You are using an out of date browser. It may not display this or other websites correctly.
You should upgrade or use an alternative browser.
You should upgrade or use an alternative browser.
What's the best way to visualize 4D data?
- Thread starter mizhael
- Start date
Quote from mizhael:
From "You have no chance with your three parameters. IT will take just a few bars to lock your system in a bifurcation mode",
are you saying that optimizing a strategy with 3 parameters has no use at all?
It may have some use I am not aware of but one thing I know for sure is that if you are using 3 parametes depending on what they are doing you can be doomed. If they are indicator parameters that introduce smoothing operators (MA, MACD, DMI, etc) your system will be constrained to a strange attractor set fast due to the fractal nature of the market.
So you have three problems, at least. (A) Smoothness of 4-D hypersurfaces in the vicinity of parameters (B) Avoiding strange attractors and (C) finding heteroclinic orbits to quarantee stability.
With three parameters, this is an excercise in futility, I warn you.
Some people may think I am talking BS because they have no clue. The less some poeple know the more they think they know. The more I know, the less I realize I know. Example to understand the practical significance of all these. Simple trend following system
2 MA parameters for cross, RSI parameter for overbough/oversold. This is easy but just an example to illustrate (A) - (C) above.
(A) You optimize and you get nice smooh hypersurfaces near optimal values, historically that is.
(B) Market gets to a pivot point which coincides with your MA cross. It keeps isolating, you get no trend and losses increase. You get no way out because RSI remains oversold, the bastard has sold you to Wilder. You need a way out fast. You are in a strange attractor, which is what the pivot point is for your type of system.
(C) You get a signal and you catch a nice trend. Your system is stable. You need to find an orbit fast to maintain this stability. Sudden reversal erases your profit. You were not able to get to a heteroclinic orbit on time. After the reversal you get to a strange attractor. Yes, it happened to me in 1996 trading copper futures in a market that we found out 2 years latter that it was rigged by some Jap market maker. I lost alll profits of about 500K plus 20K more. Now, you are downhill. A pink slip is waiting for you on your desk if you are professional or your girlfriend finds someone else if you are single. If you are married, you should know that 3 things are certain in life as of recently
(1) Your lawyer will screw you royally (this holds eternally)
(2) America is becoming socialist state unless we take drastic action
(3) Your wife will cheat on you with a much younger guy
I hope you got the message(s). Good luck.
Quote from mizhael:
hi intradaybill,
Thanks for the horror stories, but what's the way out?
Would you please share with us your experiences in tackling this problem after those 15 years?
The answer is easy and it works. Use rules with no parameters for optimization. Although even the act of selecting a system is some form of optimization, according to many clowns, the impact is much less when there are no parameters to optimize. A triviall example is the system C > C[2]. Although the clowns who confuse notions will assert this selection is optimization, even in that case it is much better than using something like f{a} > g{b} because the former is just price action itslef.
The problem is that it is difficult to find enough robust rules that are not optimized and are high probability setups so to get a sufficient edge, numerically. It is becomiing more difficult each year. if you read what market wizards did they all used simple rules, not the fancy things you are trying with 4-D data analysis. It takes simple rules for entry and then the emphasis is placed on risk and money management.
Quote from mizhael:
Okay I have tried, but it's too hard - due to my poor space imagination...
Lets reduce the difficulty a bit.
Essentially we want to find a few good local maxima in the sampled function:
f(p1, p2, p3)
What are good local maxima?
1. Of course they have to be local peaks;
2. But they cannot isolated points, i.e. the peaks surrounded by abyss. They should be the peaks that have nice smoothly decaying neighborhood...
Any thoughts? esp. in Matlab?
Thanks!
As a rought and ready first approximation I would be calculating mean value at each point using the point itself and all its immediate neighbours (or first and second neighbours, etc). You could experiment with weightings (eg depending on 'distance' from the immediate point). Unless there are real pathological values in your initial data set, I think the points with maximum mean values would pretty much identify the regions of interest - then you look more closely around there, and satisfy yourself that the region is uniform enough. Your may decide that your optimal value is not one of the local peaks - ie you may decide to take a value located in the centre of a region, rather than a local peak that occurs near the edge of that region.
Quote from intradaybill:
The answer is easy and it works. Use rules with no parameters for optimization. Although even the act of selecting a system is some form of optimization, according to many clowns, the impact is much less when there are no parameters to optimize. A triviall example is the system C > C[2]. Although the clowns who confuse notions will assert this selection is optimization, even in that case it is much better than using something like f{a} > g{b} because the former is just price action itslef.
The problem is that it is difficult to find enough robust rules that are not optimized and are high probability setups so to get a sufficient edge, numerically. It is becomiing more difficult each year. if you read what market wizards did they all used simple rules, not the fancy things you are trying with 4-D data analysis. It takes simple rules for entry and then the emphasis is placed on risk and money management.
Are you talking about systems containing dependent variables? Would you make the same assertions if the 3 variables under consideration are all independent?
Quote from intradaybill:
The answer is easy and it works. Use rules with no parameters for optimization. Although even the act of selecting a system is some form of optimization, according to many clowns, the impact is much less when there are no parameters to optimize. A triviall example is the system C > C[2]. Although the clowns who confuse notions will assert this selection is optimization, even in that case it is much better than using something like f{a} > g{b} because the former is just price action itslef.
The problem is that it is difficult to find enough robust rules that are not optimized and are high probability setups so to get a sufficient edge, numerically. It is becomiing more difficult each year. if you read what market wizards did they all used simple rules, not the fancy things you are trying with 4-D data analysis. It takes simple rules for entry and then the emphasis is placed on risk and money management.
In C>C[2], is this "2" a parameter?