Machine learning in trading: theory, models, practice and algo-trading - page 1336

[Ivan Negreshniy]

Maxim Dmitrievsky:

I don't understand the thought... rearrange the charts mentally and everything will be fine

I do so because I'm too lazy to rewind the charts in the tester all the time, there is no difference and it's even better, the training is up to date instead of 3 years ago

The point is that the connection of future prices with current prices is logically performed through the fact that all trading is planned on the basis of current prices, which in turn affects future prices, i.e. the argument-function connection goes from left to right, but not vice versa.

[Ivan Negreshniy]

Aleksey Vyazmikin:

Why? Not only the stakes are interesting, but also the rationale.

answered above

[Ivan Negreshniy]

Maxim Dmitrievsky:

No, I have a Markov process, the solution depends only on the current state, there is no planning when learning in the form as in classical q-learning, for example

there is nothing to learn in the tester - go straight to realtime:)

but if you use a tester, then do everything in one pass

[Yuriy Asaulenko]

Ivan Negreshniy:
The matter is that the connection of future prices with current prices is logically performed through the fact that all trading is planned on the basis of current prices, which in turn affects future prices, i.e. the argument-function connection goes from left to right and not vice versa.

He has short-term deals, judging by the chart. Nothing will change from shifting. Even better, because training on the latest data, there is a probability that the real will be better than when training long before the real.

[Ivan Negreshniy]

Maxim Dmitrievsky:

I'm already racing a little on the exchange, feeling it out.

Well, maybe on the exchange, where the agent can influence the process with his actions there in the glass, but here in Forex, where there is no influence and treat it as a Markov...

[Yuriy Asaulenko]

Maxim Dmitrievsky:

There is a unique development of bots with artificial intelligence, which will conquer not only the market, but also the whole world.

Napoleonic plans. Let me remind you, Napoleon finished on St. Helena.

[Forester]

Yuriy Asaulenko:

Napoleonic plans. Recall that Napoleon ended up on St. Helena.

Everyone ends up somewhere.

[Женя]

Maxim Dmitrievsky:

This is a bicycle for you, but for others a kernelized SVM (about "Reshetov's Vector Machine")

I doubt that"kernelized", its output is linear, dumb hyperplane, It's not SVM either, personally I haven't analyzed in details Java code that Yury had posted on cyberforum before he disappeared, but one "graale-writer" I know had this code analyzed for a client to tweak it and said that it was some kind of custom stochastic search of linear model coefficients, some kind of genetics or ostracism, the result was a vector of linear coefficients, in short, a more advanced version of this masterpiece:

AI Expert Advisor for MetaTrader 4

double perceptron() 
  {
   double w1 = x1 - 100;
   double w2 = x2 - 100;
   double w3 = x3 - 100;
   double w4 = x4 - 100;
   double a1 = iAC(Symbol(), 0, 0);
   double a2 = iAC(Symbol(), 0, 7);
   double a3 = iAC(Symbol(), 0, 14);
   double a4 = iAC(Symbol(), 0, 21);
   return(w1 * a1 + w2 * a2 + w3 * a3 + w4 * a4);
  }

In short, it's a more advanced version of this masterpiece: AI Expert Advisor in MetaTrader 4.

[Aleksey Vyazmikin]

Finished processing the models, seed from 1 to 200 - I swapped the training sample into validation sample and vice versa. In the table I kept the logic, i.e. where there was 10% became 90% and vice versa - this is because of the swap of samples.

Table with balance score results

Table with metric indicators

Table with the number of models that meet selection criteria in the independent sampling

Table with the number of models corresponding to the selection criterion in all three samples

Few models were selected - it is not clear why it happened so - we have to think.

Graphs of the models (I didn't make figures because of the small number of selected models)

60%

70%

What can be concluded here - the sampling ratio again turned out to be over 30% (the old samples of 70%, but swapped places!). The graphs have at first glance more differences, but still not enough for an effective parallel run, to cover a larger percentage of completeness (Recall).

[Aleksey Vyazmikin]

And this is how you can combine the two models from the last and penultimate experiment - though the separation was strengthened to 0.55. Entry on the principle if one of the models has a signal, then enter the market.