Our algorithms - FERT experienced solutions

Go to content

Our algorithms

Ideas

It's easy and intuitive to say that it's logic that market fluctuations  are not guided by time series, but by others such as fundamentals, world  markets and by the sentiment of people, operators, bankers, politics  and so on.
The first question we faced is how to put all this  potential inputs into an algorithm that can use them to predict the  future movements? Not mathematics or statistics, it's clear to evaryone  that we can't add pears to apples! How to merge data from Twitter, with  time series? How to compare sentiment and market indexes?
Before  the introduction of AI to financial calculations this was possible only  by creating multiple indexes and see their internal and external  correlations, but there's usually poor mathematical correlation between  them. But our common sense says no, it's not true. There must be some  correlation!
The answer we wanto to provide is driven by  Machine Learning algorithms and AI. These are more than mathematical  algorithms, their are something that can start learning (create weights  for variables) and learn more, day after day (repeat calculations n  times and progressively reduce the errors), such behaviour fairly  simulates that one of our brain.
Our brain knows that there  must be some correllation between those entities, and AI does the same.  The difference is that we cannot put on an algorithm the weights created  by our brains (our synapsis) but we can do so with those one created by  our computers.
Here's where we started.


AI algorithms models make use of trained models.

Training data can be of the most variable types, e.g. historical  data of shares fluctuations, market indexes, foundamental data,  sentiment data, etc. that can be found and retrieved from the most  relevant and reliable fonts. Once data is collected it has to be  organized into feature vectors, or better feature tensors, that means  that they are normally n-dimensional matrixes.
Learning is  then done by running the neural network(s) for a determinate number of  epochs on the training dataset or for an indeterminate nr of epochs, for  example continously once the model has reached the desired level of  accuracy.


Once the model is well trained (this means that the error curve is  minimized), e.g. on historical data, predictions can be run. At this  point calculations are far faster as the output of the learning phase is  a model that could be somehow reduced to a mathematical formula with  addictions and weighted multiplications, where weights are given by the  training phase (what we earlier compared to our synapsis).



The final output is given by prediction, that could be a one or more  days scenario, according to how the training was run and the features  set built.


Our systems.


The architecture is based over a dozen of different neural networks, with recalculate the data every day and more times per day.
Our data  is constantly retrieved from various sources on the web and stored on  high potentials non-relational database. All our algorithms are built on  open source languages, mostly python, and lybraries such as Tensorflow,  Pandas and NumPy.
One server is dedicated to data gathering  and organization, another one  is dedicated to train the networks, which  is repeated on weekly and daily basis, according to type of nework, its  target and the weekly schedules. All the sistem is the redunded on a  backup mirrored architecture. The two architectures are keeped separate in order to avoid security issues to the entire system.

Our  intention is immediately to provide a very resilient and structured  service and in the times to costantly improve the algoritms and  introduce at least two new networks every semester, in order to  guarantee a constant improvement of performances.
FERT experienced solutions is a registered trade mark. All rights reserved.
Contacts: fert.experienced.solutions@fertes.eu
Back to content