Thanks for the tip, was there something specific you were referring to, or just a general statement?
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.
Avoiding Curve fitting
- Thread starter maninjapan
- Start date
Each bar, and more appropriate if you are using daily bars, has 4 components. Additionally you have a "lookback period". This gives a assortment of combinations. You don't need too long a lookback period. The rest I'll leave to you. Its right there.
Thanks Muskoka, Ill be checking that out
Quote from maninjapan:
What I am interested in now though, is on any given period one set of parameters will outperform others, and this will change, so if I took 3 or 4 sets of parameters (for example, a 5, 10 and 15 ATR period) and split up my positioning across these 3, would that not help smooth out the equity curve slightly more??
Yes that would smooth your equity. But you need to be sure that each combination of parameters is profitable on the OOS period.
The next step is to develop another system on a totally new approach. You have breakout system, then research on counter-trend systems...
RedRat, looks like we are on the same wave length here. Im actually working on a long term trend following, short term counter trend as well as the break out system. I know this will smooth out my overall equity curve. In this case I am jsut interested in smoothing out the curve for this particular system. As I said the profitable parameter range I tested is above water across a firly broad range.Â@I just figure that using 3 different sets from across that range will give me smoother results over a period of time ( providing the system itself remains profitable)
Quote from Code7:
I agree. Let's say you buy if the close is greater than the previous close. Now you need to determine what the *previous* close is. As said earlier, any delta amount of time or volume or price is a parameter. If you adjust that parameter somehow to improve any performance measure, you are already curve fitting.
- If your model is c > c(n) and you then vary n for optimum performance then you are optimizing model selection, you are not curve-fitting a particular model by parameter optimization. The result can be something like c > c(2), and this CANNOT be changed afterwards by any parameter optimization.
- If your model is c > c + x, and you vary x then you are optimizing a particular model and you are probably curve-fitting it to the data.
- If your model us c > c(n) + x and you vary both n and x, you are optimizing model selection and curve-fitting at the same time, and these operations are not independent, you are essentially curve-fitting.
Quote from intradaybill:
- If your model is c > c(n) and you then vary n for optimum performance then you are optimizing model selection, you are not curve-fitting a particular model by parameter optimization. The result can be something like c > c(2), and this CANNOT be changed afterwards by any parameter optimization.
- If your model is c > c + x, and you vary x then you are optimizing a particular model and you are probably curve-fitting it to the data.
- If your model us c > c(n) + x and you vary both n and x, you are optimizing model selection and curve-fitting at the same time, and these operations are not independent, you are essentially curve-fitting.
============
IMO any input that can vary is an optimzable variable or parameter. There are no marks from the market for using what might be the "correct" term in a university math class. In a practical sense both n and x are optimizable.
No matter name you put on the variable, parameter or processes, if you test it using different input values your are optimizing it in a practical sense. And that is not a bad thing, it is a necessary thing. Optimizing is the only way to determine what is useful and what is not. Curve fitting is just optimization done poorly.
In terms of optimizing/ curve fitting
If close > close (N days ago) then ....
is the same as
If Close > (open + X) then ...
One optimizes time, one optimizes price. Either will need to show charateristics of a valid optimization.
Quote from Muskoka Joe:
In terms of optimizing/ curve fitting
If close > close (N days ago) then ....
is the same as
If Close > (open + X) then ...
One optimizes time, one optimizes price. Either will need to show charateristics of a valid optimization.
Not correct. One test of the falsity of a proposition or thesis, like the one you stated, is determining whether it leads to an absurd conclusion.
Your proposition is false because it leads to the absurd conclusion that every conceivable mathematical model is curve-fitted. However, curve-fitting only refers to the particular act of varying an independent variable for the purpose of optimizing some objective function. In your sense any model of the form
M {price, volume, indicator, algorithm, rule, etc.) is curve-fitted because you can make selection of the indicators, algorithms, etc.
It is an absurd thesis that abuses the mathematical notion of curve-fitting which is well-defined and specific.
http://en.wikipedia.org/wiki/Curve_fitting
Many people confuse curve fitting and model selection. Your simple rule should be as follows
If your parameter selection results in a model with no free parameters then you are involved in model selection. If your parameter selection does not eliminate free parameters then it is (possibly) curve-fitting.
Quote from intradaybill:
Not correct. One test of the falsity of a proposition or thesis, like the one you stated, is determining whether it leads to an absurd conclusion.
Your proposition is false because it leads to the absurd conclusion that every conceivable mathematical model is curve-fitted. However, curve-fitting only refers to the particular act of varying an independent variable for the purpose of optimizing some objective function. In your sense any model of the form
M {price, volume, indicator, algorithm, rule, etc.) is curve-fitted because you can make selection of the indicators, algorithms, etc.
It is an absurd thesis that abuses the mathematical notion of curve-fitting which is well-defined and specific.
http://en.wikipedia.org/wiki/Curve_fitting
Many people confuse curve fitting and model selection. Your simple rule should be as follows
If your parameter selection results in a model with no free parameters then you are involved in model selection. If your parameter selection does not eliminate free parameters then it is (possibly) curve-fitting.
==============
What is absurd is that you somehow from my statement come to the conclusion I believe or said that "every mathematical model is curve fit"!!!
In a practical sense for developing trading strategies, not handing in a university math paper, any time you introduce an input with a range of possible parameters you will be optimizing your trading system. Call it whatever you like.
Under your scenario "model selection" and "parameter selection", come with the same risks as they apply to the practical development of trading systems.
Quote from Muskoka Joe:
==============
What is absurd is that you somehow from my statement come to the conclusion I believe or said that "every mathematical model is curve fit"!!!
I am not responsible if you do not comprehend the ramifications of your own statements. They are quite absurd and false.