HARNet: A Convolutional Neural Network for Realized Volatility Forecasting¶
Bibliographic record. Follow the original-source link for the publication.
| Field | Value |
|---|---|
| Primary domain | Volatility |
| Other domains | — |
| Methods | Financial Ml |
| Facets | — |
| Authors | Rafael Reisenhofer, Xandro Bayer, Nikolaus Hautsch |
| Published | 2022-05-16 |
| Source | arXiv Quantitative Finance History |
| Identifiers | arxiv:2205.07719 |
| URL | Open original source |
Editorial synthesis¶
Why it matters¶
This work is operationally relevant because it bridges HAR and CNN forecasting, yet claims of substantial gains should be scrutinized for sample dependence and hyperparameter sensitivity before productization. (abstract:S1, abstract:S2, abstract:S3, abstract:S5, abstract:S7)
Main author claims¶
- The authors claim HARNet uses dilated convolutions for exponential receptive-field growth while retaining HAR-equivalent predictions before optimization. (
abstract:S3,abstract:S4) - They claim QLIKE-driven optimization is more stable and that HARNets substantially improve forecasting accuracy over the HAR baselines across three indices. (
abstract:S5,abstract:S6,abstract:S7) - The authors report: Qualitative filter analysis indicates yesterday’s volatility contributes most among recent days and monthly lag effects decay roughly linearly. (
abstract:S8,abstract:S9,abstract:S10)
Data, method, or discussion scope¶
Evidence is confined to abstract-level claims, with mention of three stock indexes but no disclosed asset codes, frequencies, loss decomposition, or hyperparameter search details. (abstract:S6, abstract:S7, abstract:S8)
Main limitations¶
There is no training/validation protocol, compute-cost context, overfitting controls, or trading-cost linkage; qualitative performance claims omit explicit decision thresholds. (abstract:S4, abstract:S5, abstract:S7, abstract:S6)
Relationships¶
- None recorded.