Engine maintenance

Spread the love

Latest days were spent revising the Amodel’s engine. For as simple as neural networks can be ideally, they cannot work if not arranged in an “engine”, a complex of parts that interact. I played with two components mainly: the preprocessor and the “booster”. The preprocessor code takes the raw data and arrange it in the proper way, the booster add self-reference to the set of data. Both are quite long, repetitive and delicate pieces if code. The quality of the vision produced by the model is totally dependent from this preprocessing phase.

I planed on the code with the machete in one hand and cleaned, rearranged, condensed and deleted. In one word, I simplified. A couple of times, the model went blind, without recovering. So I did restart from the beginning, and again. But the result were interesting: I have seen that the model can still improve and one of the secrets is that raw data must be manipulated as little as possible.

The Amodel engine as a whole has become nicely complex through development and I have suspects that it is very near to the hardware/software limits of the the platform it is developed on. It’s a surprisingly old piece of hardware that crunches the nuts!

And this brand new polished model, what does it say about the future of the S&P 500? Well, I can abstract this way: stay hungry, stay long!

 

 

Leave a Comment

Your email address will not be published.