SOM: cooking methods - page 6

New comment
 
The SOM is fundamentally incapable of correctly classifying the state of the markets
 
Debugger:
The SOM is fundamentally incapable of correctly classifying the state of the markets
On what basis is this statement based?
 
Debugger:
SOM is fundamentally incapable of correctly classifying market states
SOM does not classify any states, it does not classify at all. SOM is used to quantise data, in this case time series.
 
Debugger:
The SOM is fundamentally incapable of correctly classifying the state of the markets
Right. That's why no one does it here.
 
alexeymosc:

Thanks for the good advice.

I'm posting the best result I got. The maximum number of trades has even increased. SL = 80 pips. The drawdown decreased many times, net profit increased. If I select a deposit twice as big as the maximum drawdown, we have 395% of profit for 10 years, and not 100%, as Sych had predicted...

I still want to improve my machine. Thanks for the constructive criticism.

The result in the tester is good, but I can't say I'm happy for you yet. I still have doubts (questions, tips, suggestions).

You got the simplest strategy - buy and hold (imho). The network seems to have nothing to do with it. It looks like a fitting of history.

1). Does the net know that you put a stop at 80 pips?

2). Can you do the same procedure only with short positions, on the same pair and the same period?

 
Sych:

The result in the tester is good, but I can't yet say I'm happy for you. There are still doubts (questions, hints, wishes).

You have got the simplest strategy - buy and hold (imho). The network seems to have nothing to do with it. It looks like a fitting of history.

1). Does the net know that you put a stop at 80 pips?

2). Can you do the same procedure only with short positions, on the same pair and the same period?

Doubt is your forte...

I see there is a misunderstanding of what I am doing...

The strategy is not buy and hold (although I started this thread with it). Exiting the market now on a SOM signal. And when it gives an exit depends solely on market behaviour. Some trades hang for days, others for hours. I have a more complicated strategy in my head.

What does fitting have to do with it? I look at the period of OoS, solely by it I decide on the value of TC.

In general: signals to enter and exit the market from the network. The network was trained on history. Do you think that a neuronet is a mega-brain that will give out something adequate without learning and testing on the history? Ridiculous. A human being works with any neuronet. A human is a link between any neuronet and the market. The entire TS comes from the head, not from the neuronet. The network does not know about stops, SOM simply converts the price series to a two-dimensional element, I do all the other analysis. I can also introduce takei, if I see a good result on a long history. I have to improvise.

I did not manage to get profit on short positions on the OoS period, I tried different variants. I did not succeed with this strategy on hourly bars.

 
alexeymosc:

Doubt is your forte.

If I say "great result, faster to the real", would that make you feel better?

I can see there's a misunderstanding of what I'm doing...

Maybe,

I agree further - a neural network is just a tool for solving clearly defined tasks, the whole TC is from the head, not from the neural network.

But in this case (imho) there is a discrepancy between set task and result obtained.

I'm also interested in neural networks, otherwise I would have passed it by.

I did not manage to get profit on short positions at period OoS, I tried different variants. I tried different variants, I did not manage to make profit on OoS period with short positions. I cannot make profit with this strategy on Hour Bars.

It confirms my doubts.

Is it possible to get decent results at least on the learning period?

 

--- But in this case (imho) there is a discrepancy between the task and the result.

I disagree. I deliberately trained the self-organising map on price trajectories normalised in the range 0 to 1, that is roughly how one sees the price chart in MT. It is always scaled from maximum to minimum in the window. And then I just made up my mind, the discussion on the forum helped me. The signals from SOM are in fact neural network indicator values, and I am looking for the best combination of them. At the same time, these very signals are not abstract for my perception, on the contrary, I know exactly that they show the state of the price series data for the last time (48 hours in this case) - it can be found out by doing a graphical analysis. And if we open when the price goes up, the profit is given by the signal, when the price changes its direction. Like this...


I understood that you are interested in direct strategy training, for example in learning how to reach take or predict level of stops and so on. You need to train with a teacher, e.g. MLP grid, but the task is not easy... I've taught it, tried it, but there's a problem - drawdowns. And going deeper, one of the neural nets is trained to open positions at oversold points (for buying) and overbought points (for selling). I found it out using glasometric analysis - I visualized the history of trades in the tester. It seems to have logic in operation, but the problem remains the same - if the price is oversold but goes down further anyway, the grid continues to open positions gladly. Stops by themselves do not solve the problem. The first one is to filter the signal with some indicator (but here I can cheat myself and fit it tightly to the history). The second - to feed the network with additional information during learning. The minus - a neuronet by its nature trains the simplest things that it may learn (for example, if you teach NS that there will be trades at crossing of muwings, but not always, the network will learn to open trades at crossing every time, because this information lies on the surface, and it will not learn more complicated patterns - again, with standard training methods). If you give her a lot of different information, it creates inconsistent examples, and inconsistent information impairs her learnability in principle.

--- At least on a test period, are you getting decent results?

Not yet. )

 

I'll try to elaborate if I can.

1). Regarding (does the network know that you set a stop of 80p?). Why do you look whether the price will be higher or lower on the fifth bar and not within 5 bars. Hence the large drawdowns. I.e. the network learns with unlimited deposit and unlimited drawdowns. It turns out that if the price is 10 points higher in a couple of bars, and the price was minus 150 points before the fifth bar, you consider this a positive result.

I suggest something like this: (Open [t+5] - 0.0010 > Open [t] && Low[iLowest(NULL,0,MODE_LOW,t+5,t)] > StopLoss) 1 else 0.

Or you have some other variants, there are a lot of them and we can experiment a bit. In the first report submitted by you we can clearly see what the network has learned, while in the second one, the learning result has little effect on the result in the tester (if it has any effect at all).

2). Why did you disable short positions. As you said: Straggy - the simplest, buy and hold. It is clear that the EUR has risen most of the time in the past 10 years, you do not need to invent a neuronet, you can use a simple Expert Advisor and disable short positions and the result will be as good as with a neuronet. I suggest to enable short positions and let the network learn and decide when to buy or sell. Who knows which way the EUR will go in the next 10 years. And if it goes down?

3). To really see what the network is capable of doing, the Expert Advisor should also work according to this strategy, accurately reflect the strategy. I.e. to eliminate the mismatch between the set task and the obtained result. If the net is trained to close a position after 5 bars, I suggest closing the position in the Expert Advisor after 5 bars as well, regardless of the result. Then we will clearly see if the network is able to do something or not.
 

Thanks for the comments. I'll try something and answer in more detail.

12345678
New comment