SOM: cooking methods - page 2
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Good afternoon!
Continuing the theme. Made an analysis for GBPUSD pair on daily bars, from 1989 to 2011. Same approach, but I made a smaller BVS (5*5), therefore the input vectors separation became coarser, but nothing...
This is the result of learning the ACS. I didn't divide it into clusters, so as not to make it even coarser. I took 5800 examples for training, the size of input vector is the same - 40.
I analyzed probability of price movement in the future with a variable number of bars (from 1 to 15 in future). Horizontally go numbers of neurons, then number of samples that hit them, then probabilities for 1-15 bars in the future, first for upward moves, then block for downward price moves.
Graph by this table. Only here on the horizontal bars to the future, on the vertical - neurons. I have taken exactly those cases that are highlighted in purple to build a trading strategy. Probabilities modulo greater than 0.6.
Results. Training:
Next I feed the data from the OOS-period to the trained SOM and get the neuron numbers. I apply the strategy.
OOS-period (from the beginning of 2010 to present).
And lastly, I've built an average input vector corresponding to the examples from the cell where the probabilities are the highest:
I can also post the file with all the data on request.
Predictability is observed, on all the tools I have tested.
Great, I would even say great! - or, you know, casino, randomness....
Now, if you don't mind, and if it's possible within your neural network technology, try predicting without using the trading forecast time limit (number of bars into the future). I'm very curious if the same predictive ability of the grid will persist.
Great, I would even say great! - You know,casino,random....
Now, if you don't mind, and if it's possible within your neural network technology, try to make predictions without using time-limit trading forecast (number of bars into the future). I'm very curious if the grid will retain the same predictive ability.
I think there is some misunderstanding of the algorithm. The self-organising map does not predict.... It splits the multidimensional example space into compact areas where similar examples are concentrated. The neural network doesn't know what the future holds. (Although sometimes future-looking data is fed to train the ACS and then it can focus on clusters where profit is maximal, while learning input vectors that precede such profitable situations). I then build a spreadsheet and look at the average future price behaviour for the cases clustered by the neural network.
And how do you make a prediction without a timeline? We predict the future but not infinity.
As far as I understand, the table is based on fact, so with no time limit you have to come up with another way to build the table.
Yes, of course. You could try to predict takeprofit achievement, but again you need to set a time limit, at least some.
We can also look not at the probability that the price will be higher or lower in n bars, but at the average profit from transactions, given the duration of an open transaction of n bars. And all of this can be looked at on the already existing breakdown of the examples into SCP cells.
I'd like to hear some ideas of what else can be predicted with this approach.
I think there is some misunderstanding of the algorithm. A self-organising map doesn't predict.... It splits the multidimensional space of examples into compact regions where similar examples are concentrated. The neural network doesn't know what is going to happen in the future.
I understand how self-organizing maps work (apparently, you do not understand why I asked the question, but if you do not know, I will not explain, otherwise they may accuse me of self-praising). I am not talking about the cards (what the cards predict), but about TC in general. And TC is engaged precisely in prognosis, no matter who says what.
And how can we make predictions without having a timeline? We predict the future but not infinity.
Yikes. Percentage of 99% of all TCs that are mused on this forum don't have a time frame for prediction. It seems strange to me (and maybe to you too), but it's true. A typical example: trade on two or one wave, enter by crossing (you enter but have no idea when exit will be, maybe never), exit/entry by opposite signal.
I'm glad there are people on this forum who understand the bolded meaning.
I would like to hear some ideas of what else can be predicted with this approach.
So are we predicting after all? :)
One can predict the probability (in general, I don't like the word "predict", and especially in combination with the word "probability") that price will stay in a given range at a given time interval (or conversely, that it will leave it) - one can make money on that too.
--- Uy. Percentage of 99% of all TCs that are mused on this forum do not have a time frame for prediction. It seems strange to me (and maybe to you too), but it does. A typical example: trade on two or one wave, enter by crossing (we enter but have no idea when exit will be, maybe never), exit/enter by opposite signal.
Yes, I understand that. If we compare it with TS where both inputs and outputs are generated by indicators, we can do the following: we feed NS with inputs, get neuron number, enter the position, wait until NS gives us another number of neuron to exit by. We select the numbers of input and output neurons in the strategy tester. For this, of course, we should write an Expert Advisor and test it. But I like to think first if this approach will be viable... I have been thinking that essentially we are dealing with transition of a system from one state to another. The states of the system are formalised as belonging to compact classes (SCS cells). Theoretically there may be situations, when transitions from state x to state y with high probability give profit... But it's just a fantasy so far ) What do you think?
--- You can forecast the probability (and I don't like the word "forecast" combined with "probability") that price will stay in a certain range at a given time interval (or leave it the other way around) - you can make money on that too.
Yes, of course, we make predictions, based on the results of analysis of the resulting groups of data, probabilistically ) Nothing can be said for sure, there will always be exceptions to the rule. I'm going to put up a data file, there are cell numbers and the OOS period is greyed out. You may also try to prognosticate yourself and make up a strategy in Excel on the training period, and then check it on OOS. For example, we analyse what maximum and minimum the price will reach in the future (by bars), but here we can say right away that the channel will extend in the future. And how to display TS in the positive mathematical expectation on this idea?
The file is attached.