Hi

How does anybody analyse probabilistic data?

Is this done on excel or can it be done on excel?

There are at least 30 probabilistic conditions to be analysed .How complex is the task?

Here is an example"if price fails to go lower by x pips over X days ,how much is the highest it travels on rebounds upwards ,on average.What is the succes rate?

This is just on of many conditions to be tested?

Anybody good at this?

OILFXPRO

For a quick look see, I export OHLC or indicator data out of Metatrader using a simple custom indicator that writes stuff to a text file. Then I import that data into Excel, or whatever package I'm using. Excels got all the basic stuff that you'll probably need, and if it doesn't, you can always hack something in VBA.

If your looking for a free stats package then take a look at The R Project for Statistical Computing The charting options are a bit better than Excels, and can be useful for visualizing some of the more complex stuff.

Some stats I find so useful, that I want to have them in real time, or at least a daily update, so in these cases, I'll code the analysis routine into an indicator so we don't need to be messing about exporting and importing stuff.

There's also a couple of good Excel TA plug in's that you can use if you cant be bothered exporting or coding indicator values. Quite a few of these are free.

I use stats a lot, and 85% of what I need is covered by Excel, and a bit of VBA programming.

I have been thinking of programming exactly this kind of a program. The question is where to get market data for about every possible pair and equity because if you use only few, you are not going to get meaningful results.

I have been thinking about doing this as a web service that would continuously calculate probabilities with bayesian probabilities but we'll see.

I have much more confidence in probabilistic values and genetic algorithms (that you can see how they work) than NN's. I have no idea why everyone is fussing about them - well except Betters performance (which I unfortunately believe to be a lucky coincidence).

My basic idea is for everyone interested:
- Get a huge list of market data including futures, stocks and spot equities if possible as comparison group
- Create a random, genetic algorithm optimizaed list of meaningful market movements based on both simple price information and indicators like "price breaking donchian channels X", "ma X crossing ma Y", "volume is over 150% of usual", "price has gone up over 20% in the last 2 hours", "time is 08:00"
- Create a random list of actions that could happen like "price goes up 20%", "price goes up over 30 pips", "volume increases", "high is over 30% higher than close" etc
- Calculate simple probabilities for X period for ALL pairs and equities I can find sufficient data
- Calculate combined probabilities for all conditions using bayesian probability theory

.. And the best of all, automatically let the program generate EAs based on these probabilities

I'm one of those guys who believe that predicting market movement for longer periods is kind of impossible but predicting the future movement for shorter period [edit: meaning bar/couple of bars] based on past probabilities is very possible.

edit: And if I wouldn't be programmer, I would probably use mathematica (I like the intreractivity and visual aspect)
market - The Wolfram Demonstrations Project

Hi

So developers will wait for 6 months to 2 years of forward tests before blowing their real accounts?

So they got time to waste?

they are better off scamming newbees indirectly

oilfxpro

Sorry to say but I have no idea what you are trying to say if this was directed to me.

Probability theory is very simple. The real problem is that calculating all permutations takes time and finding the combinations that really do have some correlation is quite problematic.

Forward testing (if you are talking about the genetic algo part of my idea) is not really needed if you separate the history to two sections where the other is the evaluation group and other is the validation group as is usually done in situations where there is a possibility of curve fitting.

By the way I already have some results for people interested, following are very simple correlation one rule combinations for couple of months of EURUSD minute bars, enjoy .

DISCLAIMER: the program still probably has bugs so the results may not be accurate!

The format is: {[rules], result, probability%, hits}

Mikkom

This is not directed at you.

This is general view on developing the better NN EA.It can't be backtested.

The developers will spend months devising the nn system,after which they will spend another 6 moths forward testing the expert advisor , they may have upto 10 versions of the ea before they finish,if they finish.

After 6 months of forward testing and a year of development ,it could still blow the accounts ,if mistakes are made.

Ask any free PROFESSIONAL programmer "do u want all this?".The answer is no

OILFXPRO

Mikkom

I have completed writing a sketch probability EXPERT ADVISOR.It can be back tested on metatrader ,provided the coder codes it correctly.It can be coded for easy testing.

It uses a probability master in the EA.Back test using true/false on every
probabilitic logic

Will this work?

OILFXPRO

Simply said: it will work if the rules are correct.

And more complex answer: It is really hard to know which rules are correct and which ones are due to curve fitting, that's why I'm trying to gather a really large amount of data so the correlations can be tested with at least some confidence that the results are correct.

If you can get a probability% instead of boolean value (should be possible), I would probably require a certain treshold level (like 60% or something, just an example) and then use different lot purchases based on the certainty level. Higher the certainty, higher the bet.

Only a professional developer of EXPERT ADVISORS can get all the above correct ,and like me they do not curve fit anything and always know the correct rules.Everything is normally taken care of

Should we go for NN or probabilistic Expert Advisor?NN can be covered by
probabilistic logic and coding and system ,so there is no need for a specific NN .I M H O

OILFXPRO

I'm not sure why you say it cant be back tested. Ive not followed betters EA particularly closely, but my understanding is this thing uses some fairly basic TA and a couple of baysian classifiers. I cant see the problem why this cant be back tested.

The reason people wont spend 18 months developing something is because they have absolutely zero self belief either in themselves or the people they work with. You wont change that, and to be honest, they aren't the type of people you need to be working with anyway so its not a problem is it ?