Optimizing studies and systems

[gspiker]

Hi all,

I have written a few self optimizing studies, and am now thinking of moving to the next step of a self optimizing system.

There are a few ways of approaching this, the easiest of which is to just make each of the studies in the system self optimizing and hope that this will make the system (all the studies working together to generate one overall system) work at the optimal level.

This seems flawed thinking however, as it seems to me it is possible that (for example) a slightly "detuned" study may work best with another slightly "deturned" study. After all it is the overall system I want to work at the optimal levels.

This implies that I would have to do some major regression testing to find the optimal levels on how all the studies work best together.

Some questions to any who have read about, or traveled this road before:

1) Are my thoughts correct. The best system may not be one where each study is giving the best signals? This seems obviously true...but thought I would check

2) Any good methodology or algorithms for quickly determining what is optimal?

To forestall the "no curve fitting" statements that may be made, typically I backtest on daily bars and go back 10 years or so...

Thanks!

Garth

[bfry5282]

Garth,
What is the mix of studies you are talking about?
And what is the goal of the system, other than the obvious of generating signals which are optimally profitable?
Always interested in new ideas like this, but I think it sounds like a good idea in theory but one which can never be achieved in practice... I'm willing to be convinced otherwise

[gspiker]

bfry5282,

>What is the mix of studies you are talking about?

Most of them are my own (though I do use AGET tools also) but I think to test the waters I might go with a system system using CCI or MACD, Stoch and MSH/MSL. This way I can use mostly built in studies and to do a quick and dirty proof of concept (POC).
And that way, if I do put something together that looks interesting I can share the initial version with all.

>And what is the goal of the system, other than the obvious of >generating signals which are optimally profitable?

Well that is the key. I'm more a swing trader/position trader so I am looking at getting in early on longer term moves. But the same methodology should work for optimizing any objective system.

>Always interested in new ideas like this, but I think it sounds >like a good idea in theory but one which can never be achieved >in practice... I'm willing to be convinced otherwise

Well this is just automating the work that many traders do manually...and since there are sooo many variables when you start combining more than two studies hand testing each possible valid parameter for each study become impossible for all but the most anal...uh diligent traders.

This isn't new ground either...many AI like programs out there do this.

The follow-on step to this is to add adaption (within ranges) to the studies so they can modify themselves for more recent market conditions. This gets even more complex, and I suspect may yield less advantage (if any advantage) over the methods I am now exploring.

Garth

[spike500]

Garth,

I’ve done what you are planning to do. I combined a trending system based on MA with faster indicators. The purpose was to overcome the problem that MA gives well the direction of the trend, but it’s always lagging behind. So you have to “speed up” your MA to not miss the biggest part of the moves. The other indicators were necessary to find good entry and exit points.

I first tried to combine these indicators in a scheme so that for each combination I knew what to do. After a while I had more than 10 000 different theoretical combinations. After eliminating certain combinations that were ilogical and didn’t make sense I was still having problem because I still had to organize more or less 300 combinations. This was due to the fact that I work in multiple time frames.

The next step was to try to manage these combinations. Therefore I used AI (Brainmaker). I thought that the computer was going to solve my problem. Unfortunately nothing useful came out after billions of calculations. The computer was calculating day and night to combine these indicators on large series of data and trying to link these indicators into a system.

The last and most desperate step to get a solution was to try to think analytical and do it manually. I succeeded, but I think I was lucky, after many different attempts.
It took me approximately 9 years before the system was completed and things started to work well. The first 5 years I worked 8 hours a day 6 days a week on it.
If I would have know in the beginning how difficult it would be, I probably would never had started. But on the other hand I am addicted to developing trading systems.

Final conclusion: it was worth it, but for each person that succeeds I think at least 50 others will never succeed although they made the same efforts.
For as far as I know I don’t think it is possible to design a good system by software. The logical and analytical part has to be done by a human being.

My experience says that you cannot optimise each indicator separately, you have to optimise them all together because some indicators have an other function than others. That makes it so complex and time-consuming. You start with a certain setup that you have in mind and you have to dig it out completely. If than the results are not good you have to start over again with the same indicators in another setup and restart the analytical thinking. Perhaps you even have to add are eliminate indicators, or interpret them in another way.
Believe me I have been many times desperate, but I’m obstinate so I never really gave up.

