(OFF TOPIC) Hi Loren, Rainier asked me to comment on the article about profit volatility in TASC. I agree with you that the article was misleading (or maybe misguided would be a better word), and I also agree it was more an exercise in math/stats.
The short answer is: the model is bogus, therefore the results are bogus.
The author makes the statement "all things being equal, the system with the lower volatility beats the system with the higher volatility." The implication here is that the proposed "system" represents a real-world trading scenario. Nothing could be further from the truth.
The long answer:
Anyone that's ever tried to model a process mathematically (for the purpose of predicting behavior for a given set of inputs) knows the results are only as accurate as the model of the system and the validity of the inputs. There is no way that a zero-mean, uniformly-distributed random variable, with a fixed range of +/- 50% (which is what he used) is even a rough approximation to market price reality (over any time span).
This essentially says a +50% return is equally-likely as a -50% return at any given time. In the immortal words of Norman Schwarzkopf, that's Bovine Scatology.
Since the model isn't a valid representation of market behavior, neither are the results (no matter what input conditions we use). It was just a math exercise in the properties of a dependent sequence of random variables that follow a specific distribution.
I would also suggest that the results depend on the properties of the "random" number generator (which is really a pseudo-random number generator, and probably a weak one), especially the distribution, balance and run length. The number of consecutive "random" numbers allowed to clump in one polarity (negative, for example) affects the results, as does the moving average length that optimizes the zero-mean assumption, as does the sum of consecutive same-polarity numbers. Many different versions of this random number generator could be built, all having different properties. When each trade is dependent on the (result from) the previous trade, as in the author's system, these properties becomes very important. Even if these properties were considered (they weren't), a uniform distribution is one of the worst choices of statistical distribution a person could make.
But the statistics of any stock's price behavior will change over time as perceptions and expectations regarding the future change, and even as history changes (a la Cendant). As well, the time frame of the dominant holders can change (for example, momentum traders pile in and dominate the price behavior over a previously longer-term dominated sentiment). Fixed distributions might be good approximations for mature, established, fixed-growth-rate (i.e. stagnant) markets, but who invests in those? :o).
So it's not sufficient just to gather statistics on the trading behavior over some arbitrary time-frame and come up with a model, because that time frame may include several behaviors that are at best weakly-correlated. Any statistics that try to aggregate the price fluctuations over a period of multiple behaviors into a single representative statistic would be in error.
That doesn't mean fixed, arbitrary statistics or indicators can't be used to make you money, but they only work if the future behavior is closely approximated by the past behavior (trend following), or, if the future behavior has a positive excess deviation from the past behavior (luck). When using fixed, arbitrary indicators or statistics, we need to have a way to bring a negative excess deviation from "normal" to our attention (since this says the model is no longer representing the stock's behavior, and the continuation of the old behavior is questionable). Your addition of downside-risk-limiting stops would help in that regard.
P.S. Sorry I got a little carried away rambling on. Haven't posted much the last couple months, so I had all this pent-up lunacy to unleash.:o) (At least I spared you from the adaptive indicator discussions).
dh
|
