In real markets, generating a smooth implied volatility surface requires an interpolation of the calibrated parameters by using smooth parametric functions. For this interpolation, practitioners do not use all the discrete parameter points but manually select candidate parameter points through time-consuming adjustments (e.g., removing outliers, comparing with the surface from the previous day, and considering daily market indexes) to generate a smooth and robust surface. In this paper, we propose neural network models that assist practitioners in generating a smooth implied volatility surface under the SABR (Hagan et al., 2002) model. Utilizing the self-attention mechanism of a transformer network (Vaswani et al., 2017) as a backbone network, we design two models: one that orders the parameter points by their likelihood to be selected as candidate parameter points and one that determines the candidate point set among the combinations of high-priority points. Experimental results from a 3-year period of real market S&P500 and KOSPI200 data show that the combination of two models can assist practitioners in the point selection task.

在实际市场中，要生成平滑的隐含波动率表面，需要使用平滑的参数函数对校准后的参数进行插值。对于此插值，从业人员不会使用所有离散的参数点，而是通过耗时的调整手动选择候选参数点（例如，删除异常值，与前一天的表面进行比较并考虑每日市场指数）以生成平滑且坚固的表面。在本文中，我们提出了一种神经网络模型，可以帮助从业人员在SABR模型下生成平滑的隐含波动率表面（Hagan等，2002）。利用自我关注机制以变​​压器网络（Vaswani et al。，2017）作为骨干网络，我们设计了两种模型：一种模型按照参数似然性将参数点选择为候选参数点，另一种模型确定以下参数组合中的候选点集：高优先级点。来自真实市场S＆P500和KOSPI200数据的3年期的实验结果表明，两种模型的组合可以帮助从业人员完成选点任务。