Arbitrage-Free IV Surface Forecasting

• Arbitrage-free IV surface forecasting is a set of techniques that predict future implied volatility surfaces while strictly enforcing no-arbitrage conditions.
• It combines deep learning, stochastic generative models, and manifold learning to deliver high-fidelity forecasts consistent with market constraints.
• These methods enhance risk management and derivative pricing by ensuring compliance with calendar-spread and butterfly arbitrage criteria.

Arbitrage-free implied volatility (IV) surface forecasting refers to the prediction of the future shape and dynamics of the IV surface while rigorously enforcing financial constraints that preclude static arbitrage opportunities—specifically, calendar-spread and butterfly-spread arbitrage. The accurate, dynamic, and arbitrage-consistent modeling of IV surfaces is fundamental to risk management, derivative pricing, and trading. Recent advancements leverage stochastic generative modeling, deep learning with explicit constraints, and manifold learning, resulting in tractable, high-fidelity arbitrage-free forecasts suitable for both short- and long-horizon simulation.

1. Core Principles and No-Arbitrage Criteria

The IV surface, indexed by moneyness $m$ (or log-strike) and time-to-maturity $\tau$, must satisfy analytic and economic properties to avoid static arbitrage:

• Calendar-spread arbitrage: Total variance $w(m,\tau) = \sigma^2(m,\tau)\tau$ must be non-decreasing in $\tau$, i.e., $\partial_\tau w(m,\tau) \ge 0$.
• Butterfly arbitrage: For each $\tau$, the IV slice must be convex in strike. In discrete settings, finite-difference approximations of convexity are used, while in continuous representation, the Durrleman condition quantifies admissibility.
• Large-moneyness behavior: Linear or sublinear tails are necessary to prevent negative densities as $|m| \to \infty$.

These conditions are central in parametric surface models (e.g., SSVI [Gatheral & Jacquier]) and hold for practically all stochastic and machine-learning-based surface modeling approaches (Jin et al., 10 Nov 2025, Andrès et al., 2023, Ackerer et al., 2019, Zhang et al., 2021).

2. Stochastic Generative Approaches

2.1 Denoising Diffusion Probabilistic Models (DDPMs)

Recent research has established conditional DDPMs as a robust class of arbitrage-free generative models for IV surface forecasting (Jin et al., 10 Nov 2025). The architecture comprises a forward noising process (forward SDE) and a learned reverse SDE (parameterized by a neural score network). Conditioning is performed on market features such as exponentially weighted moving averages (EWMAs) of past surfaces, returns, squared returns, and volatility indices (e.g., VIX).

The key innovation is a composite loss: $$\mathcal{L}(\theta) = \mathcal{L}_\mathrm{MSE}(\theta) + \lambda\,\mathcal{L}_\mathrm{arb,\,weighted}(\theta)$$ where $\mathcal{L}_\mathrm{MSE}$ is the standard diffusion score-matching objective, and $\mathcal{L}_\mathrm{arb,\,weighted}$ is a dynamic arbitrage penalty, SNR-weighted to avoid noisy gradients in early diffusion steps. The arbitrage penalty enforces calendar and butterfly constraints at the surface level.

A formal convergence analysis shows the introduction of this penalty induces a controllable $O(\lambda^2)$ bias in the trained model, leaving predictive distribution almost unaltered but steering outputs onto the arbitrage-free manifold.

2.2 Variational Autoencoders Conditioned on SDE Models

Hybrid methods, notably VAEs trained on parameter vectors of arbitrage-free SDE models, ensure intrinsically arbitrage-free surfaces by construction (Ning et al., 2021). Here, each day's observed surface is calibrated to a parametric SDE (e.g., regime-switching or Lévy additive process), and VAE models the empirical distribution on the space of calibrated parameters. The generative process samples from the VAE's latent prior, decodes to model parameters, and reconstructs the option price surface (and hence IV surface) via characteristic function inversion.

This ensures that both in-sample and synthetic forecasts are strictly arbitrage-free, as the SDE itself generates only admissible option price functions. Conditional VAE extensions incorporate market observables such as VIX for scenario-conditioning (Ning et al., 2021).

2.3 Tangent Lévy Model Simulations

Tangent Lévy models project current market prices onto a flexible Lévy model and evolve the local Lévy density $\kappa_t(T,x)$ via stochastic dynamics constrained by drift-restriction and compensator-matching (Carmona–Nadtochiy theorem), guaranteeing dynamic absence of both calendar and butterfly arbitrage (Carmona et al., 2015). The approach combines static calibration to current surfaces with PCA-based volatility factor extraction for dynamic evolution.

Key advantages include arbitrage-free surface evolution for any simulation horizon and accurate joint simulation of IV surface paths compatible with portfolio risk optimization.

3. Deep Learning Methods with Explicit Arbitrage Penalties

3.1 Deep Smoothing Neural Networks

Deep smoothing architectures model total variance as a product of a parametric prior (e.g., SSVI) and a neural corrector. No-arbitrage is imposed by adding soft penalty terms for calendar and butterfly conditions across a dense evaluation grid. The composite loss incorporates both data fit and arbitrage penalties, and the network can be used as a universal plug-in to correct classical models (Ackerer et al., 2019).

