Model

The Model: Proposed Framework, Design, and Methodology

Architecture

VANGA Crypto model has come a long way of transformations and constant functional improvements.

Сonsequently, it has inherited all the best properties from traditional ML algorithms like LR (linear regression), SVM (support vector machine) or Gradient Boosting, econometric models like GARCH (Generalized Autoregressive Conditional Heteroscedastic models) and ARIMA (Autoregressive Integrated Moving Average models), and recurrent neural network-based models, such as LSTM (long short-term memory) and MLP (Multi-layer perceptron).

Together with that VANGA Crypto model included completely new innovative data science methods and techniques designed by CREAN Labs researchers.

VANGA Crypto model is based on a modified and improved recurrent neural network (LSTM). The model consists of two LSTM layers for extracting temporal dependencies and two fully interconnected layers for combining information and generating predictions.

The length of the sequence used to create a forecast is 31 days. To assess the quality of forecasts CREAN Labs data scientists team used standard metrics (MAE, MAPE, MDAPE, R^2).

Special Libraries

CREAN Labs data scientists used a number of ad hoc programming libraries most suitable for creating predictive models. These are libs like Prophet - programming library from Meta specifically for working with time series, Etna - a library from Tinkoff Bank IT Dept. for time series, and *tsfresh - a library for features generation.

The evolution process of our models’ elaboration included the following milestones:

First trying out and assessing the results of some simple traditional models: LR, Random Forest, Gradient boosting, SVM, ARMA, ARIMA, Garch, etc.
Assaying different neural network architectures. For time series, testing GRU, TCN, and some others. Identifying LSTM as a fairly optimal model for our objectives.
Using a composition of all listed models in different ways: majority voting, using parameters in a fully connected layer as features for ensemble methods, etc.
Experimenting with the model: trying different numbers of layers, various methods, different numbers of neurons, etc. In some cases it led to model retraining, but eventually showed excellent results.
Implementing a classification model in parallel, that is, predicting not the exact price for the next day but a price movement direction.

PreviousData NextResults

Last updated 6 months ago

Architecture

VANGA Crypto model has come a long way of transformations and constant functional improvements.

Together with that VANGA Crypto model included completely new innovative data science methods and techniques designed by CREAN Labs researchers.

The length of the sequence used to create a forecast is 31 days. To assess the quality of forecasts CREAN Labs data scientists team used standard metrics (MAE, MAPE, MDAPE, R^2).

Special Libraries

The evolution process of our models’ elaboration included the following milestones:

First trying out and assessing the results of some simple traditional models: LR, Random Forest, Gradient boosting, SVM, ARMA, ARIMA, Garch, etc.

Assaying different neural network architectures. For time series, testing GRU, TCN, and some others. Identifying LSTM as a fairly optimal model for our objectives.

Using a composition of all listed models in different ways: majority voting, using parameters in a fully connected layer as features for ensemble methods, etc.

Experimenting with the model: trying different numbers of layers, various methods, different numbers of neurons, etc. In some cases it led to model retraining, but eventually showed excellent results.

Implementing a classification model in parallel, that is, predicting not the exact price for the next day but a price movement direction.