Just from a historical perspective, I can add a few comments. This might help you understand why kelly isn't generally used to size bets automatically with JBT.
1) I think Eugene (nonlinear) only trades ES, and JBookTrader is his project to do this. He has been kind enough to add support for other instruments, but when you start playing with it, it really is tailored for ES. For example, for back testing data, JBT assumes a constant bid/ask spread - something that you only see on the most liquid of securities, or with the smallest (worst) price unit resolution. I would argue that this is not a great model for stocks - but great for ES. 2) Positions in ES are pretty heavy. If memory serves me correctly, you are trading approximately a $50k position on margin for each contract - possibly leveraged 5-10x +. For automated sizing based on kelly to be useful, I would imagine you would need to have the ability to wield many contracts - something that most of us simply don't have the ability to do. I am no expert on this topic, so feel free to correct any of my deductions. I hope this helps you see "why" things are the way they are. On Thu, Sep 27, 2012 at 2:07 PM, B <[email protected]> wrote: > That's all well and good, and there's nothing about what you wrote > about finding N and K that I disagree with. You're just performing a > search to find what parameters maximize expected returns in a single > asset portfolio. We're in agreement on this point. Parameter > searches can indeed take substantial computing power. > > But that has nothing to do with my issues with the use of the kelly > fraction. > > The question that the math behind kelly criterion answers is, "If I > have information, how much do I bet?". > > That's it. That's all it's designed to do, because the kelly fraction > is a direct consequence when the question above is posed as a > mathematical equation. Given this, the kelly fraction for any given > asset changes when combined in a portfolio with other assets, and > other trading strategies. > > What you're doing is using the answer posed by the above question as a > solution to a different question. Namely, "How do I know when I have > information?". And that requires a different set of math, because > it's fundamentally a different question. > > I suspect you're using Kelly to gauge the presence of information, and > not as a tool for allocation, which flows from the math. > > Now if the motivation behind jbooktrader is to facilitate automated > trading of assets using systems/strategies with positive expectation > of returns for the purpose of maximizing growth of your capital base, > I don't see how the use of kelly criterion to determine position size > could be overlooked, much less deliberately excluded. > > Anyway, Ed Thorpe explains it better than I ever could. > http://web.williams.edu/go/math/sjmiller/public_html/341/handouts/Thorpe_KellyCriterion2007.pdf > > If there are no plans for implementation, I hope you don't mind if I > fork JBookTrader at some point in the future. > > > > On Sep 27, 2:54 pm, nonlinear <[email protected]> wrote: >> > In either case, calculating the kelly fraction, and *not* using it to >> > determine bet/position size is a pointless calculation. >> >> I disagree. In my previous comment, I used the comparison between strategy >> A and strategy B, but what actually happens in JBT is that it's the same >> strategy that is being evaluated by the optimizer with respect to the >> optimization parameters. So, let's use Lawrence's example of a simple >> crossover system. For example, the rules may be that we buy when EMA(N) is >> above EMA(K), and we sell otherwise. In this particular instance there are >> two parameters, N and K, and the same strategy can be optimized with >> respect to different combinations of these parameters. Furthermore, if you >> plot these combinations of parameters against the corresponding performance >> metric (such as Net Profit, Profit Factor, and Kelly), there is often a >> pattern with what I call a "sweet spot", which is the area of bets >> performance which looks like a plateau, rather than a spike. Now, my >> argument is that identifying these areas using Profit Factor is just as >> useful as identifying them using Kelly. If you are using Kelly, the plateau >> is that elevated surface where a particular combination of parameter values >> (and the surrounding parameter values) yields high Kelly, meaning that in >> the neighborhood of certain values for N and K, the strategy is stable and >> is performing well. Now coming back to my previous 30-60 example, we are >> essentially saying that we have twice more the amount of confidence in >> strategy B, compared to that of strategy A, and as such, we'd rather trade >> strategy B. So, again, in JBT, Kelly is used for >> a non-conventional purpose: instead of using it to calculate the bet size >> relative to the account size, it's being used for ranking strategy >> parameters under optimization. >> >> Now, having said that, there is nothing in JBT that prevents one from >> implementing Kelly the way you seem to suggest, namely calculating the >> position size based on the strategy historical performance. The IB API >> transmits account information, so it's straightforward to capture it. >> However, I have no plans for implementing it at that point. > > -- > You received this message because you are subscribed to the Google Groups > "JBookTrader" group. > To post to this group, send email to [email protected]. > To unsubscribe from this group, send email to > [email protected]. > For more options, visit this group at > http://groups.google.com/group/jbooktrader?hl=en. > -- You received this message because you are subscribed to the Google Groups "JBookTrader" group. To post to this group, send email to [email protected]. To unsubscribe from this group, send email to [email protected]. For more options, visit this group at http://groups.google.com/group/jbooktrader?hl=en.
