Can be very important or not so important, depending on your problem, in my experience at least (it might not be a general conclusion). As you have done I regularly use Euclidean distance (simplest to implement) where this doesn't work the problem ...

Machine Learning with algoTraderJo

Well, machine learning methods find relationships that might not be obvious enough to be found through a more transparent method (like those you suggest). What can be found also depends greatly on the model you use which allows you to control ...

Machine Learning with algoTraderJo

I am using direction (1) bullish, (0) bearish. You're indeed right that this is an extremely basic K-NN case, which is what makes it a good case for the building of a basic practical example.

Machine Learning with algoTraderJo

Your questions are quite fundamental !! If you trained a machine learning model with data from 2000 to 2014 and then executed a simulation of that same period using the model you have trained you would have effectively curve fitted the ...

Machine Learning with algoTraderJo

Agree to disagree. I have been working with machine learning methods for years and I have achieved good success in live trading. You do need to be aware of the statistical biases present and you need to ensure that you do have something that has ...

Machine Learning with algoTraderJo

This is quite more complex than what I am doing here. The K-NN and linear map methods I have showed here have the advantage of always giving the same result (they are in this sense very deterministic) the neural networks used in the Asirikuy Brain ...

Machine Learning with algoTraderJo

In my experience using close prices for predictions is always a poor choice because close prices are strongly auto-correlated. This is a reason why you often get predictions that are just the last close price when using close prices on machine ...

Machine Learning with algoTraderJo

I also do machine learning for a variety of problems (in industry) besides trading and I can tell you that trading is perhaps the worst of them all. In academia/industry most of the problems you tackle are deterministic (like computer vision and ...

Machine Learning with algoTraderJo

I have been trading like this for at least 4 years (using machine learning based systems that is). Although I have only been trading using the framework from the people at Asirikuy for the past year and a half. I am currently doing prop trading for ...

Machine Learning with algoTraderJo

If we continue to explore our K-NN algorithm and attempt to find the optimum amount of information for this model we find that using 4 bars as input gives the best overall result (remember that we are always rebuilding the model on each bar using ...

Machine Learning with algoTraderJo

Deep learning turns out to be too expensive computationally for me. Measuring data-mining bias often involves running extensive computational trials using random data obtained through bootstrapping with replacement of the original data in order to ...

Machine Learning with algoTraderJo

Regarding the open source libraries you suggest, my main problem with using them - mainly the python ones - is that going to live trading and doing simulations is not so straightforward. I do not have the time to program/debug my own simulator and ...

Machine Learning with algoTraderJo

When you add new information that contributes with predictive power you actually get better results. The graph below shows you our K-NN test using 2 bar directions as input and a second test using 18 bars as inputs. As you can see the addition of 16 ...

Machine Learning with algoTraderJo

We can now take our K-NN model and attempt to add more information to see if it improves results. Inputs can be as varied as you want and the effect of the inputs can be determined by comparing your system's overall trading results. When you add ...

Machine Learning with algoTraderJo

Yes, I have tried some things with Bayesian networks but I often find simpler models to work better (often depends on what you want to predict as well). In my practical experience successful ML can depend heavily on your model parameters/structure ...

Machine Learning with algoTraderJo

I have already started a thread on machine learning here. We may eventually reach the topic of building models to determine trade continuation probabilities (although it might take a while!).

Some silly ways of looking at P/L

Oh yes, you can attempt to train to predict trade exit targets. Which is something very interesting as well (this is somewhat how some of the strategies I trade live actually work). I will continue exploring the case of daily bar predictions to get ...

Machine Learning with algoTraderJo

You would think that! However it depends heavily on how your right/wrong answers are distributed. Knowing the net directionality of candles that have little movement is of little use, while being wrong in candles that have large values is a bad ...

Machine Learning with algoTraderJo

I used to code my machine learning strategies using R exclusively, but then things were a pain to run live due to problems with linking R to live execution platforms and APIs. Right now I use the programming framework provided by Asirikuy which ...

Machine Learning with algoTraderJo

Yes, you're right! This is a big part of the reason why we are getting poor results when using the linear mapping algorithm. Because our profitability is poorly related with our prediction. Predicting that days are bullish/bearish is of limited use ...

Machine Learning with algoTraderJo