This [unedited] guest post is by a student in my PSTAT262MC class (background post). Please praise/critique/comment on its quality and importance to you.
varvara [thumbnails].jpgVarvara Kulikova says:

In my previous post the multivariate time series are presented for three stocks.

To model this high frequency data a simple multivariate Dynamic Linear Model had been used in the form of SUTSE (Seemingly unrelated Time Series Equations) model with linear trend component. Unknown parameters in the model are observation variance-covarince and evolution variance-covarince matrices. To perform the analysis a Bayesian approach in the form of Gibbs sampler had been employed. Prior distributions for the parameters are wishart with initial values and certainty defined based on the data. Missing observations had been imputed for time points with at least one missing observation (out of three) in the Gibbs sampler aswell.

Forecast (in red) for 5 seconds ahead with corresponding confidence intervals (green dashed lines) is illustrated on the picture along with actual values for these future 5 seconds (in black). The forecast is based on the first 300 observations (5min) of the original dataset. Values for DELL,IBM and MSFT stock prices at 9:35:01am EST with 95% confidence bounds are 15.21492 (15.14671,15.28314) , 119.4038 (119.3249, 119.4828), 25.40109 (25.32586,25.47633). Although, the actual values at 9:35:01am EST are NA (DELL), 119.4233 (IBM) and 25.42 (MSFT).


blog comments powered by Disqus


18 March 2010