A few fundamental questions about modeling

C

chawwal

First post here, so be gentle. :)

I am trying to create my own automated trading system. I'd like to bounce some ideas with you guys, if possible.

A few questions that came to my mind while implementing this (and my answers):

1. How frequent should the system try to trade? I chose once or twice per month since I have 0 experience with this stuff. Basically I want a system to predict the price so I can execute trades myself in a semi-automated way.

2. What should the system predict? I chose to try to predict exact price instead of arrows (whether it will go up or down or flat).

3. What should it trade? For now, just one high-volume security.

4. What should it use to "train" or "learn"? I chose a random period in daily price action of a security.

5. How to measure success of the prediction (when used in training/validation)? I used sum-of-absolute differences.

Attached is a figure that shows my results. Blue is the actual price. Red is the price this system predicted 25 days in advance.

Notice how the red price closely follows the blue price for an interval in the middle. That was actually used to train the system.

It performs rather poorly in the other regions (where I actually want it to succeed, since those are the regions where it will actually execute trades).

I want to know what the "experts" around here think of how I answered the above questions. Also, if you do design/build systems to predict exact price action, how much is it usually off by?

I guess if it is not off by much, you would be a billionaire sitting on a yacht right now and wouldn't be bothered to read this forum...

I'd like to know how accurately people can predict price action if they do use it. I know some people predict direction, not price.
 
Here is the chart.

Blue=actual, red=predicted.

Notice the curve fitting in the middle region that was used to train the system.

Comments? Too much curve fitting?
 
You should break your post up into smaller and more manageable pieces :), posting one question at a time and thoroughly read and reply :)
 
Your questions are not fundamental.

It sounds like your are confusing the difference between modeling market action and trading on the predicted results of the model.

vs.

Modeling a trading system and back testing against actual market data.

Which is it?
 
Quote from chawwal:

Here is the chart.

Blue=actual, red=predicted.

Notice the curve fitting in the middle region that was used to train the system.

Comments? Too much curve fitting?
More...

No it's not :D

Besides that.. if you are asking yourself such questions as "too much curvefitting", then YES! You are curvefitting, a good advice on the road.. avoid curvefitting, optimization etc. The more lagging indicators you are using, the more curvefitting is involved and the more illusions you have regarding the holy path to wealth.

You are being tricked.... :eek:
 
Quote from chawwal:

First post here, so be gentle. :)

I am trying to create my own automated trading system. I'd like to bounce some ideas with you guys, if possible.

A few questions that came to my mind while implementing this (and my answers):

1. How frequent should the system try to trade? I chose once or twice per month since I have 0 experience with this stuff. Basically I want a system to predict the price so I can execute trades myself in a semi-automated way.

The market makes that decision. I like up to 40 times a day, for example.[/colorr]. Profits come in segments and you can reverse at the end of each segment.

2. What should the system predict? I chose to try to predict exact price instead of arrows (whether it will go up or down or flat).

You can simply replace prediction with a binary indicator for ,aking money or going to the corrct side of the market. Why would you be taking the risk of predicting when it is not necessary

3. What should it trade? For now, just one high-volume security.

Try price change. it works in all markets. You can rank the rate of retrun for any instrument in the world. Stocks do 100 turns per year @ 10%. You can begin with 1,000 dollars and have 1,024,000 dollars in 100 turns.. This is beginner stuff but then you can improve and go to instruments that operate at higher money velocities. The ES is 50 times greater than high Beta stocks.

4. What should it use to "train" or "learn"? I chose a random period in daily price action of a security.

Train bar to bar. There are 10 cases. As you will see only one pattern results. No other pattern is possible. Set your trainer to "deduction"; DO NOT USE INDUCTION EVER AT ALL

5. How to measure success of the prediction (when used in training/validation)? I used sum-of-absolute differences.

Lets say 1, 2, 3, and 4 are working. for stocks your sharpe ratio is over 60. Other things get better. What replaced predictiction was your discover that the pattern you found was composed of an invariant oreder of events. Knowing this, you always "know that you know" and you go to a state of antivcipation in your coding of your ATS.

Attached is a figure that shows my results. Blue is the actual price. Red is the price this system predicted 25 days in advance.

Notice how the red price closely follows the blue price for an interval in the middle. That was actually used to train the system.

It performs rather poorly in the other regions (where I actually want it to succeed, since those are the regions where it will actually execute trades).

These three paragraphs allow you to conclude nothing works for you. Great conclusion and you reached it on your own

I want to know what the "experts" around here think of how I answered the above questions. Also, if you do design/build systems to predict exact price action, how much is it usually off by?

The performance of prediction is well known. Google: financial industry.

I guess if it is not off by much, you would be a billionaire sitting on a yacht right now and wouldn't be bothered to read this forum...

"Yachting" as you see it is underrated. Sitting is not what skilled sailors do. They race and get their names on things. Google US122 (King's Cruiser) for example. The name is "The Consort". any trader can purchase a ship over a two year period (interesst only) then pay off the loan and have an eaqual amount of capital as the original value left over , i. e., all sailing ships are free.

I'd like to know how accurately people can predict price action if they do use it. Since volume leads price, it is possible to arrive at any reversal price quite accurately just before the time comes. Keeping tabs on this contuing success is not difficult: logs are used. I know some people predict direction, not price. It is a sport called gambling.
More...
 
Quote from chawwal:

A few questions that came to my mind while implementing this (and my answers):
More...

1. Good luck but keep in mind that it will take some time.

2. You can try to predict exact price although "prediction" is a world used by astrologers and card readers these days. A better world is "forecast" but most traders do not like that either. Most traders neither try to predict nor to forecast. They use probabilities. This is because the market is not a deterministic system. It is highly stochastic, with high unpredictability.

3. Better to start with mini futures contracts. Stocks sound easy but there are many caveats.

4. Are you using a NN?

5. In this methods, small deviations are penalized as much as large deviations. I would use squared deviations.
 
You must log in or register to reply here.
Back
Top