[gspiker]

Hi Spike500,

Thanks for the info. The manual time consuming stuff is what I want to avoid, and anything I can do by hand (and least when it come to numerical/stat analysis) I can get the computer to do...even if I have to export data to a program I create.

But you do confirm that the system has to be tuned as whole, instead of tuning each specific study and assuming that they will work togther as a whole.

I think we are trying for similar goals, so it should be interesting.

Thanks for the reply!

Garth

[spike500]

Garth,

This is an example of the problems that you can encounter when doing research.
The example is not realistic but has been manipulated a bit to be able to show what I mean. This example does not represent any attempt to making a trading system.
The question is: how are you going to “automate” the research instead of analysing it manually? I never found any other solution than doing it manually.

We see on the chart a MA line. We consider this lines as the trend indicator. Above the line is a long trend, under the line is a short trend.
The indicator in the lower part is a stochastic indicator.

I want to check if the following rules work out well:
when the trend is long I buy under the level of 50 and I exit when the stochastic drops again under the 80 level.
When the trend is short I sell above the 50 level and I exit when the stochastic raises again above the 20 level.
Now comes the most difficult part: these trades can only be done if on monthly charts the trend is long. But the stochastic has to be lower than 80 to go long and higher than 20 to go short on monthly charts.
So you have to combine different timeframes in 1 analysis. I don’t know of any software that is able to do this. Second problem is that you have to be sure that the programming and the generated trades are correct. This can only be checked by manually checking certain trades. On top of that the more complex the rules, the more difficult to check the correctness of the results.

The example above is a very simple one, in reality the rules I use are much more divers and much more complicated. So also much more complicated (if not impossible) to automate.
I have however a lot of experience in manipulating data in TXT, CSV and other formats and putting them in Excel or Acces files to analyse them. But I came always to the same conclusion: do it manually.

[spike500]

What i forgot to mention is that even in this very simple example you will have a lot of work to do.

Which parameter will you use in your MA?
Which parameter will you use in your stochastics?
Is it perhaps better to enter or exit at other levels than 20, 50 and 80?

If you combine these parameters in all possible combinations you will end up with millions of different combinations.
That's where the real work starts.

[sc1317]

I'm new to writing a system, without any success I may add, but I am interested in learning..Maybe someone here can help me in understanding code..

I have really been searching for the ultimate trigger point indicator, but also I feel you need more indicators to support the entry indicator in order to obtain a real sweet spot entry..

After many, I lost count, I have found that:

Stochtasics work well, but each individual stock has a different characteristic, thus leading to many false entrys and or exits..

MA's, although excellent can trigger to many entrys if set tight and entrys if set wider are way to late or no trigger at all..

MACD is the same..

Best I have found so far is Parabolic SAR, but it also has many drawbacks..

Feeling that a combination of Stoch, MA and SAR would make excellent entry points..Time frame is important..SAR really does not work well in short timeframes, less than 2 minutes gets you way to many false exits..

After much consideration and thought, I think a SUPER "wildcard" indicator would be based on either ES or NQ..

Is there an indicator based solely on the mini's?

I want to learn and this forum seems to be the place to be..I have so many pages of written things that I feel will work and actually does when looking at it, but I can't get it written in the code needed to test it or use it..

Thanks and I hope I can learn from all of you!!

Racer

[bfry5282]

Volume seeking behavior

You ask whether there's an indicator based solely on the eMinis.

eMini contract trading follows the same principles as equities trading, where the big players respond to the same factors.

The biggest factor is that price movement seeks inflection points of maximum volume. Volume seeking is the number 1 behavior of market makers in nearly any issue. They move price in response to volume from retail traders. The reason for this is that they are making the spread, so the more volume, the more they make.

My advice is to pay careful attention to where volume enters, and to make sure you do not anticipate a price pivot until big volume enters. Big players will buy from the public, and lower price; thus averaging into long positions, but losing money on paper as they do so. Price will continue to move in the current direction until the maximum amount of volume has been extracted from the trend. Trade the trend until volume begins to dry up, and then anticipate pivot.

[spike500]

Re: Volume seeking behavior

<The biggest factor is that price movement seeks inflection points of maximum volume.

I never found any reliable volume indicator. The only importance of volume to me is that there must be enough of it to let me get in or out at a reasonable quote.


< Volume seeking is the number 1 behavior of market makers in nearly any issue.

Is there a marketmaker in E-minis? I thought it was a completely computerized matching system. The spread is always 0.25 to me.

