The efficient market hypothesis states that an efficient market incorporates all available information to provide an accurate valuation of an asset. Presently investors and researchers attempt to forecast future returns (profit/loss if the asset is held for a certain period) and volatility (variance of the returns) of the asset based on past trading behaviour, and commonly ignore non-numerical information. It is almost impossible to forecast future returns for frequently traded assets such as stocks, bonds, and currencies, so many institutional investors prefer to forecast future volatility. Volatility is frequently used by traders and fund managers to measure the risk of continuing to own the asset. Most volatility forecasting models completely disregard the arrival of news and therefore theoretically violate the efficient market hypothesis. The aim of this research is to investigate how the inclusion of details of the arrival of asset specific news (news which is relevant to the asset) can improve the volatility forecasts of a model. The problem is that the efficient market hypothesis indicates that only new information will cause the market to react, and therefore it is necessary to determine whether the news contains any new information. Most news does not include any new information and therefore assuming all news will trigger abnormal market behaviour is unlikely to improve the performance of a model. Furthermore news which causes a shock, i.e., news which contains highly unexpected new information, will cause a greater change in volatility than news which contains expected information. Therefore to produce a model that factors in the arrival of news into volatility forecasts, it is beneficial to examine the content to predict the reaction to the news. This research combines the field of econometrics with machine learning and intelligent data analysis. All hypotheses tested within this thesis are tested on a large collection of stocks traded in the US, UK and Australia. To my knowledge, this is the largest dataset used for the types of experiments conducted in this thesis. In this thesis evidence is provided to suggest that asset specific news is correlated with abnormal returns, volatility, and volatility forecast errors. There is also evidence to suggest that abnormal volumes and trading activity correlate to asset specific news. This confirms the findings of previous studies though in most cases only a small dataset was used and often only one or two time series (i.e., return, volatility, volume etc.) were used. Furthermore many studies did not investigate the intraday effect of news (i.e., the reaction on the day the news was released). The studies which investigated the intraday effect tended to focus on macroeconomic news, which is scheduled and eagerly anticipated by investors. Therefore the behaviour is easier to detect that for asset specific news. It is demonstrated that the content of news can be used to forecast abnormal returns and forecast periods when the given volatility forecasting model exhibits abnormally large errors (the difference between the realised volatility and the volatility which the given model forecast) with a high degree of accuracy. This was achieved by analysing the content of past news which correlated with abnormal market behaviour. For this research a new method for ranking terms is introduced and demonstrated to be very effective. Previous studies have revealed that the content of news can be used to forecast abnormal returns but, to my knowledge, no study has investigated the volatility forecast error. Furthermore, most previous studies have used a small dataset, and to forecast at relatively low frequencies (most are daily, though one is hourly). To the best of my knowledge no previous study has use such a large dataset to predict the high frequency (as little as 5 minutes) market reaction to news. Nor has any previous study achieved classification accuracies as high as those achieved in this thesis. Finally, a news aware volatility forecasting model is produced and the evidence demonstrates that the performance is better than an alternative model which does not account for news under certain circumstances. Furthermore it is demonstrated that using the content of news to choose documents which are more likely to cause the market to react yields better forecasts. Very few researchers have included the arrival of news in a volatility forecasting model, and all of these have used small datasets. Furthermore, to my knowledge, none of these researchers have used the content of the news to choose news which is more likely to cause the market to react.
| Identifer | oai:union.ndltd.org:ADTP/265601 |
| Date | January 2008 |
| Creators | Robertson, Calum Stewart |
| Publisher | Queensland University of Technology |
| Source Sets | Australiasian Digital Theses Program |
| Detected Language | English |
| Rights | Copyright Calum Stewart Robertson |
Page generated in 0.0018 seconds