Machine learning in trading: theory, models, practice and algo-trading - page 1544

[Maxim Dmitrievsky]

Grail:

I have tried many things, but IMHO it is a sad idea to focus the sign of the future increment, at least I never learned to do anything good with it, it is not about the subtleties of configuration, but about close to zero predictability, which is completely leveled by the trading costs.

Increments are "micro"-level, like movements of atoms, and it's necessary to focus something less noisy, something like trendiness/flatness, than to filter the usual TS of pullbacks and impulses.

so not necessarily one increment, then the costs do not play a big role

If I have acurasi model ~0.7 on validation, then what prevents me from adding random realizations of the process and see what happens. The model will find something in between, acuras will fall (because some examples will overlap), but generalization will increase (in theory).

That's my way of thinking :Z

we're not running after the grail, we're trying to understand randomness

[Igor Makanu]

Maxim Dmitrievsky:

How many pips or how long are the orders in the market on this TS?

I think it's a mistake to refuse the MT5 tester, it will show you the specific moments in the trade better.

I'm running the tester MT5 "along and across", I can confidently state that the quality of order execution has a significant influence on the result, half of the excellentoptimization results show results close to zero if I set the trade mode - arbitrary delay, but what will stand this test will work fine on the forward

[Maxim Dmitrievsky]

Igor Makanu:

How many pips or how long are the orders in the market on this TS?

I think it's a mistake to refuse to use the MT5 tester, it will show you the specific moments in the trade better.

i'm running the tester MT5 "along and across", i can confidently state that the quality of order execution strongly affects the result, half of the excellent optimization results show results close to zero if i set the trade mode - arbitrary delay, but what i can withstand in this test will work fine for forward

2500 trades for 20000 bars on the test screen

I'm not giving up on the tester, it just doesn't work with sockets. I'm too lazy to remake it for pips

any markups I can add to mine, the tester is very easy

[Igor Makanu]

Maxim Dmitrievsky:

2,500 trades on 20,000 bars on the test screen

Yeah, I saw, recalculated, seems to add up on M15 about 208 days of testing, ie a year

what i mean - i look at the screenshot, again the same grayablits as in the theory - 2-day trades, i.e. again waiting for the loss - if not, why take data from M15, and the TS may hang out for a couple of days in the market?

imho the drawdown and recovery factor should be looked at

SZY: pampering ....vot 35 000 orders per year, 2 screens - one screen optil year, the second screen a year forward, oddly enough the dynamics is the same, and the mode of arbitrary delays sustained this TS ... And no magic, optim-testing-optim-testing - tests on M1, switched to H1 to show how orders can close the chart ))))

[Maxim Dmitrievsky]

Igor Makanu:

Yeah, I saw, recalculated, seems to add up on M15 about 208 days tested, i.e. a year.

I'm looking at the screenshot, again the same grayablits as in the theory - 2 days trades, i.e. again overlapping losses - if not, why the data from M15 is taken, and the TS may hang out for a couple of days?

imho the drawdown and recovery factor should be looked at

SZY: pampering ....vot 35 000 orders per year, 2 screens - one screen optil year, the second screen a year forward, strangely enough the dynamics is the same, and the mode of arbitrary delays sustained this TS ... And no magic, optim-testing-optim-testing - tests on M1, switched to H1 to show how orders can close the chart ))))

I tried to notice, why days, I have an average of 2 hours for 15 minutes. Which is almost a hft)

as i wrote above i can make any number of deals, it's not a big deal

i have a fully working model the same on mt5 but with forest... and i wanted one with mother of pearl buttons and on python

something to do with your ticks, it won't work... maybe just a bug in the tester ahahaha) i used to be able to do this kind of stuff on mt4

Z.Y., I'm going back to what I've wanted to do for a long time - MO + stoch

http://www.turingfinance.com/random-walks-down-wall-street-stochastic-processes-in-python/

[mytarmailS]

The Grail:

tried many things, IMHO forecasting the sign of the future increase is a sad idea, well at least I did not so ........

Try to predict not a sign of increase, but for example the price of the next zigzag knee or something better, or for example the next maximum of 30 successive candles or something like that, then to use regression instead of classification, but regression not one step ahead, but to look for the extrema. I think you will be pleasantly surprised.

[Igor Makanu]

Maxim Dmitrievsky:

Something to do with your ticks, will not work... maybe just a bug tester ahahah) It used to be that on mt4 they were easy to do

when i started working with mt4 it was not ticks, but M1 bars and 10 TS that the optimizer sets up by himself - the most important thing here, as i've written many times before, is stoploops - they are there!

I haven't done any requotes in MT4... If I noticed that the main trick was MT4 did not show the green equi line until the orders were closed, that's why the "loss aversion" was going on - the TS was closing right away when the profit was taken, such a trick does not work in MT5 - it always draws the equi line, but as you see, I don't have "green snot" in my screenshots ;)

[Maxim Dmitrievsky]

Igor Makanu:

I have not done it with ticks, but with M1 bars and 10 TS, which the optimizer sets himself - the most valuable thing here is, as I wrote a hundred times before, the stoploops - they are there!

I haven't done any requotes in MT4... If I noticed that the main trick was MT4 did not show the green equi line until the orders were closed, that's why the "loss aversion" was going on - the TS was closing right away when the profit was taken, this trick does not work in MT5 - it always draws the equi line, but as you see, I don't have "green snot" on my screenshots ;)

Well, I need to add the demo to the monitoring immediately) so I don't have to guess

I don't think they were made on unreal ticks. If your stops or pending orders are executed by ticks then it may help to make such nice pictures too.

I'm doing pure econometrics, if I may put it that way. That is, I work with random processes.

[Алексей Тарабанов]

mytarmailS:

Try to predict not the sign of the increment, but for example the price of the next knee of the zigzag or something better, or for example the consequence of 30 successive candles or something like that. or something like that, you should use regression instead of classification, but regression not one step ahead, but look for the extremum. I think you will be pleasantly surprised

regression not one step ahead

I'll tell you a terrible secret, regression does not have steps forward.

[mytarmailS]

Alexei Tarabanov:

regression is not one step forward

I'll tell you a terrible secret, regression doesn't have steps forward.

??