多任務類神經網路於高效的美式選擇權定價

Abstract

傳統美式選擇權定價方法的計算效率不足。本論文使用類神經網路推理來定價美式選擇權，並提出多任務學習以提升準確度。此外，我們採用神經架構搜索技術來自動選擇有效的網路架構。研究發現，我們的方法能在推理階段減少CPU時間，且提供跟傳統方法相當的結果。

Conventional methods for pricing American options can be computationally inefficient. This thesis uses neural network inference to price American options. We propose multi-task learning to enhance accuracy. Furthermore, neural architecture search is adopted to automate choosing a competitive neural network architecture. This thesis finds that our methods not only provide results comparable to conventional ones but also require less CPU time during inference.