Machine learning in trading: theory, models, practice and algo-trading - page 138

[СанСаныч Фоменко]

mytarmailS:
Thank you, I would not have figured it out.

It would be very interesting to see the result.

I used to do it, I even have the code somewhere.

1. The given code is not enough to make decisions. You need thresholds and floating thresholds.

2. It is possible to obtain very profitable systems with a profit factor of up to 10.

BUT.

Profit factor is less than 5 pips. Spread accounting kills all the beauty. And besides the spread, there are gifts.

[mytarmailS]

SanSanych Fomenko:

It would be very interesting to see the result.

The subject of arbitrage is off-topic here, but I'll answer briefly if you're interested...

I already had a working arbitrage, I wanted to improve it by filtering signals for input in such tricky way - in sliding window we consider cointegration strength (closeness of two BPs) and only when BPs are strongly cointegrated between each other we start arbitrage...

the result of such filtering is ambiguous, i can't classify it as good or bad

with the first variant the system yielded on average - 7% profit per month with a recovery factor of 3

the system with filtering gives 5.7% a month and 5.5 PV

it earns less, but it is more stable...

interesting links with examples on "R" in russian:

about correlation difference with cointegration

http://www.algorithmist.ru/2011/08/time-series-similarity-measures.html

http://www.algorithmist.ru/2011/09/time-series-test-for-cointegration.html

about pairing and arbitrage

http://rforfinance.ru/pairs-trade/

http://rforfinance.ru/rolling-window/

[Alexey Burnakov]

mytarmailS:

The topic of arbitrage is off-topic here, but I'll answer briefly if you're interested...

I already had a working arbitrage, I wanted to improve it by filtering signals to the input in such a simple way - in sliding window we count cointegration strength (closeness of two BPs) and only when BPs are strongly cointegrated we start arbitrage...

the result of such filtering is ambiguous, i can't classify it as good or bad

with the first variant the system yielded on average - 7% profit per month with a recovery factor of 3

the system with filtering gives 5.7% a month and 5.5 PV

it earns less, but it is more stable...

interesting links with examples on "R" in russian:

about correlation difference with cointegration

http://www.algorithmist.ru/2011/08/time-series-similarity-measures.html

http://www.algorithmist.ru/2011/09/time-series-test-for-cointegration.html

about pairing and arbitrage

http://rforfinance.ru/pairs-trade/

http://rforfinance.ru/rolling-window/

Thank you. The description on the first link is hilarious. Two alkonafts coming up.

[mytarmailS]

Alexey Burnakov:
Thank you. The first link smiled at the description. Two alkonafts coming.

:)

[Vladimir Perervenko]

Alexey Burnakov:

By the way, Mr. Perervenko did not say anything about this type of network in his article about neural networks. Well, I found only one mention in the whole article. And it would be possible to reveal the question of applicability to time series (thoughtfully).

Alexei

There will be a separate article(or two) about recursive neural networks (RNN, CNN and LSTM). Examples using mxnetR and maybe mxnet(Python).

By the way the mxnet package is present in the CRAN repository. Really need to do some gymnastics to add from GitHub optimizers (RMSProp, Adam, AdaGrad and AdaDelta). I'm testing these features right now.

Good luck

[Alexey Burnakov]

Vladimir Perervenko:

Okay, it will be interesting to read.

CNN is not a recursive network per se.

install.packages("drat", repos="https://cran.rstudio.com")
drat::: addRepo("dmlc")
install.packages("mxnet")

[СанСаныч Фоменко]

Under the influence of the article I read, a question arose:

Why are we all, including me, stuck with predictors derived from a single currency pair? There are a lot of currency pairs, and non-currency pairs are a dime a dozen...

[mytarmailS]

SanSanych Fomenko:

Under the influence of the article I read, a question arose:

Why are we all, including me, stuck with predictors derived from a single currency pair? There are a lot of currency pairs, and non-currency pairs are a dime a dozen...

What makes you think that "we all"? :)

[Dr. Trader]

For me it's all about the risk - I try to take at least a small risk. You may create an Expert Advisor that successfully trades a dozen pairs over many years, but what for? The profit will probably be a couple percent a year, you may as well put money in the bank on a risk-free deposit.
If you take just one pair and teach the model to trade well on it for one year you can expect profit of 10% a month, which is better.
If you teach the model for only 2 months of data - it will probably work successfully for a week, but it will bring these 10% for an even shorter period.

The narrower the data set (in terms of time and number of pairs) a model has for training - the more profitable it will be on new data. But it will become obsolete and will cease to be profitable much earlier. And it's much riskier, as some news item may break the entire model and the model parameters will have to be searched anew.

[Alexey Burnakov]

And for me, if I can't say in what range your trade will be in the future, it's already guessing by coffee grounds. It is necessary to know in advance where you will be approximately, and this knowledge should be objective.