Extensive benchmarking confirms that for realistic penalty weights ($\lambda$), the neural model predicts accurate, smooth, and arbitrage-free IV surfaces, outperforming SSVI benchmarks, with test RMSE $\approx 0.5\%$ (versus SSVI's 4–7%).

3.2 Two-Step Time Series and Reconstruction Frameworks

Another established methodology decomposes the IV surface forecasting into two steps: (1) feature extraction and time-series forecasting, and (2) arbitrage-free surface reconstruction (Zhang et al., 2021). Feature vectors (via sampling, PCA, or VAE) are predicted using LSTMs, and a surface-valued deep network reconstructs the full IV surface while penalizing arbitrage violations.

This ensemble delivers RMSE $\approx 0.025$ out-of-sample and provable elimination of arbitrage violations by design. The modular separation allows for enrichment with volatility factors, market returns, or macro variables, and supports scenario simulation via innovations in the latent space.

4. Path-Dependence and Structural Manifold Approaches

Contemporary research has established that the dynamics of the IV surface—especially at-the-money and term-structure slopes—display explicit path-dependence on historical underlying returns and squared returns, extending over multi-year horizons (Andrès et al., 2023). Parsimonious SSVI parameterizations with feedback rules on historical returns explain 62–77% of out-of-sample ATM vol variation (over 1M-2Y), with full-surface MAPE $\sim 1.7\%$.

Key features:

• Feedback models for SSVI parameters controlled by trend and volatility clustering features over up to four years of past price evolution.
• Residuals governed by semi-Markov diffusions capture regime clustering.
• Strict enforcement of SSVI no-arbitrage parameter inequalities at each step ensures validity of all simulated paths, even under regime switches.

5. High-Frequency and Functional Principal Component Models

High-frequency forecasting is addressed via multivariate Hawkes processes for jumps in IV grid points, with parameter reduction and scaling constructs delivering arbitrage by design (Baldacci, 2020). When correctly parameterized, these models capture the impact of order-flow clustering and risk-factor cross-impact, with empirical validation from market-making backtests.

Functional principal component and neural SDE methods (Choudhary et al., 2023) offer another dimension, where the entire IV surface is projected onto a finite FPCA basis, and the time evolution of FPC coefficients is governed by a (non-Markovian) neural SDE with GRU-drift and diffusion. Without hard-coding constraints, simulated surfaces lie in the empirical arbitrage-free manifold, verified by calendar and butterfly metrics on all simulated scenarios.

6. Empirical Performance and Calibration

Method/Model	Empirical RMSE (Out-of-Sample)	Arbitrage Compliance	Special Features
DDPM + SNR-Penalty (Jin et al., 10 Nov 2025)	3.00% (MAPE)	Tight: $\mathbb{E}[\Phi]$ near minimum	Confidence intervals calibrated
VAE on SDE params (Ning et al., 2021)	$W_1$ 0.037–0.053 (AUDUSD)	Guaranteed by SDE design	Conditional VIX support
Deep smoothing NN (Ackerer et al., 2019)	0.5% (RMSE)	Zero/near-zero violation	SSVI/BS prior; rapid updating
Two-step LSTM + DNN (Zhang et al., 2021)	0.0245 (RMSE), 9.90% (MAPE)	Zero observed violation	VAE/PCA/Sampling for features
Path-dependent SSVI (Andrès et al., 2023)	1.6% (surface rel. error)	Tight inequalities during simulation	4-parameter, regime feedback
Tangent Lévy (Carmona et al., 2015)	Portfolio risk-improvement	By HJM-type drift constraints	Dynamic surface simulation
Hawkes-based (Baldacci, 2020)	Not given (PnL/risk backtest)	Blockwise scaling: strict	High-frequency, cluster effects
FuNVol (Choudhary et al., 2023)	0.006–0.008 (median RMSE)	Historic-level or fewer violations	Multi-asset, neural SDE

Empirical metrics as reported in the referenced works. All models above achieve strong compliance with static arbitrage, with several guaranteeing strict enforcement by design.

7. Directions, Limitations, and Extensions

The contemporary consensus is that arbitrage-free IV surface forecasting requires a balance of modeling market microstructure, feature-rich conditioning (macroeconomic indicators, prior returns, volatility clustering), and explicit or structural regularization aligned with financial no-arbitrage. Extensions are being pursued in:

• Integrated pathwise and feature-conditional generative models for extreme scenario simulation (Jin et al., 10 Nov 2025, Andrès et al., 2023).
• Multi-asset and joint risk-factor modeling (Neural SDEs on concatenated FPCA bases) (Choudhary et al., 2023).
• Higher-order dynamic arbitrage constraints (beyond static) and "pathwise" arbitrage-restoring projections (Zhang et al., 2021).
• Robust uncertainty quantification—including ensemble, posterior last-layer, and MC-dropout algorithms—remains crucial in data-sparse and extrapolation regimes (Ackerer et al., 2019).
• Unified frameworks for high-frequency tick-level and daily grid prediction.

The field converges toward flexible, conditional, and simulation-compatible models that are both computationally tractable and enforce a high degree of financial admissibility, with empirical superiority to classical interpolating or parametric families. Importantly, the arbitrage-free manifold approach—projecting generative processes onto the space of admissible surfaces—remains at the core of state-of-the-art forecasting paradigms.