Frequency of Re-optimization

S

sidm

Hey y'all,

I am trying to design some moving-average based trend-following strategies. I am a little stumped when it comes to re-optimization though. How often do you guys re-optimize parameters for your strategies?

On the one hand, frequent re-optimization might amount to over-fitting (too reactive to noise).

On the other hand, having too long an interval between re-optimizations might result in parameters end up being out-of-date.

One could possibly optimize the frequency of re-optimization itself, but then that is quite a slippery slope - at what level do you stop?

Any thoughts/discussions would be welcome.

Sid.
 
You could use adaptive moving averages. They reoptimize themselves continuously, so you never have to optimize again. You would then include the adaption process in your backtesting.
 
I use the MACD. It incorporates 2 moving averages of different lengths. When divergence ends and convergence begins, this is the beginning of the overlap of trends. The overlap of a down trend to an uptrend or uptrend to a downtrend. As the overlap begins, congestion sets in, so you can trade the lateral trend that is the overlap and get situated for the ensuing change of trend. Then when Divergence of the MACD occurs, the trend has changed.

The above is something that I posted yesterday. You may find it helpful. Instead of taking the trade as a xross over....look at the construct of the system and make a strategy from that. Hope it helps a little.
 
Here is an example to better explain my dilemma.

Consider a really simple long-only strategy wherein you buy when prices start trending up and sell when they start trending down. Say you use a simple moving average to determine the trend. That is, if the slope of the moving average is positive, we go long. Else, you we liquidate our position.

This simple strategy can be shown to work decently if we get the RIGHT PARAMETERS.

The parameters here are:

(1) Size of the MA window, and
(2) Threshold used to determine how much positive slope is sufficient to deem the trend as worth jumping into.

It is easy to see if MA window or the threshold are too low, we risk being too reactive to the noise. If they are too big, then we risk being too late to the party.

No matter what strategy you use (MACD, adaptive MA) there are always some parametric decisions that need to be made. I am trying to figure out a way to make these parametric decisions as scientifically as possible. One way to do it is to optimize the parameters through backtesting.

So the question is: how often should one re-optimize these parameters?

I am not looking for suggestions in terms of strategy. Every strategy is going to have this issue.



Quote from sidm:

Hey y'all,

I am trying to design some moving-average based trend-following strategies. I am a little stumped when it comes to re-optimization though. How often do you guys re-optimize parameters for your strategies?

On the one hand, frequent re-optimization might amount to over-fitting (too reactive to noise).

On the other hand, having too long an interval between re-optimizations might result in parameters end up being out-of-date.

One could possibly optimize the frequency of re-optimization itself, but then that is quite a slippery slope - at what level do you stop?

Any thoughts/discussions would be welcome.

Sid.
More...
 
Quote from rwk:

You could use adaptive moving averages. They reoptimize themselves continuously, so you never have to optimize again. You would then include the adaption process in your backtesting.
More...
AMAs readjust themselves but they still have to be independently optimized.
 
Quote from sidm:

Here is an example to better explain my dilemma.

Consider a really simple long-only strategy wherein you buy when prices start trending up and sell when they start trending down. Say you use a simple moving average to determine the trend. That is, if the slope of the moving average is positive, we go long. Else, you we liquidate our position.

This simple strategy can be shown to work decently if we get the RIGHT PARAMETERS.

The parameters here are:

(1) Size of the MA window, and
(2) Threshold used to determine how much positive slope is sufficient to deem the trend as worth jumping into.

It is easy to see if MA window or the threshold are too low, we risk being too reactive to the noise. If they are too big, then we risk being too late to the party.

No matter what strategy you use (MACD, adaptive MA) there are always some parametric decisions that need to be made. I am trying to figure out a way to make these parametric decisions as scientifically as possible. One way to do it is to optimize the parameters through backtesting.

So the question is: how often should one re-optimize these parameters?

I am not looking for suggestions in terms of strategy. Every strategy is going to have this issue.
More...
You ask an excellent question.

Off hand I would say you find a backtest length, say N, that give you a satisfactory performance and then reoptimize every N/2 data points after that for N data points.

So the optimization would be from point 1 to point N, the first reoptimization would be from point (N/2 + 1) to point 3N/2, the second reoptimization would be from point N+1 to point 2N, and so on.

There's probably a better way to do it and I hope some other member will post it.
 
On 2-4 years of data, optimization's good between 1-3 months before having to scale the system to what is likely a different "range" of percentages.

Stocks aren't as important in terms of needing to re-optimize a portfolio basket watchlist, but in futures the ticks of percentages are disproportional if you have a big change in the value of a contract following those optimizations.
 
Strike 1. Data for monitoring and analysis is limited to the width of the window in which each trend resides.

Anyone who violates this principle is fucked.

strike 2. another aspect of science and mathematics that is important is that there are independent and dependent variables. by four grade most people learn the the behavior of the dependent variable is a function of the independent variable.

strike 3. In any seeking or problem solving adventure, you have to use the correct kinds of math to solve the problem (or even see the opportunity.

re-read the thread to see who is striking out. Batting .300 for a potential trader is striking out. trading is not baseball or base running.
 
Quote from jack hershey:

Strike 1. Data for monitoring and analysis is limited to the width of the window in which each trend resides.

Anyone who violates this principle is fucked.

strike 2. another aspect of science and mathematics that is important is that there are independent and dependent variables. by four grade most people learn the the behavior of the dependent variable is a function of the independent variable.

strike 3. In any seeking or problem solving adventure, you have to use the correct kinds of math to solve the problem (or even see the opportunity.

re-read the thread to see who is striking out. Batting .300 for a potential trader is striking out. trading is not baseball or base running.
More...

Right, but you can only hit .900 if you use ma's correctly. Sma for 1 avg is only a first derivative and the third ma is a change in concavity which is way more important.

You get a channel, a trend, and then concavity change immediately when the market's direction changes. Averages show the markets move and not when it is about to change.
 
Quote from sidm:

(2) Threshold used to determine how much positive slope is sufficient to deem the trend as worth jumping into
More...

Slope based on what scale? By zooming on the chart the slope changes.
 
You must log in or register to reply here.
Back
Top