TS IS FRESH

Time Series Importance based Selection and FeatuRe Extraction on basis of Scalable Hypothesis tests

This is the documentation of ts-is-fresh. Algorithm implements the idea of combining the selection of features by their importance and the generation of features using the tsfresh library :)

Project came from the task of building important features for high-frequency trading.

Here is an idea that combines:

• analysis of the similarity of price behavior for different currencies

• automatic generation of statistical statistics by time windows (from tsfresh)

• feature selection through statistical hypothesis tests (from tsfresh)

• reducing the dimension of the feature space due to the clustering of highly correlated features

• counting feature importance values (including shap values) through block cross validation

• feature selection based on their importance values

Contents

• Introduction
  • Overview
  • What problem does ts-is-fresh solve?
  • How does it solve this problem?
  • And why exactly in this way?
• Algorithm
  • Preprocessing
  • The basic version of the algorithm - working with a single time series
  • Advanced version of the algorithm - take into account the context
• Toy Example
  • Data
  • Block Cross Validation
  • Stats Selection
  • Importance Selection
  • Remarks
• Source code
  • extraction_utils module
  • preprocessing_utils module
  • selection_utils module

Indices and tables

• Index

• Module Index

• Search Page