Dear all,
In financial time series specially of stock market we have to deal with data before modelling for missing values and outliers.
Missing values for any one day can be found out by averaging of values nearer that required value. But what we do in case more then 12 days data is missed in any series?
(suppose i have closing price data from 1st to 30th date for any month but in this month, data for date 11th to 17th is missed. what we do over here whether we should insert the same value of date 10th in these days from date 11 to 17.)
Before modelling of time series data we have to check for outliers also in series.
What are the strategies to detect them in excel and Eviews?
How can these outliers be treated except to remove them from data?
So that data can be used for modelling.
Hello
Presence of outliers is an indication of structural breaks in the data. Plotting the time series data can be an easy and informal way of detecting the outliers. However one can apply the formal test like chow test( in which case break date is single and known ) or bai- perron test (when there are multiple breaks and break date is unknown and either there are additive or innovative outliers). One can apply the dummies to take account of presence of outlier values in the data for the comparative purposes.
In your case, I would first try to find out whether the data show a weekly pattern, which should not be too difficult. If there is such a pattern, data could be created by taking the averages of the weeks before and after the missing data. Nevertheless, for a time series regression, one should include dummies for all days of this time span. The number of dummies could be reduced, if the coefficients for several day dummies are simmilar. An extreme dummy coefficient could be a hint that something special happened on this day.
For the missing values you can, e.g. use ARIMA. Whatever method you will use, you have to be aware of its assumptions, i.e. limitations.
As for outliers, one easy statistic is Cooks distances. Another, is to standardize the data and everything that is above 3 standard deviations can be considered an outlier.
Prior to removing any points in data you should consider consequences. And so, if you have time series, removing data is a breach in the series, e.g. it is like have a week without a Thursday in it - does not make much sense. A better solution is to search for variables that can allow you to account for the outliers, e.g. a specific date, season, event etc.
As to missing values (of the dependent variable - "y"), there is some literature - try searching on "time series imputation". In principle, your imputation method can be informed by the model that you believe applies. For example, if your model includes exogenous variables then their values during the missing block are relevant. Maybe this topic can be usefully seen as an example of forecasting - you want to forecast the missing values of y on the basis of whatever relevant information you do have. If your model is dynamic (i.e. using lags of y) then values of y after the missing block are also relevant (as well as values before the missing block). In this case you may combine "backcasting" with forecasting. The averaging of the two y-values at the edge of the missing block is a simple example of combining forecasting with backcasting but it potentially ignores much relevant information. The larger is the block of missing values, the greater is (in theory) the motivation for employing a better-informed method of imputation.
One should not remove outliers, unless they are clearly caused by computational or measurement errors. Outliers have, in general, more informational content than "normal" data. They can give you a hint that there is some important variable missing in your model or that there could be discontinuities in relations. In any case, one should try to find a good explanation for outliers. One way would be Tomasz's proposal to "search for variables that can allow you to account for the outliers". If not possible (e.g. because it is a unique or non-measurable event), I would prefer to use dummies in econometric estimations, instead of just remove data.
Dear,
First of all, we have to differentiate between three issues (1. outliers, 2. Missing, 3. Structural Breaks), for the outliers, you can downloading in add-ins icon in the main menu in the E-views desktop (Trimming and Winsorising), and you will overcome the value of the outlier by using Trimming, and then you can to compensate outlier value through Winsorising.
In addition, to make sure that you overcome the issue of outliers, open as a group the three series (original series, series solve outlier problem, series after solve missing data problem), then plot a graph for that using Boxplot, and you will see the results.
for missing value, in E-views you click on (pro) in the object window, then click on interpolation, and using linear methods, and to make sure that you overcome the issue of missing value, open as a group the two series (original series with missing value, series after compensating the missing value), and then you will see the results.
in terms of the structural Breaks, you can use the chow breakpoint test and then using dummy variables to solving the third issue of structural breaks.
Good luck.
Agree withDr Asaad one suggestion.If outliers are many then data isa suspect. Also ifdelete one or two ouiers in the middle use pro in Eview fir those deleted data or use extraploation.
- Badges
- Science topic
- Similar topics
- Social Psychology
- Attitude