Operator Deep Smoothing for Implied Volatility

Dear Reviewer Xa1Y, we thank you for your review and for the clear suggestions of how to improve the paper. We reworked the relevant sections and included ablation studies, aiming at improving the explanation of our architectural choices and validating them. Please have a look at revised PDF and our summary below. ### In-Neighborhood Sets You mention: > The choice of in-neighborhood sets is arguably key model choice in this work. [...] This paper would benefit from an additional analysis on changing the $\bar{p}$ and $K$ parameters. To address this point, we added the new paragraph "The Choice of In-Neighborhoods" in Appendix C.2 (as well as an ablation study in Appendix C.6, see below). In it, we expand on the following points: - The truncation of the in-neighborhoods in $\rho$-direction by $\bar{\rho} = 0.3$ is motivated by concrete domain-specific considerations: The maximum distance between option slices (for our data), is around 0.269. Our choice then ensures slices to be able "communicate" with each other in pairs of two, which is enough to enforce the monotonicity constraint of Theorem 2.1 of the paper. On the other hand, it is not motivated to increase $\bar{\rho}$ above this level, since our deep architecture allows information to propagate over the whole length of the domain in the $\rho$-direction (but also we wouldn't expect the error to go up when doing so – we hypothesize the network will pick the relevant information even from a larger neighborhood; see the next point). Moreover, an ablation study for $\bar{\rho}$ would require retraining the whole GNO from scratch, which time-/resource-constraints forbid us at this point. - Our neighborhood sets are not limited in size in the $z$-direction; this is a defensive choice that, in particular (relating to previous point), shows that the GNO is able to extract the relevant information even from an untruncated in-neighborhood (along $z$). Moreover, this choice readies the model to arbitrarily extrapolate in $z$-direction. - We identified the subsample-size $K=50$ by manual experimentation. Prompted by your remarks, we performed an ablation study for different values of $K$, discussing the interesting results in newly introduced Appendix C.6 (adding Tables 6 and 7). The results ultimately validate our choice. - We add a description of our subsampling heuristic instead of solely referring to the code material. ### Loss Function Weighting You mention: > I would kindly ask the authors to include more detail how the weighting of the loss function has been performed. As it is stated right now, there is no information on how to select/adjust the weightings on a new problem. We reported in Appendix C.3 that we retrieved our weighting of the loss terms by manual experimentation, led by the findings of Ackerer et al. (2020). Indeed, this has allowed us to produce effectively arbitrage-free results. Prompted by your remark, we performed an ablation study for the weight of the arbitrage penalty terms and discuss its results in newly introduced Appendix C.7 (adding Tables 8 and 9). ### Other model choices You mention: > Again, there is little information about all the other model choices and/or instantiations. In particular, the choice of the kernel functions as MLP (as opposed to other choices) would be interesting for the reader. I'm also wondering whether the MLP can output negative values? [...] the choice of a MLP would benefit from further explanation/ablation. We use MLPs for all remaining components of the GNO to keep things simple: The universal approximation theorem for neural operators in Kovachki et al. (2023) is based on using plain MLPs as well, relying on their universality properties as approximators between finite-dimensional spaces. While we agree that it would be interesting to investigate the effectiveness of different architectures in this context, it has been our strategy to stick to straightforward GNO-components for this paper. As for the configuration of the MLPs, we used manual experimentation led by conventional wisdom: For example, for the kernel-MLP, we chose the number of nodes of the hidden layers at the power of two closest to three times the number of nodes of the input layer, and we picked GELU as hidden activation for its smoothness properties, and use no output activation since the kernel parametrizes a linear transformation and definitely should be able to take negative values (answering your question). Thank you very much again for your review which we believe has helped us substantially improve the transparent presentation of our methodology (and it is reassuring to see our model choices validated by the ablation studies you suggested).

