Mirocana. Inside Look.

As the main theory of making predictions we use historical extrapolation. System structured to unify all possible data types to one common format. In oder to create self-learning system with the ability to adjust fast-changing market conditions, we decided to build AI that will take into account universe of strategies paying attention the particularities of each individual signal.

Mirocana is a platform for strategies, it's designed to make due diligence and take advantages of each strategy of collection. Strategies may take as input different kind of data to deal with, such as: quotes, news, economic calendar events, data releases, other traders activity, tweets and etc. The complexity of strategies may vary, but the output always fit the same format. That allows us to calculate wights among strategies properly and shortly.

There's a few different sources of strategies. The main one is strategies manually generated by quants of Mirocana Team. Some significant amount of strategies are coming from Strategy Generator that automatically explore data correlation. In near future we are going to open intuitive web-interface for writing new strategies, so our partners and individuals can contribute and get paid. This will unlock full power of crowdsourced learning.

for Telegram
Tech Stack.
This instruments we use in our work mostly:
ClickHouse’s performance exceeds comparable column-oriented DBMS currently available on the market. It processes hundreds of millions to more than a billion rows and tens of gigabytes of data per single server per second.
as a primary database
Python 3.4 is very fast language when it comes to writing code on it. Python allows us to to express our concepts in fewer lines of code than any other programming language.
Python 3.4
as primary language
TensorFlow was open-sourced by Google in 2015. Definitely, it's the most powerful machine learning low-level library for Python.
for numerical computation
To identify certain market situations when given strategy shows it's best performance we use market images technology. Market Image is a vector of more than 200 params that describes current market situation on symbol. As soon as new signal arrived from the strategy we can find a few signals in the past with approximately same market image, that bring us more insights on how model should calculate the wight for this signal.
We constantly trying different approaches of machine learning to build models that make most accurate predictions without overfitting. We still on our way to find out gold-standard model and we hope that Mirocana Predicting Tournament will help us improve our models. One of the method of making predictions is based on family of models called neural nets. It's easy to understand it on real tasks, you can find them on TensorFlow demonstration.
TensorFlow Playground
Can help you to understand the main principles of neural nets:


Which dataset do you want to use?


Which properties do you want to feed in?

Click anywhere to edit.
Weight/Bias is 0.2.
This is the output from one neuron. Hover to see it larger.
The outputs are mixed with varying weights, shown by the thickness of the lines.


Test loss
Training loss
Colors shows data, neuron and weight values.

We deeply believe that Mirocana will become extremely accurate predicting system of the future. Detailed, global and modern.