Generating virtual scenarios of multivariate financial data for quantitative trading applications

TL;DR

This paper introduces a method to generate realistic, customizable multivariate financial data scenarios for testing trading strategies, preserving key statistical properties and correlations of real market data.

Contribution

The paper presents a novel PCA-based approach to synthesize multivariate financial data that maintains realistic distributional and correlation structures, allowing flexible scenario creation.

Findings

Simulated data closely matches real financial data in statistical tests.

The approach outperforms classical and state-of-the-art methods in data realism.

Generated scenarios are useful for testing quantitative trading strategies.

Abstract

In this paper, we present a novel approach to the generation of virtual scenarios of multivariate financial data of arbitrary length and composition of assets. With this approach, decades of realistic time-synchronized data can be simulated for a large number of assets, producing diverse scenarios to test and improve quantitative investment strategies. Our approach is based on the analysis and synthesis of the time-dependent individual and joint characteristics of real financial time series, using stochastic sequences of market trends to draw multivariate returns from time-dependent probability functions preserving both distributional properties of asset returns and time-dependent correlation among time series. Moreover, new time-synchronized assets can be arbitrarily generated through a PCA-based procedure to obtain any number of assets in the final virtual scenario. For the validation of such simulated data, they are tested with an extensive set of measurements showing a significant degree of agreement with the reference performance of real financial series, better than that obtained with other classical and state-of-the-art approaches.