaches can yield returns, based on good features and rolling predictions looking back 30 days, or the LSTM machine learning type of models (but you unfortunately need a really good computer for this). The most important thing is to actually understand what you are doing.
From signal/noise ratio perspective, financial timeseries is very noisy, with low signal/noise ratio. So you'll likely fit your model to noise. ML can work, but OOS validation must be very rigorous.
Machine Learning and Algorithmic trading, in my opinion, will lead to a period of very mechanical like trades. Where the prediction improves as more institutions and "Algorithm for Hire" services come online. I am not sure my thoughts on the outcome of this, but a quick thought leads me to believe it will become a brief era of lower margins and higher volumes.
I think machine learning can make strategies based on short-term anomalies disappear or drive their margins even lower. But strategies based on the exchange of value are likely to remain profitable, IMO.
In my experience, the biggest problems lie in the fact that markets change quite rapidly; you can't make a model based on 2 years of past data, as the market has changed by then. I do believe more dynamic approaches can yield returns, based on good features and rolling predictions looking back 30 days, or the LSTM machine learning type of models (but you unfortunately need a really good computer for this). The most important thing is to actually understand what you are doing.
From signal/noise ratio perspective, financial timeseries is very noisy, with low signal/noise ratio. So you'll likely fit your model to noise. ML can work, but OOS validation must be very rigorous.
Machine Learning and Algorithmic trading, in my opinion, will lead to a period of very mechanical like trades. Where the prediction improves as more institutions and "Algorithm for Hire" services come online. I am not sure my thoughts on the outcome of this, but a quick thought leads me to believe it will become a brief era of lower margins and higher volumes.
I think machine learning can make strategies based on short-term anomalies disappear or drive their margins even lower. But strategies based on the exchange of value are likely to remain profitable, IMO.
I like this, and hope that the value trades and traditional value investing wins out long term.