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

Battle Count: 50

Highly relevant for quantitative trading research as it provides a diagnostic framework for understanding why forecasting models fail in financial contexts. Key insights include: (1) simple models often outperform complex Transformers in financial forecasting, (2) mechanism-specific model selection matters more than universal sophistication, (3) probabilistic calibration must be validated independently from point accuracy, (4) neural models require 2-3x more data than classical baselines. However, it is a diagnostic benchmark rather than a trading strategy generator, and does not directly produce trading signals or portfolio allocations.

The benchmark framework itself is moderately complex, requiring implementation of six parametric data-generating processes (GARCH, HAR, Student-t, Markov-switching, Hawkes, Zero-Inflated Poisson) with multivariate panel structure. However, the evaluation protocol is standardized and the code is open-source. The main complexity lies in training and evaluating 15 diverse models across 30 environments with learning curve analyses. For practitioners, using the provided code significantly reduces implementation burden.

Code is available on GitHub (https://github.com/jiazeee/FinStressTS). The paper provides full parametric specifications for all 30 environments, standardized evaluation protocols, and reproducible experimental settings. All data-generating processes are fully specified with equations. The benchmark is designed as an open-source extensible artifact.