To create a ts-object in R, one has to specify a data frame, a start date and the frequency of the time series.
When searching the internet (e.g. Role of frequency parameter in ts), I get the impression that by choosing the frequency, one can emphasise whatever periodic pattern one believes is the most important in the data. However, I doubt that this is actually true. My impression is that it is solely used to compute the dates of the time series on-the-fly. E.g. when I set the start date “2015-08-01”, R automatically transforms it into a decimal date and I get something like 2015.58. If I now choose a frequency of 365 (or 365.25), R divides one unit by 365 and assigns this fraction to each day as one unit ahead, so the entry 366 days later is exactly 2016.58. However, if I choose frequency=7, the fraction assigned to each day is 1/7th, so the date assigned to the 8th day after my start date corresponds to a decimal number between 2016 and 2017. So the only choice for a data set with 365 entries per year is 365, isn’t it? And it is only used to actually create the time series?
Otherwise, if I choose the xts-class, an xts-object is built from a vector and a matrix where the vector has to be created in advance. So here there is no need to compute the date on-the-fly using a start date and a frequency and that is the reason why no frequency has to be assigned at all.
In both cases I can apply forecasting packages to either ts or xts objects (such as ARIMA, ets, stl, bats, bats etc) without specifying anything else so this shows that the frequency is actually not used for anything else. Or am I missing something here?
Thanks in advance for your comments!
I am trying to cluster customers consumption behaviors using time series techniques. Customers buy tokens and use them whenever they want (a max of 4 tokens per day).
This is a sample of what the customers journeys time series (x = days after first order , y = number of tokens consumed per day) and it look alike the image below.
I tried clustering with derived variables (median delay between two events, standard deviation of the delays, total number of tokens, time between first and last consumption, mean number of tokens consumed per consumption event ...). I used K-means and this gave me some good results but it wasn't enough to spot all patterns in data. I looked at some papers about the use of Dynamic time warping in such cases but I have never used such algorithms..
Is there any materials (demos) on the use of such algorithms to cluster such time series ?
Yes.
There are many techniques that can be useful here.
The obvious approach from literature would be HAC with DTW.
This is more of a methodological (rather than a programming) issue, yet it feels SO is the right place for it. Following the ups and downs after Yahoo changed its defaults in May 2017 for fetching daily data (discussed on https://github.com/joshuaulrich/quantmod/issues/174, http://blog.fosstrading.com/2017/06/yahoo-finance-alternatives.html and also on SO Why Open,High,Low prices are wrong when using quantmod?) I am probably not the only one not 100% certain which data to use in a backtesting procedure and whether quantmod getSymbols.yahoo and adjustOHLC still provide the relevant data for quality backtesting.
Quantmod 0.4.11 also includes AlphaVantage as (adjusted stock) data provider, but I am not familiar with their reliability.
How to prepare the (stock and index) data obtained from getSymbols calls? Which data ((stock & dividends) adjusted or unadjusted) should be used? Which transformations do you use? The adjustOHLC function also contains a bug, as it is not split adjusted (easily seen on AAPL by calling
getSymbols(AAPL)
chart_Series(adjustOHLC(AAPL))
and observing a jump in 2014.
You should always use adjusted prices. Most of the time when data provider doesn't have adjusted prices then usually provider's close prices are adjusted. There is no point doing backtests on a raw close prices data. I've once made a mistake by downloading close prices instead of adjusted and at the end of backtesting, my strategy told me that among all S&P composites Master Card was the worst performer. After looking at the MA chart it was obvious why.
Beacuse of a split on January 22, 2014 my data had a single return over -90%! In conclusion raw close data for backtesting might give you utterly false results.
How to deal with splits
Divide every price before a split by split ratio. For example Master Card had 1:10 split ratio so you should divide every price before 21.01.2014 by 10. It's very easy to find splits in a data, you just have to look for returns around or below -50%.
Dividends
Subtract from every price before dividend day dividend amount. To find dividends days you need dividends calendar, it's impossible to find them by yourself.
All,
I'm looking to download stock data either from Yahoo or Google on 15 - 60 minute intervals for as much history as I can get. I've come up with a crude solution as follows:
library(RCurl)
tmp <- getURL('https://www.google.com/finance/getprices?i=900&p=1000d&f=d,o,h,l,c,v&df=cpct&q=AAPL')
tmp <- strsplit(tmp,'\n')
tmp <- tmp[[1]]
tmp <- tmp[-c(1:8)]
tmp <- strsplit(tmp,',')
tmp <- do.call('rbind',tmp)
tmp <- apply(tmp,2,as.numeric)
tmp <- tmp[-apply(tmp,1,function(x) any(is.na(x))),]
Given the amount of data I'm looking to import, I worry that this could be computationally expensive. I also don't for the life of me, understand how the time stamps are coded in Yahoo and Google.
So my question is twofold--what's a simple, elegant way to quickly ingest data for a series of stocks into R, and how do I interpret the time stamping on the Google/Yahoo files that I would be using?
I will try to answer timestamp question first. Please note this is my interpretation and I could be wrong.
Using the link in your example https://www.google.com/finance/getprices?i=900&p=1000d&f=d,o,h,l,c,v&df=cpct&q=AAPL I get following data :
EXCHANGE%3DNASDAQ
MARKET_OPEN_MINUTE=570
MARKET_CLOSE_MINUTE=960
INTERVAL=900
COLUMNS=DATE,CLOSE,HIGH,LOW,OPEN,VOLUME
DATA=
TIMEZONE_OFFSET=-300
a1357828200,528.5999,528.62,528.14,528.55,129259
1,522.63,528.72,522,528.6499,2054578
2,523.11,523.69,520.75,522.77,1422586
3,520.48,523.11,519.6501,523.09,1130409
4,518.28,520.579,517.86,520.34,1215466
5,518.8501,519.48,517.33,517.94,832100
6,518.685,520.22,518.63,518.85,565411
7,516.55,519.2,516.55,518.64,617281
...
