logo
RU
RU
EN icon

TECHNOLOGY

Have a close look inside of the Mirocana structure and internal design.

The system structured to unify all possible data types to one common format. We build a system that takes into account as many factors as possible: technical indicators, chart patterns, correlations, fundamentals, trading activity of other traders, analysts and hedge-fund managers, news, tweets, articles and other data that may correlate with assets' prices.

Mirocana has a multi-layers structure that allows us to increase the level complexity in each of these layers in parallel.

HOW DATA TURNS INTO INVESTMENT PRODUCT

LAYER №1

DATA SOURCES

We receive the raw financial data from a vast number of data sources in real-time.

Also, we store historical data from each of these data sources.

01

LAYER №2

STRATEGIES

We have many strategies that interpret the data we receive from our data sources.

Every data source has at least a few strategies.

02

LAYER №3

SIMULATIONS

Strategies may work and may not. Simulation is resposible for taking all the predictions from all the strategies and calculating the weights among them using deep-leaning neural nets in order to maximise profits.

03

LAYER №4

BEST PERFORMING SIMULATION

When we know the performance results of each simulation we select best performing simulation and deploy it into production to let it manage the funds.

04
product

LAYER №1

DATA SOURCES

We receive the raw financial data from a vast number of data sources in real-time.

Also, we store historical data from each of these data sources.

01

LAYER №2

STRATEGIES

We have many strategies that interpret the data we receive from our data sources.

Every data source has at least a few strategies.

02

LAYER №3

SIMULATIONS

Strategies may work and may not. Simulation is resposible for taking all the predictions from all the strategies and calculating the weights among them using deep-leaning neural nets in order to maximise profits.

03

LAYER №4

BEST PERFORMING SIMULATION

When we know the performance results of each simulation we select best performing simulation and deploy it into production to let it manage the funds.

04

LAYER №5

PRODUCT

Our investment products are based exclusively on best-performing simulations. They are connected to a monitoring unit that is able to stop all activity in case of an accident or emergency situation.

05

LAYER OF DATA SOURCES

We collect, process, store and analyze huge volumes of financial data that we receive from 40+ data sources. Each data source is a public or private site or API service that has information that we think may correlate with the prices of assets we predict or will predict.

We pay for the data we receive. With some data sources we have a non-disclosure agreement and they are not listed below.

Quotes, fundamental and macroeconomic data.

Quotes, economic calendar and orderbook.

Quotes, market data and fundamentals.

Quotes and market data.

Quotes and market data

Quotes and crypto-currencies market data.

Quotes and market data, sentiment data and news.

Graphical chart patterns.

Traders’ activity and market data.

Traders’ activity on the platform

Traders’ activity on the platform

Traders’ and algorithms activity.

Traders’ and algorithms activity.

Traders’ and hedge-fund managers’ activity

Searches and trends data

TTraders’ activity, announcements and news.

Analytics and hedge-fund managers activity and insides.

News, articles and traders’ sentiment.

Economic calendar events, news, traders’ sentiment data.

Reports and data releases. Market activity data.

DIRECT PARSING

News, articles and other information.

We will be adding new data sources in the future. Members of our team permanently negotiate new partnerships and integrations that could positively affect our investment performance.


LAYER OF STRATEGIES

Strategy is the code that describes the logic of how we interpret data from the data sources. It indicates what we should do when we receive a new data point from the data source - should we buy, should we sell, should we stay neutral.

Each strategy in the system generates a prediction, during its backtesting. These predictions are recorded in the database. Each prediction contains information about the direction (to buy or to sell), confidence level, duration and meet data of the market conditions in which this signal has happened.

We have strategies that are only based on quotes data. These strategies can be applied to any financial instrument and to any time-frame, since they are just a denormalization of historical series of quotes. Some strategies are based on the trading activity of other market player (traders, experts, hedge-fund managers). These strategies predict how successful a newly placed trade based on the past performance of the trader will be. Some strategies are based on tweets, news, and articles. These strategies use our own and open-sourced tools of syntax analysis to determine the sentiment level of the text. Additionally, we receive prepared sentiment results from some of our data sources.

EACH STRATEGY CAN BE CLASSIFIED TO ONE THE 9 CATEGORIES

Based on activity of traders and insiders
Based on technical indicators
Based on fundamental data
Based on news and articles
Based on macroeconomic data and events
Based on graphical chart patterns
Based on symbols’ correlation
Based on social media activity
Autogenerated strategies

LAYER OF SIMULATIONS

Simulation is the program that takes as an input all predictions from all the strategies in the system and calculates weights for each strategy based on its performance. Then using this vector of weights, it calculates a cumulative prediction that is used to place buy and sell orders in the simulation environment.

