如何處理時間序列中的日期間隔 (with gaps)問題？

01-29

Author: Ashish Rajbhandari, Senior Econometrician
Title: Handling gaps in time series using business calendars
Source: Stata blogs

轉載+部分翻譯：Stata連享會

編者按：
在分析時間序列資料時(如股票收益數據)，由於在周末或重要節日里休市，導致日期數據往往是不連續的。若使用 Stata 默認的日期格式，會導致我們無法連續地計算收益率。為此，我們應該做些適當的調整，而不是把這些差距看作是缺失的值。這這篇推文中，作者使用 Stata 的商業日曆舉例說明了處理不規則間隔的日期一個簡便方法。

Introduction

Time-series data, such as financial data, often have known gaps because there are no observations on days such as weekends or holidays. Using regular Stata datetime formats with time-series data that have gaps can result in misleading analysis. Rather than treating these gaps as missing values, we should adjust our calculations appropriately. I illustrate a convenient way to work with irregularly spaced dates by using Stata』s business calendars.

In nasdaq.dta, I have daily data on the NASDAQ index from February 5, 1971 to March 23, 2015 that I downloaded from the St. Louis Federal Reserve Economic Database (FRED).

. use http://www.stata.com/data/nasdaqnn. describennContains data from http://www.stata.com/data/nasdaq.dtan obs: 11,132 n vars: 2 29 Jan 2016 16:21n size: 155,848 n-------------------------------------------------------------------------------n storage display valuenvariable name type format label variable labeln-------------------------------------------------------------------------------ndate str10 %10s Daily datenindex float %9.0g NASDAQ Composite Index (1971=100)n-------------------------------------------------------------------------------nSorted by: n

date is the time variable in our data, which is a string format ordered as year, month, and day. I use the date() function to convert the string daily date to a Stata numeric date and store the values in mydate. To find out more about converting string dates to numeric, you can read A tour of datetime in Stata.

. generate mydate = date(date,"YMD")n. format %td mydaten

I tsset these data with mydate as the time variable and then list the first five observations, along with the first lag of index.

. tsset mydaten time variable: mydate, 05feb1971 to 23mar2015, but with gapsn delta: 1 daynn. list date mydate index l.index in 1/5nn +------------------------------------------+n | L.|n | date mydate index index |n |------------------------------------------|n 1. | 1971-02-05 05feb1971 100 . |n 2. | 1971-02-08 08feb1971 100.84 . |n 3. | 1971-02-09 09feb1971 100.76 100.84 |n 4. | 1971-02-10 10feb1971 100.69 100.76 |n 5. | 1971-02-11 11feb1971 101.45 100.69 |n +------------------------------------------+n

The first observation on l.index is missing; I expect this because there are no observations prior to the first observation on index. However, the second observation on l.index is also missing. As you may have already noticed, the dates are irregularly spaced in my dataset—the first observation corresponds to a Friday and the second observation to a Monday.

I get missing data in this case because mydate is a regular date, and tsset–ing by a regular date will treat all weekends and other holidays as if they are missing in the dataset instead of ignoring them in calculations. To avoid the problem of gaps inherent in business data, I can create a business calendar. Business calendars specify which dates are omitted. For daily financial data, a business calendar specifies the weekends and holidays for which the markets were closed.

Creating business calendars

Business calendars are defined in files named calname.stbcal. You can create your own calendars, use the ones provided by StataCorp, or obtain them directly from other users or via the SSC. Calendars can also be created automatically from the current dataset using the bcal create command.

Every stbcal-file requires you to specify the following four things:

the version of Stata being used
the range of the calendar
the center date of the calendar
the dates to be omitted

I begin by creating nasdaq.stbcal, which will omit Saturdays and Sundays of every month. I do this using the Do-file editor, but you can use any text editor.

version 14.1npurpose "Converting daily financial data into business calendar dates"ndateformat dmynrange 05feb1971 23mar2015ncenterdate 05feb1971nomit dayofweek (Sa Su)n

The first line specifies the current version of Stata I am using.
The second line is optional, but the text typed there will display if I type bcal describe nasdaq and is good for record keeping when I have multiple calenders.
Line 3 specifies the display date format and is also optional.
Line 4 specifies the range of dates in the dataset.
Line 5 specifies the center of the date to be 05feb1971.

I picked the first date in the sample, but I could have picked any date in the range specified for the business calendar. centerdate does not mean choosing a date that is in fact the center of the sample. For example, Stata』s default %td calendar uses 01jan1960 as its center.

The last statement specifies to omit weekends of every month. Later, I will show several variations of the omit command to omit other holidays. Once I have a business calendar, I can use this to convert regular dates to business dates, share this file with colleagues, and also make further changes to my calendar.

