Control points and open price should give you the exact same results if your systems is coded correctly in MT4. This is probably an issue with your code executing multiple times for the same bar or something along these lines. Make sure you're only ...

Machine Learning with algoTraderJo

The above results also make us wonder about the need for such complexity on the SVM model. Would a K-NN perform in a similar manner? You can see a comparison of the above same input/target structure using a K-NN. Predicting a 5 day direction Vs a 1 ...

Machine Learning with algoTraderJo

Let's experiment with how the results change when adding more inputs. The following results show you the same model as on my last post although the inputs have been changed to include a series of input variations. We try making predictions with EU ...

Machine Learning with algoTraderJo

By increasing the number of examples used for training, the number of inputs used per example (using more past bar directions) and placing the C at 10 and the gamma at 0.001 (train much harder at a finer level of detail) we can get much better ...

Machine Learning with algoTraderJo

So far when using classifiers to predict returns (1 bullish, 0 bearish) we have only looked at predicting the direction of the next bar's return. However we can also look at attempting to predict the return of the next N bars, making a prediction ...

Machine Learning with algoTraderJo

It's very nice to see your results They actually closely match what I get for that same period using an SVM and the settings you have suggested. I was in fact very suspicious when I saw your 1H results as in my experience the lower TFs ...

Machine Learning with algoTraderJo

Yes you're right. You cannot use incomplete past examples. This means that if your target looks into the next X bars the first example you can use is at least X bars in the past. This effectively discards bars between 0 and X-1 bars in the past. ...

Machine Learning with algoTraderJo

I'm back now! I will take some time to answer your questions and will then continue posting on the thread. It's great to see all your replies while I was gone

Machine Learning with algoTraderJo

I am leaving for the holidays today will see you guys again in a couple of weeks, leave any questions or thoughts on the thread! I hope you have enjoyed the ride up until now

Machine Learning with algoTraderJo

Suppose you have the following 2 examples, with 2 bar inputs that attempt to predict the return for the next bar. Here are the examples: First Example Closing prices = 1.2510, 1.2520, 1.2530, 1.2540 Example in terms of prices: (inputs) 1.2520, ...

Machine Learning with algoTraderJo

Yes, but I want to reserve Fuzzy classifiers for when we move to lower time frames when they become more important.

Machine Learning with algoTraderJo

You are most welcome. Keep the questions and discussion coming

Machine Learning with algoTraderJo

We're talking about the same thing. You're saying that the next prediction is always close to the last price (a zero return) because a zero return is the most likely scenario. This is precisely the problem that is exacerbated when you attempt to ...

Machine Learning with algoTraderJo

This is the auto-correlation of Oanda daily close prices for lags from 0 to 26 on the EURUSD. It is huge. Returns are very poorly correlated but prices are very heavily correlated because variations from one price to another are extremely low. This ...

Machine Learning with algoTraderJo

Yes, returns have very little auto-correlation. Do the same exercise with close prices (not returns) you will see a very heavy auto-correlation.

Machine Learning with algoTraderJo

This is the manner in which most academic papers approach machine learning. However it creates many problems related with the selection (what data is used for training, how often to train, when to train, etc), in my experience this leads to mining ...

Machine Learning with algoTraderJo

On the last posts using random forests for regressions we implemented a model using returns from the EUR/USD, GBP/USD and USD/JPY to make predictions. In the same manner we can create a more complex classifier that uses past bar directions (1 ...

Machine Learning with algoTraderJo

I'm glad you have found the thread interesting! Yes, this is a well known and somewhat expected problem. If you try to predict the next day's close using closing prices machine learning algorithms will always predict the best prediction to be the ...

Machine Learning with algoTraderJo

Consider that HFT firms do not trade the same type of systems we're concerning ourselves with here. The HFT firms trade statistical arbitrage strategies for which there is no need to demonstrate an inefficiency exists, because the inefficiency they ...

Machine Learning with algoTraderJo

Data means information. You can think about information as the amount of ticks that were used to build the data you are using. This is why 300 daily candles contain more information than 300, 5 minute candles. To build a daily candle you need far ...

Machine Learning with algoTraderJo