Dear Reviewer o4xg, thank you very much for your review. It is encouraging that you find our paper well-written and easy to follow, and that you agree with the strengths of our method with respect to input subsampling, no-arbitrage conditions, and elimination of data pre-processing. As for mentioned weaknesses, however, we believe there might have been an oversight. You mention: > 1. The contribution is limited. It is focused on applying the Graph Neural Operator (GNO) architecture to the specific task of nowcasting implied volatility, without modifications to the operator itself. > 2. It's important to discuss various neural operators and clarify why the Graph Neural Operator (GNO) was chosen for this task. However, our employed graph neural operator is, in fact, modified compared to the conventional version, and we devote the paragraph "Interpolating Graph Neural Operator" in Section 3.2 to this (as well as Appendix B). More precisely, our modification consists of dropping the local linear transformation in the first layer and careful construction of the graph structure to enable data interpolation. We state: > We therefore propose a new architecture for operator deep smoothing leveraging GNOs’ unique ability to handle irregular mesh geometries. We use a purely non-local first layer (dropping the pointwise linear transformation), and use it to produce hidden state at all required output locations, enabling subsequent layers to retain their local transformations. (Page 7 of the paper.) While our tweak to the classical GNO architecture is subtle, it is crucially important to enable its use for interpolation tasks. We explain --and argue that this moreover addresses your second mentioned weakness--: > Various neural operator architectures exist, mostly arising from the kernel integral transform framework of Kovachki et al. (2023). Most prominently, these include Fourier neural operators (FNO), delivering state-of-the-art results on fixed grid data, as well as graph neural operators (GNO), able to handle arbitrary mesh geometries (both reviewed in Appendix A.1). While highly effective with documented universality, these neural operators are not directly applicable for interpolation tasks as their layers include a pointwise-applied linear transformation, which limits the output to the set of the input data locations. (Page 7 of the paper.) Finally, we want to reiterate on your claim that our contribution "is focused on applying the GNO architecture to the specific task of nowcasting implied volatility". However, the entire Section 3.2 is kept general and not specific to implied volatility. Instead, we introduce the concept of *operator deep smoothing* as the generally applicable idea of using discretization-invariant neural operators approximating the identity operator to solve irregular interpolation tasks and explain the required modifications necessary to current neural operator architectures. Our practical experiments and much of the rest of the paper then indeed remain limited to implied volatility smoothing, which however is a massively important task in the financial industry, and which we believe benefits dramatically from our foundational, data-focused operator approach and is highly suited to showcase its unique advantages. We hope the above explanations helped to dispel your doubts regarding our contribution. Please do let us know if any unclarities or suggestions for improvement remain. For example, we could expand our review of existing neural operator architectures in Appendix A, to support our argumentation in Section 3.2. Thank you very much!

Dear Reviewer GefV, thank you very much for your insightful review, including an exhaustive summary of the strengths of our work. As for your mentioned weaknesses, they have prompted us to rework portions of our paper and include additional material which we believe has improved and rounded off its presentation. We address in more detail below. ### Abstract Treatment of Discretization-Invariance You mention: > The explanation of discretization-invariance is largely theoretical. Providing concrete financial scenarios, such as handling abrupt market shifts, would better illustrate its practical significance and strengthen the narrative. Thank you for pointing this out. We have newly introduced Appendix C.8 with two new plots detailing the spatial and dimensional irregularity of observed options data. It includes Figure 9, which plots the time-to-expiry/log-moneyness trajectory of three options over their lifetime (including the Covid crash) as well as Figure 10 which graphs the development of the size of the input data (i.e. the number of listed options) over the training period. Both plots are meant to illustrate the significance of the discretization-invariance of neural operators in the applicational context of volatility smoothing. ### Clarity of Experimental Figures You mention: > Figures 3 and 4 have unclear legends and insu!cient annotations, which reduce their interpretability. Improving their clarity and labeling would make the results more accessible and impactful. Thank you for bringing this to our attention. We made adjustments to Figures 3 and 4: We improved the legends and reworked their captions, hoping to have eased their interpretation. ### Analysis of Computation Complexity & Synthetic Benchmark Comparisons We thank you for the detailed comments regarding complexity and comparisons with other models: > The paper benchmarks its approach against SVI [1] and Ackerer et al. [2] but does not include comparisons with other key methods, such as SSVI [3] and VAE-based approaches [4]. Incorporating these would provide a more comprehensive evaluation. Additionally, using synthetic data, as in [2], could further strengthen the experimental validation. > While the paper highlights the elimination of instance-by-instance recalibration, it lacks a detailed discussion on computational complexity, including runtime and memory requirements. These metrics are crucial for assessing the method’s practicality in high-frequency financial contexts Regarding these considerations, we would first like to point out that SSVI is an extension of the SVI parameterisation, by making the SVI parameters functions of time-to-expiry. In other words, SSVI is a smooth arbitrage-free interpolation scheme for SVI slices that produces an entire surface with fewer parameters. In this sense, SSVI is in fact a subset of SVI, and any SSVI-benchmark is beaten by the SVI-benchmark, which explains why we focused solely on the more general SVI class. You rightly point out that our paper lacks the VAE-method as a benchmark. While [4] is analogous to our method, it is not applicable for index options data, which includes our S&P 500 data: "Standard [VAE] methods are limited to certain option markets (e.g. FX markets where options spread out on fixed rectilinear grids)" (Section 1, page 2). It is a unique feature of our method to be able to handle the very high-dimensional and variable-size nature of the S&P 500 input data. [4] uses a fixed input-size of 40, while the input size of our data increased from around 700 in 2012 to over 3000 in 2021 (this development is plotted in newly introduced Figure 10). We argue that this qualitative comparison is the most impactful and compelling argument in favor of our approach, while resource constraints at this stage forbid us the pursuit of retraining the GNO on synthetic data, which would have made a direct quantitative comparison possible. Finally, we agree that the implementation of a highly optimized variant (our code has prioritized prototyping and feasibility) from which to draw robust statements about HFT feasibility would be an important next step.