Using a business calendar

. bcal load nasdaqnloading ./nasdaq.stbcal ...nn 1. version 14.1n 2. purpose "Converting daily financial data into business calendar dates"n 3. dateformat dmyn 4. range 05feb1971 23mar2015n 5. centerdate 05feb1971n 6. omit dayofweek (Sa Su)n(calendar loaded successfully)nn. generate bcaldate = bofd("nasdaq",mydate)n. assert !missing(bcaldate) if !missing(mydate)n

To create business dates using bofd(), I specified two arguments: the name of the business calendar and the name of the variable containing regular dates. The assert statement verifies that all dates recorded in mydate appear in the business calendar. This is a way of checking that I created my calendar for the complete date range—the bofd() function returns a missing value when mydate does not appear on the specified calendar.

Business dates have a specific display format, %tbcalname, which in my case is %tbnasdaq. In order to display business dates in a Stata date format I will apply this format to bcaldate just as I would for a regular date.

. format %tbnasdaq bcaldatenn. list in 1/5nn +---------------------------------------------+n | date index mydate bcaldate |n |---------------------------------------------|n 1. | 1971-02-05 100 05feb1971 05feb1971 |n 2. | 1971-02-08 100.84 08feb1971 08feb1971 |n 3. | 1971-02-09 100.76 09feb1971 09feb1971 |n 4. | 1971-02-10 100.69 10feb1971 10feb1971 |n 5. | 1971-02-11 101.45 11feb1971 11feb1971 |n +---------------------------------------------+n

Although mydate and bcaldate look similar, they have different encodings. Now, I can tsset on the business date bcaldate and list the first five observations with the lag of index recalculated.

. tsset bcaldaten time variable: bcaldate, 05feb1971 to 23mar2015, but with gapsn delta: 1 daynn. list bcaldate index l.index in 1/5nn +-----------------------------+n | L.|n | bcaldate index index |n |-----------------------------|n 1. | 05feb1971 100 . |n 2. | 08feb1971 100.84 100 |n 3. | 09feb1971 100.76 100.84 |n 4. | 10feb1971 100.69 100.76 |n 5. | 11feb1971 101.45 100.69 |n +-----------------------------+n

As expected, the issue of gaps due to weekends is now resolved. Because I have a calendar that excludes Saturdays and Sundays, bcaldate skipped the weekend between 05feb1971 and 08feb1971 when calculating the lagged index value and will do the same for any subsequent weekends in the data.

Excluding specific dates

So far I have not excluded gaps in the data due to other major holidays, such as Thanksgiving and Christmas. Stata has several variations on the omit command that let you exclude specific dates. For example, I use the omit command to omit the Thanksgiving holiday (the fourth Thursday of November in the U.S.) by adding the following statement in my business calendar.

omit dowinmonth +4 Th of Novn

dowinmonth stands for day of week in month and +4 Th of Nov refers to the fourth Thursday of November. This rule is applied to every year in the data.

Another major holiday is Christmas, with the NASDAQ closed on the

25th of December every year. I can omit this holiday in the calendar as

omit date 25dec*

The * in the statement above indicates that December 25 should be omitted for every year in my nasdaq calendar.

This rule is misleading since the 25th may be on a weekend, in which

case the holidays are on the preceeding Friday or following Monday. To

capture these cases, I add the following statements:

omit date 25dec* and (-1) if dow(Sa)nomit date 25dec* and (+1) if dow(Su)n

The first statement omits December 24 if Christmas is on a Saturday,

and the second statement omits December 26 if Christmas is on a Sunday.

Encodings

I mentioned earlier that the encodings of regular date mydate and business date bcaldate are different. To see the encodings of my date variables, I apply the numerical format and list the first five observations.

. format %8.0g mydate bcaldaten. list in 1/5n +-----------------------------------------+n | date index mydate bcaldate |n |-----------------------------------------|n 1. | 1971-02-05 100 4053 0 |n 2. | 1971-02-08 100.84 4056 1 |n 3. | 1971-02-09 100.76 4057 2 |n 4. | 1971-02-10 100.69 4058 3 |n 5. | 1971-02-11 101.45 4059 4 |n +-----------------------------------------+n

The variable bcaldate starts with 0 because this was the centerdate in my calendar nasdaq.stbcal. The business date encoding is consecutive without gaps, which is why

using lags or any time-series operators will yield correct values.

Summary

Using regular dates with time-series data instead of business dates may be misleading in case there are gaps in the data. In this post, I showed a convenient way to work with business dates by creating a business calendar. Once I loaded a calendar file into Stata, I created business dates using the bofd() function. I also showed some variations of the omit command used in business calendars to accommodate specific gaps due to different holidays.

aHR0cDovL3dlaXhpbi5xcS5jb20vci83VWptLXRmRUhJcGpyWk9kOXgzLQ== (二維碼自動識別)