[bfry5282]

Spike,

The term "market maker" has a specific meaning in equities, but eMinis have defacto aggregate big money players who fulfill exactly the same function, and behave in the same way.

eMini trading is like what the Nasdaq would be if, say, everyone were forced to trade on Island.

MM's in eMinis trade against the small trader in 2 ways: 1) by making the spread and, 2) by accumulating net long or short positions against the retail trader. Most small traders are retail traders, either through inexperience or through a practical inability to get a wholesale fill due to crowding scalpers. This is most obvious in issues such as ES (eMini S&P 500). You will usually be unable to get a timely wholesale fill in ES.

PM me if you want to discuss further, since this is probably getting off-topic. But the right analysis reveals volume patterns which are highly predictive of short-term price movements.

[michaelm]

Re: Optimizing studies and systems

Originally posted by gspiker
...
Some questions to any who have read about, or traveled this road before:

1) Are my thoughts correct. The best system may not be one where each study is giving the best signals? This seems obviously true...but thought I would check

2) Any good methodology or algorithms for quickly determining what is optimal?

To forestall the "no curve fitting" statements that may be made, typically I backtest on daily bars and go back 10 years or so...

Thanks!

Garth

Hey Garth!

1) Yes. Best to test system as a whole...not the sum of individual parts.

2) Great question. However, how do YOU define optimal? Is it simply testing a system to produce the most winners and least losers? Is it risk/reward? Is it finding consistency? Is it scooping most of a trend?

In essence, is it optimizing the TA, money or time management?

Look forward to what seeing what you unveil.

[theplumber]

Originally posted by bfry5282
Spike,

The term "market maker" has a specific meaning in equities, but eMinis have defacto aggregate big money players who fulfill exactly the same function, and behave in the same way.

eMini trading is like what the Nasdaq would be if, say, everyone were forced to trade on Island.

MM's in eMinis trade against the small trader in 2 ways: 1) by making the spread and, 2) by accumulating net long or short positions against the retail trader. Most small traders are retail traders, either through inexperience or through a practical inability to get a wholesale fill due to crowding scalpers. This is most obvious in issues such as ES (eMini S&P 500). You will usually be unable to get a timely wholesale fill in ES.

PM me if you want to discuss further, since this is probably getting off-topic. But the right analysis reveals volume patterns which are highly predictive of short-term price movements.

Your exactly right, how do you know this? I didn't know this until I became part of the industry(will give no explaination) and was sent to Chicago to observe the locals and players tricks so to speak first hand. Funny thing is when I said almost the same thing you said to someone on Elite I was told I'm an idiot and knew nothing about trading, how ironic since I traded one day last month 12K R/T's . The human element is always excluded with systems, why?

[bfry5282]

I'm interested to hear that you have inside knowledge of some of the tactics used by big money to manipulate markets. I know it second-hand through data analysis.

I don't usually make negative posts, so if I did, in your case, then accept my apology.

Let's PM if anyything further, and thanks for the info.

[gspiker]

Spike,

I don’t know of any software that is able to do this. Second problem is that you have to be sure that the programming and the generated trades are correct. This can only be checked by manually checking certain trades. On top of that the more complex the rules, the more difficult to check the correctness of the results.

Combining multi-timeframes isn't an issue, I already do that and backtest a system that uses the results from higher time frames to influence lower.

I agree, verifying that your rules are being generated correctly can be difficult, but I have a methodology I use (and may, once I clean it up a bit, post to the group) that makes the visual inspection easier. It still requires time, but after you verify your rules you can then backtest on very long timeframes and on different issues fairly quickly.

Now imagine this, I can test my rules visually fairly quickly even though I have a method of power ranking my signals...which means (this isn't a real example) that if I get a stoch bounce from 0 that crosses the 10 line I could rank that a power of 10, but if it bounced from 5 to cross the 10 only rank it a 8. I then use a formula I have that takes the total of all these signals and gives me an overall power rank which determines if it is a strong enough trigger to enter. It is even more complex than this, and ultimately the power rank for each studies specific signal may end up being dynamic...but that is far down the road.

I suspect that the best way to attempt to do what I am doing would have been to export all the data and make a modeling language for it...but I am still doing all of this with efs and starting to suffer from some of its limitations and the complexities in the code I am forced to use to work around them.

EFS2 will be a huge help when it is finially stable and released.

Garth