Browsing by Author "Laakso, Tomi"

Flash crashes are one of the most prominent market inefficiencies which are recognized in stock and index prices. They violate the hypothesis of efficient markets and affect the real economy as well. Their proper forecasting has not been possible with conventional methods due to their seeming rarity and extremity. Furthermore, they are difficult to detect from the noise of price processes. By augmenting the HAR model with company-specific news data aim is to improve volatility estimates on days when these extreme events occur. These days are first identified from the price processes by a novel statistical method called V-statistic, which detects statistically significant flash crashes. Data is every fulfilled trade for six stocks from New York Stock Exchange, and company-specific news flow data which is obtained from RavenPack News Analytics service. Both data sets cover the years 2014 to 2016. HAR model is estimated for all six stocks with and without the news data and the in-sample-model estimates are compared both on the full data and on the days when flash crashes occur. Results are ambiguous but they give slight signs that news data could be useful for improving volatility estimates on the days when flash crashes occur. Model volatility estimates are better on average when augmented with news data. Mean absolute percentage error is 0.1% smaller on average across all entities when augmenting the model with news data. However, there are differences across companies on how much and if at all news data improves models' performance. In conclusion, further work is needed to verify the usefulness of news data in forecasting flash crashes.