Research On The Forecasting Method Of Stock Risk Based On Bursty Event Detection

After the2008financial tsunami, both the governmental ministries and thescholars have paid more and more attentions to the problem of financial risks. Withthe rapid development of the social media networks, bursty events can be quicklyspread across such a tremendous social network, and forwarded by multiple users,the effect of such bursty events can then be largely amplified. This scenariostimulates a short-term, strong disturbance to the price of corresponding financialassets, which calls for a novel approach to estimate the financial risks brought by thebursty events.However, the traditional approaches cannot accurately predict the financialrisks if the bursty events occur. The reason lies in that the traditional approaches,such as macroscopic model and time series analysis, only learn common law ofstock volatility by performing analysis based on either fundamental factors or aseries of price change, which is proved to be ineffective when the strongdisturbances occur. Therefore, a novel approach which could identify the riskscontained by the breaking financial news is urgently needed. Proved by earlyresearchers, there exists the relationship between the stock price and the burstyevents. Based on this, this thesis proposed to forecast the stock risk by modeling therelationship between the topic distribution of bursty events and changes of stockprice. The main contents of this thesis are as follows:First, to bridge the association gap between bursty events and stocks, weproposed to adopt the Gaussian distribution to filter the news which is notco-occurrent with the abnormal stock price changes. Therefore, the many-to-manymapping relationship between stocks and bursty news could be identified.Second, to predict the stock risks brought by bursty events, we adopted theunsupervised topic model to cluster topics of bursty events. With the learnedmapping relationships, a random walk model was built to predict the stock risks.Last, to integrate the forward effect of the stock price change into the stock riskpredicting model, we proposed to train the supervised topic model simultaneouslymaking use of the coupling relationship between the topic distribution of burstyevents and the stock risks.Experimental results showed that the proposed approaches are more accuratethan the traditional stock risks forecasting algorithms when the bursty events occur.