A matter of frequencies: a new model

Spread the love

It all started as a standard revision routine. The code analysis tool showed some issues (it shows a lot of issues, definitively) that could be easily corrected. Once inside, my attention has been captured by the sequence of the steps that build the model. They were correct, but not completely correct.

I realized that the model is the correlation of the system of instructions on one side and the spectrum of frequencies on the other side. The market is a living beast and manifests its energy fluctuations through cycles. Now, all in with brute force. Now cover ALL FREQUENCIES.

For newbies to r.Virgeel, I would like to remember the following: all evaluations and predictions are generated from a model that has no trace of the SPX or even other stock exchange indices. A large variety of financial data is pushed inside the model and correlations are evaluated. The SPX index is virtualized from the components and correlated to the learning system applied to the SPX records.

This process works so well that adding the data from SPX or other stock market indices does not change a single digit on the output side. So, they stay removed. That also applies to the Technical Analysis tools I tried to insert and then deactivated: a Standard Deviation, a Signal Noise Reduction and the beloved mhDMI, a personal variation of the DMI. They all seem irrelevant, once again, to the evaluation of the output. In the past, I’ve also tried the COT data without any difference. The COT also is irrelevant (and it makes sense, as the COT data is delayed).

The hot bun out of the oven:
here is the first forecast chart from the new brain,
directly from the calculus tool.

“Out Of Memory” messages from the console. This moved the avalanche. Revision of code line by line and more than once. Reconfiguration of the working space for best optimization and fingers crossed. Yes. Now the new “brain” is larger than anything before: during development, it reached 52 Mn (52 million neurons) and it is now stable at 48 Mn. Calc time is now consistently higher. But, a new big step has been achieved: the model is now, for the first time, totally deparametrized.

But it does something I never tried before: facing the instructions with ALL the spectrum of frequencies. In other words, the total and irreversible surrender to inspect optimal frequencies. Forget it. Just passing all of them and the black box makes the trick. Where the surfaces contact, there something happens.

If the intuition is correct, it will be sharper than before. I’m here to watch it. After two intense days of coding and decoding, the stability is here again. Just need to check the incoming bars as they come in from the future and see materialize the analysis from r.Virgeel. As usual.