has anyone ever backtested a method and then traded it profitably?
- Thread starter 1a2b3cppp
- Start date
actually this “hand-testing” sounds like a recipe for curve fitting
either a system checks out over many trades and different market conditions using a basket of securities or it doesn’t
either a system checks out over many trades and different market conditions using a basket of securities or it doesn’t
I always hand test, probably 8 hours EVERY WEEK.... after years, you become absolutely GIFTED in pattern recognition. And Trading in real time is like hitting a basketball with a baseball bat.
If you are going to be an earner, you cant half a$$ this sheet
If you are going to be an earner, you cant half a$$ this sheet
Last edited:
Interesting what you mention.
Not quite sure on your definition of hand back testing, but I created a new algo a coupla months ago and have run it several hundred times while tweaking, along with eyeballing the results.
Well I've been trading 30+ years now and this lastest algo (supalgo-running a journal on it) has taught me stuff I wasn't aware of before.
There is however issues of biases in the brain, so need to forward test in realtime, real money.
In reality it is just the opposite. Hand testing has to follow strictly the rules of the trading system. But you see at least what happens, when it happens and you can find out why it happens.
Letting the computer do the job will never give as much information as hand testing. Hand testing allows me to see/find relations between things that you never see if you are "numbercrunching" as you will only see the final result. I watch my charts for hours/days, while you just watch the program code. Charts give far more info then source code.
Compare it with driving a car or sitting next to somebody who drives the car. The driver understands better what to do as he is driving. He is more concentrated on the road and sees all the problems. The passenger is more in sleep mode.
Last edited:
Actually, as usual, it depends .... on whether your algo is created from scratch, or whether it's pre-packaged, etc.
If the trader is creating the algo from scratch, then the trader firsts drives the car, then tries to figure out how to program a robot to drive the car. Then, they sit in the passenger's seat and observes the robot driving the car.
Then, they investigate what needs improving; and drives the car themselves again with the corrections in mind; then figures out how to code those improvements into the robot ... rinse repeat.
You see, this approach is actually the most educational. You hear folks say all the time how teaching others helps them learn more. Well so does teaching a computer...to an even greater degree.
What many seem to not realize is that when you create an algo from scratch, you are simply automating what you would do yourself, manually. The algo doesn't create itself from the ether. The manual aspect of trading never goes away. The from-scratch algo is the trader's methodology; reduced to computer code.
The other concern is that manual-only keeps you so close to all the minutia, you may fail to see bigger pictures/relationships in the data. You more easily can miss the forest for the trees.
The trip looks different from an airplane at hundreds of miles per hour, than it does by foot.
From-scratch algo creators are forced to continually look at things from both perspectives.
If the trader is creating the algo from scratch, then the trader firsts drives the car, then tries to figure out how to program a robot to drive the car. Then, they sit in the passenger's seat and observes the robot driving the car.
Then, they investigate what needs improving; and drives the car themselves again with the corrections in mind; then figures out how to code those improvements into the robot ... rinse repeat.
You see, this approach is actually the most educational. You hear folks say all the time how teaching others helps them learn more. Well so does teaching a computer...to an even greater degree.
What many seem to not realize is that when you create an algo from scratch, you are simply automating what you would do yourself, manually. The algo doesn't create itself from the ether. The manual aspect of trading never goes away. The from-scratch algo is the trader's methodology; reduced to computer code.
The other concern is that manual-only keeps you so close to all the minutia, you may fail to see bigger pictures/relationships in the data. You more easily can miss the forest for the trees.
The trip looks different from an airplane at hundreds of miles per hour, than it does by foot.
From-scratch algo creators are forced to continually look at things from both perspectives.
Like you say:Actually, as usual, it depends ....
Solutions might need a combined improvement that is covering several algo's at the same time that are continuously influencing the results in both ways. I speak about building a modular system in which modules work independently from each other, yet interfere in each other as feedback for further calculations and to come to a result. For the average trader not that easy to achieve.
I use pattern recognition. These patterns are created based on a logical decision tree. Programming a computer to recognize different patterns and combine them in a tradable signal is not that easy. I never succeeded even to write down all the rules to be programmed. There are too many and they are too complex. I have all the patterns in my head and I am trained to see in a fraction of a second what to do.
This one is completely wrong in my case. I found things that I could never find using algo’s on a computer. But it might depend on the skills or missing of certain skills of me.
I think that only a small majority is able to use algo's as they should be used. Just using a copied free indicator and think you are creating an algo because you change a parameter is cheating yourself.
Not really the topic I replied to, but...
One way to achieve this is to have an 'overseer' algo apply weights to the independent algos, resulting in a final score/decision. A genetic algo, for example could do this.
Again regarding topicness.
It would be much easier, imo, to train an algo with the patterns that you've seen; rather than trying to reduce these patterns to rules. You would train the algo by feeding it the data (OHLCV 5min/EOD/whatever) as you manually identify as many patterns as you can; and then as you come across them going forward.
The algo could then 'look' for matches/near-matches going forward.
Of course, a decision tree may not be the best algo for this solution.
But I didn't speak against that. I said that you may not be able to see things from an 'algo' perspective, without going through the process of coding an algo from scratch. Maybe one's approach won't benefit from an algo perspective. But I believe every trading system can benefit from the speed and data an automated version of their trading methodology can generate.
I've had algo's that have generated tons of telemetry data that was/is extremely useful. This level of data generation can't be achieved manually.
My trade setup is based on astrological approach. Back tests prove to be quite accurate. The most difficult part, for me, is the entry & exit points.