We have many simulations with the diffrent logic of weights distribution. Some of them use deep-learning neural nets, decision trees, gradient boosting, reinforcement learning, quality learning (based on Markov's decision process) and many other machine learning models to properly calculate the weights. Our Core Team and Mirocana Research Group experiment every day to find best performing predictive models. After all the results for all the simulations are calculated, the best performing simulation is selected and allowed to manage the funds.
In our work, we use our own and open-source tools for machine learning and data science. Some of them are:

CREATE YOUR OWN NEURAL NET

Epoch

Data

Which dataset do you want to use?

Features

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.

Output

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

To better understand the concept of a neural net please use the interactive playground.

CROWD-SOURCED LEARNING PLATFORMS

Our Core Team and Mirocana Research Group are working every day to add new data sources, strategies and simulations, but we believe that there's a lot of talented people who are ready help us. We develop three crowd-sourced learning platforms to improve all three layers of the system.

a

alpha

Create a new strategy

A web platform, where any person with basic coding skills will be able to create, test and evaluate his/her own strategy and get paid in MIRO tokens.

s

sigma

Create a new prediction model

A web platform, where any data scientist will able to create, validate and score his/her machine learning model based on our encrypted data for predictions.

t

target

Add your forecasts

Android and iOS app, as well as a web platform, where any talented individual will be able to add his/her own predictions for stock, currency and crypto-currency markets.

LONG-TERM COMPETITIVE ADVANTAGES

01

FLEXIBLE CODEBASE

We are a team of engineers and data scientists. And we solve a problem of predicting the markets as a data science problem. The majority of traditional trading firms use third-party software to write and backtest their strategies, therefore they are limited by that software. We wrote everything from scratch. We have developed our own engine to backtest strategies and to run the simulations. This allows us to quickly scale our system and unlocks endless field for new exciting experiments.

02

HIGH SCALABILITY

To constantly improve accuracy of system’s predictions we must add new data sources, new strategies and new simulations. We do that by our own but also we develop three products for crowd-sourced learning that will attract smart people from all over the world to help us. As well, our Mirocana Miner product will solve the problem of constantly increasing needs for computing power.

03

CUMULATIVE EFFECT

It is known that many traditional investment firms use a few dozens of strategies and manually allocate capital among them trying to achieve balance. Our system bases its final predictions on all the strategies that we have at once. Each strategy by itself is able to generate profit, but combining them together using deep-learning neural nets and other ML models creates a cumulative effect, some sort of super-strategy that has the results that are better than the results of the best performing strategy.

04

MARKET ADJUSTMENTS

Trading robots that are used by investment banks and trading firms may be very complex, but the code that describes their logic is usually no more than 2 thousand lines. Such algorithms may work well for extended period of time, but when the behavior of market is changed, this robots do not change, they start losing, and they are manually replaced by another working algorithm. Mirocana system is very different: if some strategy stopped working it will be given a low weight after some iterations, and this strategy will not affect the final prediction.

05

ANALYSING HUMANS

Some of our strategies are based on the activity of humans traders, analysts and hedgefund managers. Our system learns strong and weak sides of this traders based on their past performance and is able to predict their reaction on certain market events. By analysing the reaction of a small fraction of traders, our system is able to extrapolate this reaction for the rest of them.

06

BACKTEST LEGITIMACY

Backtest Legitimacy: If you ever had a try to code a trading strategy, you may noticed that your strategy shows very decent results on the backtest, but is not working quite well on the real data. This happens because you adjust and validate your strategy parameters based on the full result of the strategy, meaning based on future data that your strategy should not know. To prevent this you will need to develop you strategy on training data and validate it on the test data. In the Mirocana system, when a new simulation is launched it iterates over history, making training and testing at the same time and it does not know anything about the future data - that ensures the legitimacy of our backtesting.

IN CLOSING

We will continue to push the level of sophistication in all three layers of the system to increase the accuracy of Mirocana predictions and to improve the quality of our investment products. We ourselves do not fully understand how far we can go. What would be if there will be a few thousand strategies in the system - what results we could get.

MIROCANA 15-YEARS LONG VISION icon