I will get to that with time. You'll see that we need to make some modifications to make them work on lower timeframes

Machine Learning with algoTraderJo

A good question is to wonder about the sensitivity of the target. Are things very different if we use slightly different targets for prediction? Keeping the frontier at 35 and using the same targets with exactly the same SVM ensemble (as mentioned ...

Machine Learning with algoTraderJo

Another thing we can play with is the prediction target. I previously selected it as 0.5*ATR(20) but you can change it to higher values in order to try to predict a more favorable move in the currency pair. Changing the target to 1.5*ATR(20) means ...

Machine Learning with algoTraderJo

Happy to hear you're enjoying it! I will try to keep it interesting

Machine Learning with algoTraderJo

We can attempt to eliminate our dependency on the number of inputs and the number of learning examples used on each bar by making decisions using a group of models instead of a single model. If the overall premise of input and target relationship is ...

Machine Learning with algoTraderJo

The number of examples used to train on each bar is also important. Below you'll see a graph with the results using 40, 60, 100 and 120 examples to train the SVM model on each bar (using 10 bar directions as inputs). This means that we actually need ...

Machine Learning with algoTraderJo

I agree with you in that MM does not generate any edge. However it does affect the results significantly so I would advice you to use the same MM to reproduce my results. The GMT offset is far more critical however, since it directly affects the ...

Machine Learning with algoTraderJo

The previous post begs to ask whether the results are simply due to more information or due to a better model for the description of the input to target relationship. Running tests for strategies using SVM models using the past 2, 4, 6, 8, 10 and 12 ...

Machine Learning with algoTraderJo

It is evident that we wouldn't want to trade something that has unstable performance under recent market conditions (gone flat or started losing money) so I will now go a bit into some measures to attempt to increase our performance statistics for ...

Machine Learning with algoTraderJo

I have to emphasize here that the money management aspects of the strategy are critical. The strategy (in fact all strategies I post) adapt their lot sizes against market volatility. Trading a constant lot size will always bring you worse results ...

Machine Learning with algoTraderJo

You talk about a valid issue. When I train the model on each bar I start looking at past data to build trading examples. The first valid examples exist 20 bars in the past (because I need to be able to look 20 bars forward to build the target) if ...

Machine Learning with algoTraderJo

I'm traveling now for the weekend! Leave your thoughts

Machine Learning with algoTraderJo

This is just a comparison of some of the things we have studied up until now: image We are yet to get a regression approach that is as successful as our classifiers and to get better performance under recent market conditions.

Machine Learning with algoTraderJo

I now want to post some experiments related with a more complicated input/target structure using classifiers. Let's say that we now want to predict something that is more useful to our trading, we want to predict whether a + or - given target will ...

Machine Learning with algoTraderJo

The Timezone used on this file is GMT +1 non-DST and GMT +2 DST. All the files I post will be using this TZ. The SL can be hit at anytime within the daily bar but the timestamp is rounded to the daily bar. Also you're right. For the 2007-2014 period ...

Machine Learning with algoTraderJo

I only use R code to make the graphs and show you the analysis. As I have posted before all simulations are carried out using the Asirikuy programming framework and tools.

Machine Learning with algoTraderJo

Since the linear regression offers a prediction of return values we can see what happens if we try to filter the results by only taking into account predictions that are above 5 pips. In theory this should eliminate all border-line predictions (+/- ...

Machine Learning with algoTraderJo

I already posted a csv file containing actual/predicted data for the EUR/USD linear regression test.

Machine Learning with algoTraderJo

We will hopefully move into this issue of the daily starting time later on, you will see that the results change depending on how you actually start/end your daily candles. I believe that this is because the return of each day must encompass the ...

Machine Learning with algoTraderJo

Yes, in FX the devil is really in the details. I use the European DST because most of my strategies depend on having the European market open accurately timed. Since the market opening times also have DST you will see that this also becomes ...

Machine Learning with algoTraderJo