Title: FinStressTS: A Parametric Synthetic Benchmark for Time-Series Forecasting in Finance

Abstract:

Accurate financial forecasting is hindered by challenges such as low signal-to-noise ratios, latent factors, heavy-tailed distributions, regime shifts, and sudden jumps. While real-world benchmarks allow researchers to observe underperformance, they rarely enable the isolation of specific failure causes because the underlying mechanisms are unobservable and deeply entangled. Furthermore, since actual financial data presents only a single realized path, it is difficult to properly evaluate tail-risk calibration or data efficiency. To address these issues, we present FinStressTS, a mechanism-aware synthetic benchmark that connects model behavior directly to controlled structural causes. This benchmark consists of 30 diagnostic environments spanning six mechanism families: volatility clustering, multi-scale persistence, heavy-tailed shocks, regime switching, self-exciting jumps, and zero-inflated processes.

We assess performance on two primary tasks: point forecasting, measured by Normalized Mean Absolute Error (NMAE) across five settings, and probabilistic forecasting, evaluated via Continuous Ranked Probability Score (CRPS) under known data-generating mechanisms. Our study benchmarks 15 models, ranging from classical approaches like HAR and VAR to Transformer-based forecasters such as PatchTST and iTransformer, as well as deep probabilistic architectures including DeepAR and TSFlow. We also utilize learning curves to gauge sample efficiency.

The evaluation yields three key insights. First, performance is heavily dependent on the specific mechanism; in environments driven by volatility, tails, or jumps, autoregressive and linear models are highly competitive and frequently surpass Transformer-based counterparts. Second, distributional alignment is crucial: while parametric probabilistic models like DeepAR demonstrate strong calibration in stationary settings, more flexible models offer advantages when distributions are multimodal or sparse. Third, neural models typically demand larger datasets to match the performance of simple baselines, with significant improvements occurring primarily when the task involves learning latent regimes or complex distributions. FinStressTS offers an open framework designed to diagnose failure modes and promote the development of risk-aware forecasting systems.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC