LFTD: Transformer-Enhanced Diffusion Model for RealisticFinancial Time-Series Data Generation

Firm-level financial statement data form multivariate annual time series with strong cross-variable dependencies and temporal dynamics, yet publicly available panels are often short and incomplete, limiting the generalization of predictive models. We present Latent Financial Time-Series Diffusion (LFTD), a structure-aware augmentation framework that synthesizes realistic firm-level financial time series in a compact latent space. LFTD first learns information-preserving representations with a dual encoder: an FT-Transformer that captures within-year interactions across financial variables and a Time Series Transformer (TST) that models long-horizon evolution across years. On this latent sequence, we train a Transformer-based denoising diffusion model whose reverse process is FiLM-conditioned on the diffusion step as well as year, firm identity, and firm age, enabling controllable generation aligned with firm- and time-specific context. A TST-based cross-decoder then reconstructs continuous and binary financial variables for each year. Experiments on Korean listed-firm data from 2011–2023 show that augmenting training sets with LFTD-generated samples consistently improves firm-value prediction for market-to-book and Tobin’s Q under both static (same-year) and dynamic (τ→ τ+1) forecasting settings, and outperforms conventional generative augmentation baselines and ablated variants. These results suggest that domain-conditioned latent diffusion is a practical route to reliable augmentation for firm-level financial time series.