讲座背景

• The 4th Women in Quantitative Finance Conference (WQF)
• 2021-06-14 Deep Hedging under Rough Volatility by Blanka Horvath Lecturer, King’s College London and Researcher, The Alan Turing Institute
• 这个一小时的讲座主要基于Generating Financial Markets with Signatures，已于2021.06.09在Risk杂志刊登 （pdf文档可以在这里下载)。这个讲座主要基于这篇文章，自己在2021.06.13的博文中也把这篇文章列入to read list。

讲座摘要

• Classical QF vs Deep Model Architectures
  • Classical: (Program or Algo; Data) => output
  • • e.g. SABR model in pricing library, match the algorithms with the data in the market
  • Now: Model = (Architecture, ObjF; TrainData) => Program, (Program, TestData) => output
  • • e.g. Machine learning architecture, choose objective function (more art than science, need to under the problem to design suitable ObjF), TrainData (data-driven aspect, important for the performance of the algorithms afterwards, motivation for data generator) .
  • • Architecture, ObjF combined as Network;
  • • Quality of training data shapes the DNN (and its performance)!
  • • What kind of training data we should use, what is the training data that will prepare our algorithms ideally for the scenarios that they will be facing in real?
  • • Test data is real life data that algorithm will face in the market.
• What are the challenges that we use the historical data as test data?
  • Historical data is one realisation of the past, just represents one out of many many realisations.
  • Unless make some extreme assumptions: e.g., market is stationary which is not the case most of times; We observe the evolution in the past does not mean that it works like this in the future.
• 个人认为这也是么trainset, forward testing重要以及model risk需要评估的原因。