I have to preface this with admitting I only know/use 6th grade math but: I just don't get how kelly and the popularized (by tharp etc) bet sizing approach is anything more than wishful thinking (and confidence builder) for traders. 1) kelly and bet sizing is applicable only IF there are known probabilities like in games of chance which specifically excludes the stock market because neither the probabilities nor expectancy can be measured for the future. 2)I may be mistaking my jargon here but I don't think trades are truly independent events. The only value as i see it is when it encourages the trader to assume less risk, those that try to maximize profit are doomed to certain failure if they continue to play the game. The following book details some statistics on winners/losers from brokerage account histories etc. for $4.00 you can't go wrong. The Futures Game: Who Wins? Who Loses? Why? http://www.amazon.com/gp/offer-list...?ie=UTF8&qid=1339251010&sr=8-1&condition=used PS It's dated but I've seen subsequent work by academics and dal that reinforce the ideas.
You're probably right. Thorp's paper also deals specifically with the stock market, but is based on risk, which can not be measured for the future. Thus, "risk-adjusted returns" can be adjusted by anything, but not by risk. You can know the risk of games, but not the risk of trading, thus Thorp uses the variance of historical returns as a proxy for risk. He also assumes a normal distribution of returns, i.e. independent trades. Both is not exactly correct, but a bad approximation is better than none at all. That's why Thorp's method is often used in trading for capital allocation algorithms. There are other methods though, f.i. by Ralph Vince, that do not require a normal distribution of trade events.
Returns can not be measured for the future either. Risk-adjusted returns are measured for the past - you know both and can safely adjust returns for whatever risk metric you desire. Do you actually do what you claim to do or you are a software salesman?
The difference is that trade returns can be measured for the past, but trade risk can not be measured at all, not even for the past - that's what I was trying to explain to you all the time. For calculating risk adjusted return, you use variance, drawdown, or some other proxy of risk. The Sharpe Ratio for instance uses the standard deviation. All this gives you very different performance data. Risk adjusted return is not some objective measurement, as you seem to assume, but it's subjective, dependent on how you define it. Therefore using it would make no sense in the Kahneman study. I do algorithmic trading, and I also sell software, but not trading software.
By the way, why are you always asking what I'm doing? In a discussion it does not matter at all what a participant is doing and how long he is trading. Only the arguments matter. Besides, it is not actually a qualification to have traded for 20 years. The statistics are probably correct that most traders are losing. So the main difference between a veteran trader and a newbie is that the veteran has lost more money.
I specifically said that you can use you favorite risk measure it - it would still be an improvement over simply taking returns. When you look at your algo strategies, you are most certainly are forced to make an assumption of your risk metric, yet you chose to do so. There are a number of assumptions in that study that are just as questionable. Assuming some random index as a benchmark is as much of a loss of objectivity as assuming some sort of risk metric. Assuming that any sample that was volunteered by some random "professional firms" is representative of the overall population is a major (and a very questionable assumption) assumption. Any scientific paper contains a number of biases that reviewers either focus on or choose to ignore. I suspect that you are swayed by the fact that Kahneman is a Nobel laureate and has written a bestseller, hence he can't be wrong. My PhD adviser was a Nobel laureate too - it was way easier for him to get papers past peer review, while personally I felt that the papers were total crap (like the majority of scientific papers). The fact that a hypothesis is true in a cross-sectional sample does not make it stable in a temporal sample. For example, the fact that 95% (I'd recon the number is higher, more like 99%) traders in any given population lose can be representative of attrition which actually would mean that traders that have traded for 20 years are the 1% that actually figured it out. I don't see anything magical about the fact that majority loses - as I said, in any activity with zero barriers to entry you would see similar attrition rates. How many people that know how to cook become profitable chefs? How many girls that take ballet classes in the childhood become professional dancers? It matters - with a professional trader I'd use a different line of reasoning then with a software salesman and yet I'd use very different arguments with someone who has done serious science.
Yes, for comparing or optimizing strategies. For this it doesn't matter when the risk metric is not objective. The same strategies that give a Sharpe Ratio 2 in my trade platform come out with a Sharpe Ratio about 3 in Ninja Trader - despite the fact that the Sharpe Ratio is based on a well defined formula. The reason is that I'm calculating the Sharpe Ratio from the equity curve, while Ninja Trader possibly uses the balance curve, or the trade returns, or a different time frame. So, even the same performance indicators are often not comparable between different platforms. When you publish something, there are good reasons to use a more objective performance measure, even if that does not consider risk. Well, in that case I'd ask you with all respect to use just your best arguments or line of reasoning in future discussions.
All the more reason to uniformly define and quantify your measurements. Gross returns alone are useless bits of trivia for example "trading competitions" come to mind.