Forecasting S&P 100 Implied Volatility Using Artificial Neural Network

Roongroje, Thipawan (2009) Forecasting S&P 100 Implied Volatility Using Artificial Neural Network. [Dissertation (University of Nottingham only)] (Unpublished)

[img] PDF - Registered users only - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Download (1MB)

Abstract

Volatility forecast is an important task in financial markets. It has held the most attention among academics and practitioners over the last decades. A good forecast of the volatility of the asset prices over the investment period is a good starting point for assessing investment risk.

Most traders consider the implied volatility calculated by Black-Scholes model a significant tool used in signaling price movements in the underlying market. The ability to forecast the volatility of the market is equivalent to the ability to establish the proper strategic position in an anticipation of changes in the market trends. To be able to forecast correctly the market volatility, therefore, is critical to traders and analysts.

Over the last decades, many statistical models are invented for volatility forecasting purpose. Those models include time series analysis, ARCH-class models, forecasting based on implied volatility using Black-Scholes model, etc. Not long after a technology of artificial intelligence has been studied, a neural network, which one of learning algorithm that mimic the ability to learn from human brain, are applied in volatility forecasting task. This method became popular and widely studied by many researchers. Many studies have shown that neural network approach outperformed the conventional time series analysis and linear regression forecasting model.

In this paper, we aim to develop a multi-layer perceptron-based model for forecasting future implied volatility of the S&P 100 (OEX) index by using past volatilities and various option market factors. Trial-and-error method is employed to explore the efficient network architecture. This including find input variables that have the most contribution to the target output, number of neurons required in the hidden layers. The final network selected is the one that have the highest out-of-sample forecasting power.

Item Type: Dissertation (University of Nottingham only)
Depositing User: EP, Services
Date Deposited: 05 Feb 2010 14:21
Last Modified: 23 Oct 2016 08:27
URI: http://eprints.nottingham.ac.uk/id/eprint/23196

Actions (Archive Staff Only)

Edit View Edit View