Kelly Betting for Dynamic Weighting of Position Sizes

For continuous returns, the optimal weight for Kelly Betting is (predicted return)/(Variance of the returns).

Volatility (or variance) is known to be persistent for stock market data and therefore can be determined using historical volatility with a sliding window.

The predicted return can be obtained using most regression machine learning methods.

This is one way a person could use machine learning’s predicted returns for the weighting of holdings. BTW, I am not suggesting that the weights should be optimal Kelly. You could easily use a more conservative fractional Kelly formula (probably should).

Of course, you would have the information to maximize the Sharpe ratio using Markowitz mean-variance optimization if you are skeptical of Kelly weighting.

Actually, there are a lot of ways that machine learning (not limited to using the predicted returns) could be used for dynamic weighting of holdings. I don't think I have even scratched the surface but having the predicted returns could be a nice start for anyone using the API.

With predicted returns, historical volatility and a correlation matrix--of the stocks you will be holing or just buying-- the potential is endless. All completely doable with the API. There are Python libraries for much of this.


1 Like

Isn't this the Merton Share?