One important thing IMHO is that you do things when backtesting exactly as you will do when live trading. I like F4 because you use the exact same code when doing both so it's impossible to introduce data snooping bias. By creating a dataset for ...

Machine Learning with algoTraderJo

Take a look at this example on the EURUSD 30M timeframe. I am using the last described trailing stop that involves no additional parameters. The machine learning algo is a linear regression which is trained after every trading decision, as I ...

Machine Learning with algoTraderJo

You just need to have a given number of examples in the training set, there is no rule indicating that such examples must be fully contiguous in the past. You can have 100 examples but you might have had to skip a few because they did not reach one ...

Machine Learning with algoTraderJo

Also redlion, the fact that you cannot negate a null hypothesis does not mean that your null hypothesis is true! That's a logical fallacy. If you cannot negate the null hypothesis that the market is random it does not make it random, it just means ...

Technical Analysis alone will not give you an edge

You have to love this guy. Exquisite example.

Technical Analysis alone will not give you an edge

Your tests definitely show that the market does not show linear correlations. The market indeed does not show first order autocorrelations and is virtually indistinguishable from random data when you look at simple statistical measurements like ...

Technical Analysis alone will not give you an edge

Ask yourself if you can perform a thorough data mining bias analysis using any of these techniques. Can you? If you can't, does it make sense to use them? If you could perform the bias measurement with any technique you wanted then your choice of ML ...

Machine Learning with algoTraderJo

Sorry I have deleted those trading results (I do tons of research and cannot keep everything that I post stored). If you want to ask questions about things I post please do so rather quickly so that I can answer while the data is fresh and had not ...

Machine Learning with algoTraderJo

That's a wrong understanding. The SL is moved so that it is always the same but it is only moved if it leads to a more favorable SL than the current one. So if you enter a trade with an SL of 100% of the ATR20 then you will only move the SL if at ...

Machine Learning with algoTraderJo

If you don't hit a target you skip that example, it is not a valid example if you don't hit both targets within the used horizon.

Machine Learning with algoTraderJo

Depends on what your outputs are. Different inputs can lead to better predictions on different outputs. Generally you will find it very hard to get any inputs with better performance than simple bar returns. In my own research technical indicators ...

Machine Learning with algoTraderJo

To avoid over fitting do two things. One, constantly retrain your algorithm before each trading decision using a moving window as shown by the systems on the thread. Two, use as much data as you can.

Machine Learning with algoTraderJo

That's a very lazy proposition man. If you want to learn the best system that has been posted then read the thread and compare system statistics. The thread establishes clear steps to allow you to create your own systems and get the best out of any ...

Machine Learning with algoTraderJo

You can post results for anything you want. Just try to be clear in what software you used and what markets/instruments are involved. To answer what you asked: Yes, all number are in absolute values. A value of 50 means a 50*100% total return. ...

Machine Learning with algoTraderJo

It's not an ensemble, in this example I am making predictions using rates from two different symbols on the input/output function. None of the functions that come with the F4 framework can do this, but you can modify them to do so if you wish. I'll ...

Machine Learning with algoTraderJo

Going down to the 30M timeframe, compounding starts to get really serious. Using the previously described trailing stop based system on the 30M TF we can get the following result (remember that the ML algo is retrained before making each trading ...

Machine Learning with algoTraderJo

By all means, try it on other markets! The techniques I have discussed on this thread can be applied to the equity, bond or futures markets without problems. There are however a few reasons why you might want to try FX first: Intraday data for ...

Machine Learning with algoTraderJo

Another way to attempt to improve your results is to include data from other symbols as inputs. You must be really careful when doing this as you must make sure that the two data sets match perfectly so that you don't get any wrong results due to ...

Machine Learning with algoTraderJo

That's cool.

Machine Learning with algoTraderJo

Random walks also don't work well to generate random data for this purpose... The data must be exactly the same as your real data except that it should be shuffled to eliminate any potential for real predictive action to happen. This means use ...

Machine Learning with algoTraderJo