Lessons Learned From Profitable System Development

Quote from 0008:

Are your methods quite subjective? It is hard to determine which one is a bad data. And some spikes are huge but some are not. I know some professors of statistics did research in this topics. But I remember that they couldn't make a very great improvment of the accuracy of the forecasting (it was not about trading).

I agree "Filtering out smaller moves is important". But I think this is very very hard, no matter what methods you use. Otherwise we could get a 100% winning system! I tried many methods, but no one could remove all the wipesaws. And you?
More...

Not subjective at all - a computer program does all the tick filtering. You just have to come up with a definition of what a bad tick is. For me, a tick that is outside my % range from the baseline price is a bad tick. If I get many ticks at that price, then those are good ticks and my baseline price needs to be altered.

It's really a simple concept. Assume that your first tick is a good tick. Store the value of that tick as the lastprice. When the next tick comes in, see if that tick is within .5% of lastprice (vary this % depending on the normal tick range of your market). If you get more than 10 ticks greater than .5% from lastprice, set lastprice equal to the current tick. If the current tick is under .5% from lastprice, set lastprice = the current tick price. Do this for all the symbols you collect.

I also have a larger 1% filter, but the # of ticks there is larger - from memory requires 30 or so ticks outside 1% of baseline before the baseline lastprice is changed and those ticks are transmitted as being "good".

HTH,
 
Filtering "bad" ticks has two aspects:
- cleaning historical data for backtesting
- filtering incoming ticks for realtime trading

I experienced that data quality differs dependant on the provider. E.g. you can get for little money raw data from EUREX but this data contains a lot of bad ticks. I applied a filter algorithm - quite similar to what TriPack proposed.

For not getting trapped in real trading I should supply the same filter on realtime data.

Other data providers apparently have high quality data. Here, filter mechanims may be applied only when trading realtime and not for backtesting.

Hope this added some value.

Dierk
 
Good systems:

- are easy to trade. They don't have rules like buy on open of next bar using 1 min. data. They give you enough time to actually get fills similar to the tested results.

- don't try to capture profits in all market conditions. Every system I have is targeted to a specific market behavior.

- have a edge superior to random entries and exits.

- are stable. Market behaviors with a edge happen with varying frequency. However they should be both frequent enough and with sufficient size moves so as to warrant taking prudent risks.

- are not curve fit. Use entries/exits versus random entries with same holding period and a chi square test to avoid fooling yourself.
 
Quote from 0008:



I agree "Filtering out smaller moves is important". But I think this is very very hard, no matter what methods you use. Otherwise we could get a 100% winning system! I tried many methods, but no one could remove all the wipesaws. And you?
More...

IMO there are 2 types of whipsaws, those caused by bad data, and those caused by the market based on how your system looks at the data. You can virtually eliminate the first type by tick filtering. Eliminating the 2nd type is more difficult, and it is a game of minimization of whipsaws which involves tradeoffs. And sometimes your system is just plain wrong, so there is no system that wins (or needs to win) every time.
 
Nice post, Smitty. What testing platform(s) meets your requirements for #2?

Quote from mail2smitty:

My take on going down the systems development road:

1. Simpler is better. Have to watch myself on this one as the love for computers & math will lead me astray. The really cool stuff like neural nets and signal filtering algorithms can help a little, but don't make the mistake of equating higher tech with higher performance in trading. There seams to be an inverse relationship here. Some folks seem to make this work, but they're usually so deep in it that its hard for others to understand what the heck they're doing (see HarryTrader!).

2. Get yourself a solid backtesting platform and use the hell out of it! You'll be amazed at how much of the stuff out there DOESN'T work. IMHO that's the most important feature of a backtesting platform - demonstrating that something doesn't work. This will also provide much needed exposure to the market's workings.

3. Search out and read all posts here and on other similar websites that are put up be systems traders. Other good websites for this include the message areas of the backtesting software maker's websites. You'll discover the serious systems traders by just browsing or searching the site. These posts provide more direct and timely information than most books I've read.

