Machine learning in trading: theory, models, practice and algo-trading - page 260

 
Dr.Trader:

It will be very difficult to determine the winner. Someone will martingale and win with money rather than accuracy of the model, and argue about winning with someone who seems to win with accuracy. Someone will just look at real prices and post the result late. Models are unlikely to be posted by anyone, so there is no way to believe or verify.

It is much easier to make your own signal and show results on it.

Theoretically the signals are good, but there are many "buts". First of all it is a MT theme, few brokers offer MT as a terminal for exchange trading. I need to get an honest flow of requests and deals to analyze, and Forex trading with brokerage companies is... you know)) You can check it in the same way as in numerai just for logloss or accuracy of direction predictions, it is enough. As a measure of the strategy portfolio success take the Sharpe Ratio or its robust analogues, there martingale and other absurd MM algorithms will not work. I just offered an idea. Then there will be a basis for discussion, at the same time we will discuss data, attributes, target functions, etc. on specific examples.
 
toxic:
Of course signals are good in theory, but there are many "buts", first of all it is a MT theme, few brokers offer MT as a terminal for exchange trading. I need to get an honest flow of requests and deals to analyze, and Forex via brokerage companies is... you know)) You can check it in the same way as in numerai just for logloss or accuracy of direction predictions, it is enough. As a measure of the strategy portfolio success take the Sharpe Ratio or its robust analogues, there martingale and other absurd MM algorithms will not work. I just offered an idea. Then there will be a basis for discussion, at the same time we will discuss data, attributes, target functions, etc. on specific examples.

Put your data, we'll predict everything we can :)

But in R trade simulation problems, I can only give a forecast for each bar - buy/sell, and then find accuracy. Sharpe ratios I can't get.

 
toxic:
I need to get an honest stream of requests and trades to analyze, and the forex via DCs is... you know))

We understand... )

Here guys have been simulating MM algorithms with market data for a long timehttp://www.quantalgos.ru/?p=1372 , and even with R , quantstrat packagehttps://r-forge.r-project.org/scm/viewvc.php/*checkout*/pkg/quantstrat/sandbox/QuantstratWorkshop.pdf?root=blotter

And in general an interesting site to readhttp://www.quantalgos.ru/

Dr.Trader:

But R has problems with simulated trading

Again, pay attention toquantstrat

It has things which even MT couldn't dream of, like that walk forward test.

Еще одно тестирование алгоритма Маркет Мэйкера | QuantAlgos
  • 2015.07.10
  • www.quantalgos.ru
Продолжая тему тестирования алгоритма Маркет Мэйкера, поделюсь своими результатами и мыслями по его работе: 1. Основной режим работы алгоритма - это маркетмэйкинг (он же арбитраж ликвидности, он же торговля спредом). И конечно же, прибыльность этой стратегии сильно зависит от рыночных условий, скорости получения данных и работы системы...
 
mytarmailS:

There are things in it that even MT couldn't dream of.

I saw there in the description the history of ohlc. This is even less than MT4 can do.
I can plot equity on ohlc values without quantstrat, but it is very weak.

I need an accurate simulation of ticks within a bar, with correct working out of stops and takes, spread widening, slippages, otherwise all these calculations of profitfactor, sharperatio, etc. on naked ohlc lose accuracy and mislead. MT5 for forex just allows you to test everything very accurately.
Valc forward in mt5 is also possible (in semi-automatic mode), with manual launching of each new step on new dates.

 

I read a curious set of predictors. Especially the last one.

Thetarget variable is a 1 week price prediction, which can take two states - UP, if the price for the week went up, and DOWN, if the price for the week went down.

The following indicators are taken as explanatory variables and take only two values:

