I have talked about many ML algos, some using MLE/MSE but many using other outputs. The latest NN algos do not use the MLE/MSE they use the hypothetical result of trades taken with a trailing stop. Read those posts again if you want to get a better ...

Machine Learning with algoTraderJo

Those differences seem acceptable to me. Don't they? What feeds did you use? There is a feed dependency you cannot escape but it never makes a long term difference in my live trading, even at the 30M timeframe. Just make sure you use some metric ...

Machine Learning with algoTraderJo

8 means 8AM GMT+1/+2. This is actually before the European open which is at 10 AM GMT +1/+2 so a very relevant time for the EUR/USD. Also 20 GMT +1/+2 is not close to the London or NY closing times. London closes at 18 GMT +1/+2 and NY closes at 23 ...

Machine Learning with algoTraderJo

Sometimes the ML algo will output long>0 and short>0 or long<0 and short<0 in that case you do not enter a trade. You also do not predict the outcome of the trailing stop, you predict the outcome of the trade that uses a given trailing stop. Since ...

Machine Learning with algoTraderJo

Your results are bound to be worse because of the distortions you are introducing. When you train the NN all your training examples must be built equally if you have problems with the hitting of targets then increase the horizon further but don't ...

Machine Learning with algoTraderJo

What does it mean that it "answers better" ? Be careful how you measure NN performance! If you are measuring NN performance by NN accuracy then evidently the close values will perform better because the series of close prices is heavily ...

Machine Learning with algoTraderJo

Yes you could try that but then you would introduce a difference in what is being predicted. The idea is to predict the trade outcome when using a trailing stop, if your output sometimes mismatches this and is instead some other value then you're ...

Machine Learning with algoTraderJo

Another parameter is the initialization of the random weights when training the NN algo. If you're a good observer and you read the last code I posted you might have noticed that weights in the NN are initialized to random values between -0.5 and ...

Machine Learning with algoTraderJo

I tend to agree with that. Even if you have some improvement in Sharpe ratio with more training it's not worth the additional time because the DMB analysis would be impractical. With the NN training 50 iterations it takes about 2x the time of a ...

Machine Learning with algoTraderJo

We can also pay a 10x penalty in simulation speed and train for 1000 iterations per NN training, before every trading decision. This would heavily impair DMB analysis. Look how it affects the last posted system. image Is it worth it? You tell me ...

Machine Learning with algoTraderJo

This course is actually taken "on your own time" now, also offline with all the lectures available when you want to watch them. So don't worry about it mate, you can just join and go through the lectures, exercises and homework as you see fit. I ...

Machine Learning with algoTraderJo

Another case on the EUR/USD where the number of returns used to build each example (C parameter) is 5 instead of 2. As in every case retraining of the NN is done before each trading decision. SL = 160% ATR20 A = 1 GMT +1/+2 B = 90 C = 5 D = 90 ...

Machine Learning with algoTraderJo

These are some initial sample EUR/USD results using the before mentioned NN function for predictions, the trailing stop as explained before and the input and output structure used for previous linear regression 1H strategies. As in all cases the ML ...

Machine Learning with algoTraderJo

Daniel's blog can be a bit overwhelming to the beginner. Try quantstart as well, the guy is more organized than Daniel and has lots of insightful articles. I also fully agree with beachboy44, binary options are terrible IMHO.

Machine Learning with algoTraderJo

I now want to talk about neural networks. Up to this point we have explored simpler machine learning algos like linear regression and we have only briefly explored SVMs, NNs and random forests (when we were doing daily data experiments if you ...

Machine Learning with algoTraderJo

OMG definitely 100,000 lots is a bit unrealistic When doing these portfolios I always like to keep total risk at 1-2% to give some more realistic simulations. But in any case the Sharpe ratio and other risk adjusted statistics should be very ...

Machine Learning with algoTraderJo

This is interesting. I have tried indexes as inputs before but I have never managed to get them to work better than simple returns. I think Daniel has posted on the Asirikuy forum with a similar experience when using indexes. However I didn't do the ...

Machine Learning with algoTraderJo

I am impressed by all the progress in the cloud enhanced machine learning in Asirikuy. Daniel recently released another 20 or so low DMB systems from these mining experiments. The portfolio has 6 systems on the USD/JPY and 28 on the EUR/USD, all ...

Machine Learning with algoTraderJo

You need both to happen. Long prediction > Short Prediction & Long prediction > 0 for a long entry and viceversa for shorts

Machine Learning with algoTraderJo

You can probably go down to 15M before any of these problems become too prevalent with the Asirikuy data. Providing your SL and TP (if any) values are large enough then you should not face any significant issues until you go to something like 5M, ...

Machine Learning with algoTraderJo