My systems: Have two countertrend and one trend follower. The countertrends only trade about 25 times a year for 1 or 2 days holding - they're my bear / choppy market winners. The trend follower is designed to sense a major trend and ride it as far as possible given its parameters - its a bull market trader.

Just my opinion,
Smitty
More...
 
Quote from mail2smitty:

My take on going down the systems development road:

1. Simpler is better. Have to watch myself on this one as the love for computers & math will lead me astray. The really cool stuff like neural nets and signal filtering algorithms can help a little, but don't make the mistake of equating higher tech with higher performance in trading. There seams to be an inverse relationship here. Some folks seem to make this work, but they're usually so deep in it that its hard for others to understand what the heck they're doing (see HarryTrader!).
[]
More...

Good post! Specifically, I found that you really want to limit the number of "conditions" to one to two per independent time series.

For instance, don't use both rsi, stoch, macd, and 7 different MAs to time an entry. They are all based on the same time series and therefore not independent. I've found one or two conditions to be the most that is statistically stable.
 
the only way to truly know if a new high frequency system is tradable is to trade it. if something looks really good, don't ponder and sim trade it for months, trade it now. keep ultra accurate records of actual vs system fills, chart them against each other, and if the system is too good to be true, you'll know in a few days.

also, if you're trading more than one strategy on the same instrument in one account, maintaining a net position of the two systems, keep in mind that slippage on one system will also come out of the other's equity curve too.
 
Quote from TriPack:

To start the thread off, let me share a few things I've learned from system development. This is from developing a counter trend system.

1) Filtering out smaller moves is important. This does two positive things for the system, reduces the # of whipsaws (entry in a flat market) and increases the % profitable. One side-effect is that whenever you filter you introduce lag. See:

http://www.elitetrader.com/vb/showthread.php?s=&threadid=21739

2) Only taking entry signals on bar close, rather than accepting intra-bar signals. This has the same positive and negative effects as Filtering moves, and is in fact a type of filter. But it is such an effective filter that it bears mentioning by itself.

3) Timing the trade: Place buy order on weakness, and sell order on strength. Even with a method like a breakout method, it is possible to implement this strategy by only entering on a pullback after the breakout has occurred. For a countertrend system, a breakout is the pullback.

4) Filtering the raw data, then applying simple analysis to the filtered data is superior to using raw data and applying complex analysis to the raw data.
More...

Agreed, filtering the small moves is usually why I add in "extra indicators". I first find a system that works then add in things to keep it out of bad trades and small trades.

Taking a signal on a partial bar is really second guessing a signal on a completed bar in another timeframe. This knowledge points to the idea of running systems on many timeframes on the same instrument.

If you cannot filter data in your particular software package you can still distance yourself from bad ticks. A bad tick is not likely to also be on the close of a bar so if you use indicators that only use the bar closing price you will sidestep a lot of bad ticks. Going to a slower bar interval even further helps with that. Filtering bad ticks in your signal code by not entering until a confirming tick is not hard to implement and won't slow your system as much as filtering will.

It should be interesting to produce some range bar data by altering it with Excel and see what happens with some tech analysis applied.

Thanks for the discussion.
 
Currently use WealthLab along with an occasional Excel sheet. Used Amibroker for a year or so before buying WealthLab, the two are alot alike in ability. Amibroker is probably the best "bang for the buck" but requires more of the user. WealthLab is easier for me to automate and is just about perfect for my style. Also really like the community built around the WealthLab website. Have gotten some good ideas just poking around with other peoples code (WL lets people share trading system code either via web downloads into the program or by surfing the website).
Also have Neuroshell and a few other C++ NNs, more for play and experimentation than trading (love that stuff - tons to learn!).
Have a friend who uses TradeStation - looks like a good platform from what little I've used it. Sure there are other good ones out there. Takes alot of web surfing and experimentation to find the perfect one for you & your style.
Smitty
 
You must log in or register to reply here.
Back
Top