The underlying premise of (supervised) machine learning is that there is predictive value present in the predictor values fed into a model, but a non-linear relationship between them and the target, making them difficult to relate through human logic. Yes, most models will fit any data you put into them even if there is no predictive value and this is exactly what we want to avoid.
Machine Learning Algo for Trading
- Thread starter stepseazy
- Start date
The underlying premise of (supervised) machine learning is that there is predictive value present in the predictor values fed into a model, but a non-linear relationship between them and the target, making them difficult to relate through human logic. Yes, most models will fit any data you put into them even if there is no predictive value and this is exactly what we want to avoid.
What exactly should be accomplished by the engine?
The best algo traders use tick charts.....which help eliminate tiny, choppy candles that result with time bars when the trade activity drops...say from 11 am to 2 pm.
Ticks ruin you
When the tick frequency exceeds your service time
When the tick frequency exceeds your service time
Not sure why that is not apparent. Take the GO competition between human and computer. The input data are merely used by the algorithm (=engine) to train the network. How it trains it and how it used the trained network to make prediction are the real secret sauce. Certainly not the training data.
Same with an image recognition network. The input data are just labeled image data. Everything else, the building of associations and the computed weights to later on recognize images is all accomplished by the actual algorithm.
Same with an algorithm that peruses time series or other financial input data...
Same with an image recognition network. The input data are just labeled image data. Everything else, the building of associations and the computed weights to later on recognize images is all accomplished by the actual algorithm.
Same with an algorithm that peruses time series or other financial input data...
This is a common misconception of how machine learning works.
Why do you think all attempts to train a neural network with GO games failed over the last 20 years? And why did AlphaGo succeed? it had simply two differences to all previous attempts, a deeper network with more neurons, and a clever preparation and preprocessing of training data. In fact a separate network was used for data preprocessing.
It's the same with financial data. If we could just throw prices to some ML algorithm and abracadabra, it predicts the next trades, we all were billionaires by now. 90% of all effort for ML prediction for financial data goes into preprocessing and selecting the features.
Last edited:
No.
Because you affect the system.
Trading is new perturbation.
Because you affect the system.
Trading is new perturbation.
That is utter nonsense. Everything changed about neural networks not just about computational power but about the inner workings of the networks, back propagation, auto differentiation, GPU computations, convolutional network components among many other algorithm specific bits and pieces that did not exist before. You clearly have no idea what you are talking about. In fact nothing about the input data changed AT ALL. Its the same historical game data that was fed into the system that existed for ages. In the same way the same financial time series that existed for decades can be fed into a neural network. Hardly anything about the data changed (other than the amount of data that can nowadays be processed and even that is made possible by more computational power and algorithms that are designed in a more clever way.).
Of course a researcher has to be knowledgeable about which data to feed into the system but even that has not changed over time. I believe you could not possibly be more wrong about your assertions.
Of course a researcher has to be knowledgeable about which data to feed into the system but even that has not changed over time. I believe you could not possibly be more wrong about your assertions.
...and that is easy to figure out by looking how predictions hold up against a validation data set. Not sure still why this makes the input data more important than the inner workings of an algorithm.
The whole point of competitions where the smartest AI researchers compete against each other to push up prediction rates is to work on better algorithms. All researchers have the exact same data set at their disposal. Hence the differences in results are 100% and completely independent of the input data.
The whole point of competitions where the smartest AI researchers compete against each other to push up prediction rates is to work on better algorithms. All researchers have the exact same data set at their disposal. Hence the differences in results are 100% and completely independent of the input data.