Constraint-Guaranteed Synthesis of OHLCV Data

Abstract

Existing generative adversarial network (GAN) based approaches to financial time series synthesis produce univariate return series only. They provide no mechanism to enforce the geometric constraints that every valid bar must satisfy: the High must weakly exceed both the Open and the Close, and the Low must weakly fall below both. We solve this through a constraint-by-construction reparameterization of each OHLCV bar into five components (close-to-close return, overnight gap, upper shadow, lower shadow, and volume) under which the bar constraints are algebraically guaranteed for every generated bar regardless of generator output. Zero violations were recorded across 1,255,500 generated bars; this result is independent of training quality and holds for any output of the trained generator. Building on QuantGAN, we extend the stochastic volatility neural network to a five-channel joint generator trained with the Wasserstein GAN with gradient penalty (WGAN-GP) and auxiliary losses targeting return drift, volume collapse, and volume-volatility co-movement. Evaluated on S&P 500 daily data from 2016 to 2025, the model reproduces the intrabar range distribution with an 8% mean error and produces a contemporaneous volume-volatility correlation of 0.23 against a real value of 0.41. Volatility clustering remains an open problem, which we discuss alongside a concrete proposed resolution. Code is publicly available at https://github.com/KambizHoma/HardConstraints.