Research map / Architectures and efficiency / RNNs and normalization
HyperIV: Real-time Implied Volatility Smoothing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HyperIV, a hypernetwork that maps 9 sparse option quotes to a compact MLP generating an implied volatility surface in about 2 ms, with auxiliary losses encouraging no static arbitrage.
Problem
Constructing complete implied volatility surfaces in real time from very sparse (fewer than 10) option observations at high frequency, where traditional calibration is too slow and data is scarce.
Approach
A hypernetwork (Transformer set encoder with mean pooling) takes a reference set of 9 option contracts and outputs the weights of a tiny MLP (337 parameters) that defines the volatility surface over log-moneyness and maturity, so no per-interval calibration is needed; arbitrage-free conditions are enforced via auxiliary loss terms rather than hard constraints.
Key results
- HyperIV achieves the lowest implied volatility MAE on 7 of 8 index datasets (e.g., VIX 0.0336 vs. GNO 0.1345 and SSVI 0.2771); GNO is better only on 1-min SPX (0.0140 vs. 0.0167).
- HyperIV constructs a 10,000-point surface in 2.09 ms on an A100 GPU using only 9 reference contracts, with a 337-parameter surface network.
- Auxiliary arbitrage losses are around 1e-8 in training and 1e-6 at test, but the paper says it provides no hard arbitrage-free guarantee.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HyperIV instead of GNO | SPX 1-min implied volatility smoothing · SPX 1-min | MAE (implied volatility): 0.0140 -> 0.0167 | Hurt |
| Use HyperIV instead of the strongest baseline | NDX 1-min implied volatility smoothing · NDX 1-min | MAE (implied volatility): 0.0162 -> 0.0156 | Helped |
| Use HyperIV instead of the strongest baseline | SPX, NDX, RUT, VIX, MXWLD, and MXEF 1-day implied volatility smoothing · 1-day index options | MAE (implied volatility) | Helped |
| Use HyperIV instead of the strongest baseline | SPX 1-min price smoothing · SPX 1-min | MAE (price): 2.2398 -> 2.5115 | Hurt |
| Use HyperIV instead of GNO | NDX 1-min price smoothing · NDX 1-min | MAE (price): 10.2187 -> 10.0043 | Helped |
| Use HyperIV instead of the strongest baseline | 1-day option-price smoothing · SPX, NDX, RUT, VIX, MXWLD, and MXEF 1-day | MAE (price) | Helped |
| Use HyperIV instead of SSVI | MXWLD and MXEF implied-volatility smoothing · MXWLD and MXEF 1-day | MAE (implied volatility) | Helped |
| Use HyperIV instead of GNO | NDX-to-SPX cross-asset transfer · NDX→SPX 1-min | MAE: 0.1272 -> 0.0349 | Helped |
| Use HyperIV instead of GNO | SPX-to-NDX cross-asset transfer · SPX→NDX 1-min | MAE: 0.0185 -> 0.0223 | Hurt |
Surprising findings
- SSVI, with only four parameters, outperforms the 300K-parameter VAE on MXWLD and MXEF.
- GNO fails catastrophically on the reverse cross-asset direction NDX to SPX (MAE 0.1272), and on VIX and MXEF.
Limitations
- Requires GPU or NPU, unlike CPU-based SSVI; inference slows to 8 ms on an M2 chip.
- Arbitrage-freeness is enforced only via auxiliary loss, not hard constraints, and the method lacks interpretability of parameters.
- Data is licence-restricted and cannot be redistributed; cross-asset generalization fails for VIX.
Experimental setup
Datasets: OptionMetrics SPX/NDX/RUT/VIX index options (1-min and end-of-day) MSCI World and MSCI Emerging Markets index options
Benchmarks: Test MAE on implied volatility across 8 index option datasets Test MAE on option prices Cross-asset generalization MAE
Metrics: mean absolute error (implied volatility) mean absolute error (price) testing time (ms) training time memory
Baselines: SSVI VAE GNO
Builds on: Hypernetworks (Ha et al., 2016) SSVI Graph neural operator (Gonon et al., 2024) VAE calibration-free surface (Bergeron et al., 2021)
Closest papers by meaning
- Operator Deep Smoothing for Implied Volatility (ICLR 2025 · similarity 0.89)
- StockMixer: A Simple Yet Strong MLP-Based Architecture for Stock Price Forecasting (AAAI 2024 · similarity 0.85)
- EarnHFT: Efficient Hierarchical Reinforcement Learning for High Frequency Trading (AAAI 2024 · similarity 0.83)
- FactorGCL: A Hypergraph-Based Factor Model with Temporal Residual Contrastive Learning for Stock Returns Prediction (AAAI 2025 · similarity 0.83)
- R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization (NeurIPS 2025 · similarity 0.83)
- Market-GAN: Adding Control to Financial Market Data Generation with Semantic Context (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.