Page 39 - CCFA Journal - Seventh Issue

加中金融 机器学习 Machine Learning

Machine Learning in Banking – Derivative Pricing with Neural Network

机器学习在银行业的应用—神经网络与衍生品定价

Ernest Lok, Manager, RBC Group Risk Management，RBC 风险管理部经理

Executive Summary:
Neural network is a machine learning/deep learning technique used often in predictive
analytics. Due to the advancement in technology, they are widely used in different industries
these days. Today, we will explore one of its applications in capital markets.

This presentation covers 4 different things including the motivations and drawbacks of using
machine learning on derivative pricing, brief introduction to neural network, application of
Monte Carol Simulation on traditional derivative pricing and application of neural network in
pricing derivative.

【提要】神经网络是一种机器学习及深度学习的技术，在预测分析中被大量使用。 由
于技术的进步，近年来它已广泛应用于不同的行业。 今天，我们将探讨其在银行业资
本市场中的应用之一。

Motivations and Drawbacks of Using Machine Learning on Derivative Pricing
First of all, derivatives such as basket options, barrier options and other kinds of exotic options are often mispriced due to low trading
liquidity. As long as the machine learning predicted price lies within a reasonable range, it is a great way for pricing options. Second,
pricing derivatives with neural network saves time. Once the model is trained, instant results can be obtained by plugging in new sets
of values for the input variables. Compared to the traditional pricing method “Monte Carlo Simulation” which can take half a day to
update prices of the options, neural network is more efficient. Third, with neural network, we can approach derivative pricing in a
more objective way without making tons of assumptions about the financial mechanics, such as assuming Log-Normal Distributions
for Stock Prices and Normal Distributions for Stock Returns. Fourth, neural network is a wonderful tool to deal with non-linear data
and large number of inputs which are the basics of derivative pricing.

Neural networks also come with some drawbacks. For instance, a lot of the time will be spent on training and validating the model
depending on how much data we are analyzing and how many variables we will include. Also, periodic update is required as the
market is changing rapidly in such a fast-paced industry. Another well-known issue is that neural network is a black box. It is very
difficult to explain and interpret the results from neural network predictions. Since original inputs and operations are not directly
observable to the users, we are simply giving the model our sets of input values and in return we get the predicted outcomes from
the neural network. As the saying goes, “garbage in, garbage out”. It implies that neural network performance will highly depend on
the quality of the training data.

衍生品定价中使用机器学习的优点和缺点

首先，篮子期权、障碍期权和任何种类的奇异期权等衍生品由于交易流动性低而经常被错误定价。只要机器学习预测的价
格在合理范围内，它就是定价期权的好方法。其次，使用神经网络为衍生品定价可以节省时间。训练模型后，可以通过为
输入新的数据来获得即时结果。与传统的定价方法“蒙特卡洛模拟”需要大半天时间来更新期权价格相比，神经网络的效
率更高。第三，借助神经网络，我们可以以更客观的方式处理衍生品定价，而无需对数据做出大量假设，例如假设股票价
格的对数正态分布和股票收益的正态分布。第四，神
经网络是处理非线性数据和大量数据的绝佳工具，这
也是衍生品定价的基础。

神经网络也有一些缺点。 例如，大量时间将用于训练
和验证模型，具体时间取决于我们正在分析的数据量
以及我们将包含多少变量。 此外，由于市场瞬息万
变，模型因此需要定期更新。 另一个众所周知的问题
是神经网络是一个黑盒子。 这让人很难解释神经网络
预测的结果。 由于用户无法直接观察到原始输入和操
作，因此我们只是为模型提供了输入数据，并从神经
网络中获得了预测结果。俗话说，“垃圾进，垃圾
出”，这意味着神经网络的性能将高度依赖于训练数
据的质量。

CCFA JOURNAL OF FINANCE February 2022
Page 39 第39页