Intelligence Authentic + Financial Analytics

GPU Enabled Stochastic Optimization for Millions of Backtests

 

Challenge

Parallelization and scale of mathematical computing is the forefront of the field today. With many algorithms taking hours or even days to utilize results, the ability to quicken computations enables the best of the best to make better decisions in the markets.

Solution

A recent client endeavored to build a backtesting system capable of processing millions of scenarios, parameters, and algorithms for equity trading purposes. We developed a scalable, efficient C++ system utilizing CUDA compatible GPUs to do so. Algorithms had numerous measures to validate their in sample and out of sample performance, mainly utilizing variants of the Sharpe, Sortino, and Kappa ratios. =

Results

Improved performance of backtesting system in terms of computational speed by over 1000x compared to other C++ systems, 1,000,000 against an original Python parallelized system.

  1. Enabled strategy selection via an advanced stochastic optimization processes on price datasets with 1000s of trading days - previous systems could only handle a few days at a time and larger trading intervals.

  2. Helped improved out of sample performance of algorithms - depending on the ratios used, it was common for individuals backtesting their strategies to find 10-30% gains over their previous chosen parameters and algorithms.