Trader's Roundtable

[sagev]

I am reviewing Schwage-Eckhardt interview in the New Market Wizards book. In the conversation, William Eckhardt suggested that only "robust" statistical techniques should be applied to one's trading system. I am not skilled in Statistics, so I could possibly hold the false presumption that all trading systems comprise some kinds of statistical assumptions. Thus, the ability to discern the instability of an instruments seems to be very important when building a system. The original text is below for your reference.

[SMKJ]

Robust - Not overtly curvefitted...

For example... if you back test... and find that the parameters for your MAC works best when it's 10, 20, 8.... but the performance drops off dramatically if you adjust the parameters slightly up or down... in that case..your system is NOT ROBUST. In this case, you prolly happened to stumble upon "well chosen examples"

On the other hand... a robust system... should be that..even if you fiddle w/ your parameters slightly, it should still work relatively well.

Now how to increase the robustness of your system
a) Back test further....
-This helps you avoid "well chosen examples"
-For example, if you were only using the GBP/USD cross rate data from 2003 to do your testing....considering how perfect the GBP's uptrend was.... any crackpot system would have worked. However, by backtesting further...you eliminate the impact of such situations and will find a system that works in more types of market situations

b) Use a longer time frame to trade
-When trading on a longer time frame... you don't have to be as precise... for example, getting in on the EUR/USD when EUR = 0.9, or 0.95 doesn't really matter..b/c eventually it went to like $1.2...
Basically, what I mean is, on a longer time frame, your entry doesn't have to be as precise... meaning that your system could have some extra leeway..making it more "robust".
Warning: Longer trading time frame also requires more patience...and wide stops
Ie) In some markets, my system only gave me 3 signals last yr and my stops are much wider than most ppl would set them...

[Nathan]

i don't think (at least from the passage above) eckhardt was talking about trading systems. He was talking about statistical estimators. He says it very plainly:

"Most classical applications of statistics are based on the key assumption that the data distribution is normal, or some other known form..."

"In general, the delicate tests that statisticians use to squeeze significance out of marginal data have no place in trading. We need blunt statistical instruments, robust techniques."

Basically you want to use statistical tools that do not rely on assumptions about the data, such as the "normal" distribution.

I have no idea specifically what he uses of course, but id think summary statistics, converted into useful metrics that do not rely on assumptions like a normal distribution, would be something of what he is talking about. Basically using statistical tools that minimise the assumptions that have to be made in order to be accurate.

The passage really explains itself though, so im not sure my comment added anything at all.

[sagev]

Thank you all guys for your comment. With your help and a search via google now I think I probably understand the meaning and significance of that concept.

[Christian Smart]

A statistical test or procedure is robust if it maintains the significance level close to a desired significance level a for a wide variety of underlying probability distributions with good power for all the distributions. I think that Eckhardt was referring to this kind of robustness, which is robustness in the face of variations in the underlying distribution. Non-parametric statistics are robust in this sense because they are distribution-free.