  • Volatility (VAR1). High volatility is usually typical of falling markets, while low volatility is typical of rising markets. This variable is based on the value of the ATR indicator compared to its moving average MA with a period of 20 days. If ATR>MA, VAR1=1, if ATR<MA, VAR1=-1.
  • Short price impulse (VAR2). Here the price value of the asset is compared to its simple moving average SMA with a period of 5 days. If price>SMA, VAR2=1, otherwise VAR2=-1.
  • Long price impulse (VAR3). For this variable the period of SMA is taken as 50 days. If price>SMA, VAR3=1, otherwise VAR3=-1.
  • Pivot (VAR4). In this case the value of indicator CRTDR (Close Relative To Daily Range) is used, which is calculated as follows: <span class="MathJax" id="MathJax-Element-1-Frame" tabindex="0" data-mathml="CRTDR=Close&#x2212;LowHigh&#x2212;Low" role="presentation" style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline; display: inline; line-height: normal; word-spacing: normal; word-wrap: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0px; min-height: 0px; position: relative;">CRTDR=Close-LowHigh-Low CRTDR=Close-LowHigh-Low, where Close is the closing price of the day, High is the maximum price of the day, Low is the minimum price of the day. If CRTDR>0.5, VAR4=1, otherwise VAR4=-1.
  • Mode of autocorrelation (VAR5). The market has two modes of autocorrelation of price increments. In one of them the autocorrelation is positive, in the other one it is negative. If autocorrelation of increments for last 5 days is more than 0, then VAR5=1, otherwise VAR5=-1.
Использование CART в предсказании направления рынка | QuantAlgos
  • 2015.04.03
  • www.quantalgos.ru
Интересный подход к предсказанию направления  рынка рассмотрен в статье "Using CART for Stock Market Forecasting". Для того, чтобы предугадать движение цены на недельном отрезке используется техника под названием CART (Classification And Regression Trees) - построение классификационного графа (дерева) с целью предсказать значение  целевой...
 
Dr.Trader:

I saw the history of ohlc in the description. It is even less than MT4 is able to do.
I can plot equity on ohlc values without quantstrat, but it is very weak.

I need an accurate simulation of ticks within a bar, with correct working out of stops and takes, spread widening, slippages, otherwise all these calculations of profitfactor, sharperatio, etc. on bare ohlc lose accuracy and are misleading. MT5 for forex just allows you to test everything very accurately.
Valc forwards in mt5 is also possible (in semi-automatic mode), with manual triggering of each new step on new dates.

Look at the links I gave, there the man is testing exactly on ticks.

Naturally, in the introductory pdf examples will be with ohlc, but how else?

 
mytarmailS:

Look at the links that I gave, there the man is testing exactly on ticks.

The article is a heap of text and empty words, not a single code example, the quantstrat package is not even mentioned (and not used). The article is about nothing, for filling the site with content, its value = 0.

 
Dr.Trader:

Put your data, we'll predict everything we can :)

But in R trade simulation problems, I can only give a forecast for each bar - buy/sell, and then find accuracy. Sharpe ratios I can't get.

I propose second "slices" for our futures: bid, offer, tick average per second, delta in the pile, volume of buy and sell, open interest; for forex currency pairs: bid, offer, tick average; for foreign indexes: price and change per day, example in the attachment.


Files:
example0.zip  67 kb
 
Theproblem is that we need a collective consensus on the basic traits and targeting:

Gentlemen, what if we improvise something like numerai but with real data and compete?

I propose to discuss conditions. For example, we will use a moderate high-frequency in the Russian market, not hard MM, take the flow of prices, volumes, open interest for local liquid futures and prices of major currency pairs of forex and a set of liquid Western indices, futures, etc. Synchronized seconds prices. The data is publicly available.

Forecasting our futures. The holding horizon is minutes, but it depends on who wants, the thing is that there will be not much data, for example, a week and it will not be possible to train trading days.

The last day of the week is not known, it is necessary to teach the system to trade it profitably.

toxic:

It is necessary to determine the channels, everyone will himself signs to form, I propose the second "slices" for our futures: flippers per second bid, offer, average tick value per second, delta in the pile, the volume of purchases and sales, separately and change in open interest, for forex currency pairs flippers bid, offer, average tick, for foreign indices price and change per day example in the attachment.

What of the provided instruments to trade? What is the output? Signals, bids, probabilities of change direction?

The first problem is immediately obvious, if the data are open, then the results can be easily falsified by a peep, and if they are closed, then the indicators have to be created by the one who gives the data, and alien indicators may not be the most effective, it will be like a pig in a poke with https://numer.ai/. So at a minimum, we need a collective consensus on the basic signs and targets.

 
Zhenya:

And what of the provided tools to trade? What is the output? Signals, orders, probabilities of change direction?

The first problem is immediately obvious, if the data are open, then it is easy to falsify the result by peeking, and if they are closed, then we will have to create a sign that will give the data, and other people's signs may not be the most effective, it will be like a pig in a poke with https://numer.ai/. So at a minimum we need a collective consensus on the basic features and targets.

Everything is negotiable. I suggested forts futures, for example Si, RI, BR etc. the most liquid ones in general. As a result I propose a signal (-1,0,1) (short, cash, long), the signal is unambiguous than probability and is not distortedby MM as applications. Postprocessing, signs and targeting is a matter of the owner, or order.