In latest days, I’ve added and checked a new improvement to the model that generates the forecasts published by spxbot.com. As you might know, the input to a neural network is usually preprocessed, for many reasons, usually to eliminate excesses in the raw data and to create a more uniform analysis environment. Even if it may seem bizarre, a lot of documents available on the web agree that adding noise to the input produces a better pattern recognition. In easy words, this process enhance the ability of the neural networks to generalize, or to extract meaning from the inputs, or simply to “see” better.
But what is exactly noise? and which noise and in what amount must be added? Here we are talking of Gaussian White Noise, which is statistical noise having a probability density function (PDF) equal to that of the normal distribution of the underlying data. So, the data is not structurally altered and the neural network is forced to consider a “blurriness” that is not present in the original data. This process is a bit time consuming, but actually worth it.
The result is encouraging: the forecast has not radically changed, but some signals previously hidden has appeared, suggesting that this improvement really makes the model sharper and better. As usual, no back testing is conducted because here at spxbot testing new features happens in real time with the collaboration of fellows subscribers.