Relevance
8/10
Highly relevant for quantitative trading as realized volatility forecasting is central to risk management, derivative pricing, and portfolio allocation. The paper provides practical guidance: (1) TSFMs are not a drop-in replacement for HAR benchmarks, (2) model selection within the TSFM class matters more than the TSFM-vs-econometric choice, (3) a simple TTM+Log-HAR combination enters the MCS for 98-100% of assets, eliminating the need for per-asset model selection. Directly applicable to volatility-based trading strategies, options pricing, and risk model calibration.
Implementation Complexity
5/10
Moderate complexity. The econometric benchmarks (HAR, Log-HAR, ARFIMA, ARMA, MEM) are straightforward to implement with standard statistical packages. The TSFM evaluation requires loading pretrained models from HuggingFace and running zero-shot inference, which is relatively simple. However, the full evaluation framework (walk-forward with rolling origin, QLIKE computation, DM tests, MCS with block bootstrap, MZ recalibration, GR fluctuation tests) adds significant implementation overhead. TTM runs on CPU (<1M params), making it accessible, but other models (Chronos-Bolt-B at 205M, TimesFM at 200M) require GPU resources.
Reproducibility
4/5
Replication code is openly available at https://github.com/Alessiobrini/tsfm-rv. All TSFM checkpoints are loaded from HuggingFace with specific model identifiers provided. The VOLARE dataset is available from the VOLARE project, though underlying tick data from Kibot is proprietary. Detailed model specifications, evaluation protocols, and robustness checks are fully documented. Minor limitation: exact pretraining data cutoffs for some models are not published.