Hacker News new | past | comments | ask | show | jobs | submit login

I've spent a huge amount of time playing with similar stuff. The goal is to predict anything that you can make money on -- this can be obviously be price, but you can also try volatility, or predicting a probability distribution. Try relaxing constraints to find your signal: price ranges instead of exact prices, date ranges instead of exact dates. If you can output a probability, distribution, or confidence interval as well, that information can be used for position sizing (see Kelly criterion.)

If you start expanding beyond a small set of specific companies, don't forget to find data for delisted companies to avoid survivorship bias.




I looked at Kelly criterion, and what's interesting is that I was just thinking yesterday about whether I should optimize

E(log(wealth)) - c * std(log(wealth))

or

E(wealth) - c * std(wealth)

The first one is better for building up wealth over a long time, but when I looked at Modern Portfolio Theory, it is suggesting to use the second formula.

I'm also planning to spend about 5% of my wealth every year, and the second number gives me lower risk in the short time frame, so I think something in the middle would be a better criterion.


Kelly sizing is optimal long term, but is highly volatile. A common practice is to size by a fraction of Kelly, to reduce variance and since risk of ruin increases fast after 1.0 * Kelly. It also allows room for error in the probability predictions, to avoid going over 1.0 * Kelly by accident.

I try not to worry about stdev and focus on the soundness of my process, but in practice I have money split into higher-risk where I try to be clever, and low-risk failsafe investments to raise the floor of the worst case scenario.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: