for what it's worth, I employ a convolutional network in order to predict co-integration properties of financial time series. I am disincentivized to go into details but I can tell you that every hedge fund, every bank, every buy side firm that I have friends working at is at the very least researching and experimenting with deep learning networks. Some are making serious inroads into utilizing every last bit of academic research in this space as well as pushing the limits further. Deep learning networks will revolutionize most every area of our lives in the next 10 years.
Neural Networks Revisited
- Thread starter maninmoon
- Start date
Actually I found the issue. I had my input as 5 days delta, but output was 5 day delta OFFSET by 1. Since I corrected it, the NN now very closely matches solution to output as I was expecting.
The reason for doing this (seemingly silly) test of same inputs as outputs was to CHECK that NN actually gave the correct result!
Hey Conduit, are able to say whether you approach works on FX? I am from FX background, but I have data for FX and Equities.
Thanks for the intro.
If you can steer me towards a profitable NN I would be grateful.
I've suspected this for awhile. thanks for the post.
Yeah, also my bad. I neglected to remember the bias unit, which could cause the activation to always fire at 1
also if the input is always >> 1, the activation could always fire full.I am working with some daily FX data. I have created 5 inputs: deltas, averages, RSIs etc. And I am trying to predict the "Forward" delta i.e. the price change 1,5 or 20 days ahead.
I have normalised all data between the min and max of a fixed period and mapped that to range 0-1.
I am looking at how to structure the layers. The simplest is 5-5-1 (Input neurons, hidden neurons, output neuron).
Now I could try 5-10-1 or 5-5-5-1. i.e. add 5 neurons to hidden layer, or add another hidden layer with 5 neurons. Obviously, I am going to test the difference empirically, BUT I would welcome any thoughts on what structure is likely to prove fruitful
I have normalised all data between the min and max of a fixed period and mapped that to range 0-1.
I am looking at how to structure the layers. The simplest is 5-5-1 (Input neurons, hidden neurons, output neuron).
Now I could try 5-10-1 or 5-5-5-1. i.e. add 5 neurons to hidden layer, or add another hidden layer with 5 neurons. Obviously, I am going to test the difference empirically, BUT I would welcome any thoughts on what structure is likely to prove fruitful
I'm working on this subject for a while, experimenting with different architectures and models.
When I started to obtain more or less satisfying results, I've tested them on my own account, rather successfully.
After that I decided to publish predictions using website - http://stocksneural.net.
As you can see, there are some instruments on which models produce stable and profitable predictions, but there are also a lot of models with lower quality. So there's a lot of work to be done, and I'm constantly busy with experimenting with new models and trying to incorporate bleeding edge academic findings.
Also there are some glitches with the representation of information, especially with recommendations, I'm working on it also.
Dmitry, do you show VAMI or equity curves for individual stocks anywhere?
Not yet, do you think it's worth adding? Does it have a great sense for individual stocks or is it better to let's say construct some portfolios and track VAMI for them?
Your opinion is greatly appreciated.