...
Note the first value of first column a1357828200, my intuition was that this has something to do with POSIXct. Hence a quick check :
> as.POSIXct(1357828200, origin = '1970-01-01', tz='EST')
[1] "2013-01-10 14:30:00 EST"
So my intuition seems to be correct. But the time seems to be off. Now we have one more info in the data. TIMEZONE_OFFSET=-300. So if we offset our timestamps by this amount we should get :
as.POSIXct(1357828200-300*60, origin = '1970-01-01', tz='EST')
[1] "2013-01-10 09:30:00 EST"
Note that I didn't know which day data you had requested. But quick check on google finance reveals, those were indeed price levels on 10th Jan 2013.
Remaining values from first column seem to be some sort of offset from first row value.
So downloading and standardizing the data ended up being more much of a bear than I figured it would--about 150 lines of code. The problem is that while Google provides the past 50 training days of data for all exchange-traded stocks, the time stamps within the days are not standardized: an index of '1,' for example could either refer to the first of second time increment on the first trading day in the data set. Even worse, stocks that only trade at low volumes only have entries where a transaction is recorded. For a high-volume stock like APPL that's no problem, but for low-volume small caps it means that your series will be missing much if not the majority of the data. This was problematic because I need all the stock series to lie neatly on to of each other for the analysis I'm doing.
Fortunately, there is still a general structure to the data. Using this link:
https://www.google.com/finance/getprices?i=1800&p=1000d&f=d,o,h,l,c,v&df=cpct&q=AAPL
and changing the stock ticker at the end will give you the past 50 days of trading days on 1/2-hourly increment. POSIX time stamps, very helpfully decoded by #geektrader, appear in the timestamp column at 3-week intervals. Though the timestamp indexes don't invariably correspond in a convenient 1:1 manner (I almost suspect this was intentional on Google's part) there is a pattern. For example, for the half-hourly series that I looked at the first trading day of ever three-week increment uniformly has timestamp indexes running in the 1:15 neighborhood. This could be 1:13, 1:14, 2:15--it all depends on the stock. I'm not sure what the 14th and 15th entries are: I suspect they are either daily summaries or after-hours trading info. The point is that there's no consistent pattern you can bank on.The first stamp in a training day, sadly, does not always contain the opening data. Same thing for the last entry and the closing data. I found that the only way to know what actually represents the trading data is to compare the numbers to the series on Google maps. After days of futiley trying to figure out how to pry a 1:1 mapping patter from the data, I settled on a "ballpark" strategy. I scraped APPL's data (a very high-volume traded stock) and set its timestamp indexes within each trading day as the reference values for the entire market. All days had a minimum of 13 increments, corresponding to the 6.5 hour trading day, but some had 14 or 15. Where this was the case I just truncated by taking the first 13 indexes. From there I used a while loop to essentially progress through the downloaded data of each stock ticker and compare its time stamp indexes within a given training day to the APPL timestamps. I kept the overlap, gap-filled the missing data, and cut out the non-overlapping portions.
Sounds like a simple fix, but for low-volume stocks with sparse transaction data there were literally dozens of special cases that I had to bake in and lots of data to interpolate. I got some pretty bizarre results for some of these that I know are incorrect. For high-volume, mid- and large-cap stocks, however, the solution worked brilliantly: for the most part the series either synced up very neatly with the APPL data and matched their Google Finance profiles perfectly.
There's no way around the fact that this method introduces some error, and I still need to fine-tune the method for spare small-caps. That said, shifting a series by a half hour or gap-filling a single time increment introduces a very minor amount of error relative to the overall movement of the market and the stock. I am confident that this data set I have is "good enough" to allow me to get relevant answers to some questions that I have. Getting this stuff commercially costs literally thousands of dollars.
Thoughts or suggestions?
Why not loading the data from Quandl? E.g.
library(Quandl)
Quandl('YAHOO/AAPL')
Update: sorry, I have just realized that only daily data is fetched with Quandl - but I leave my answer here as Quandl is really easy to query in similar cases
For the timezone offset, try:
as.POSIXct(1357828200, origin = '1970-01-01', tz=Sys.timezone(location = TRUE))
(The tz will automatically adjust according to your location)
I would like to use the twitteR package in R to compare the number (count) of mentions of two competing hashtags from 11/14/2012-11/22/2012 (i.e. an 8-day time period). For example, I would like hourly comparisons of two hashtags: #A vs #B.
I am wondering if there is a way to use the twitteR package in R to do this. Something using the searchTwitter function:
searchTwitter(searchString, n=25, lang=NULL, since=NULL, until=NULL,
locale=NULL, geocode=NULL, sinceID=NULL, ...)
I am not interested in grabbing all tweets, just getting an hourly count comparison for #A vs. #B over the specified time period. I know I have to be cognizant of the rate limit and maybe will have to do some clever sampling of tweets to avoid the rate limit. Any ideas if this is feasible, and how I would go about coding it?
I would pull 100 tweets for each hash tag every 2 minutes. Use #TweetsReturned / (TimePulled - TimeOfOldestTweet) to get a Tweets per Unit Time estimate. You can plot these to get a moving average type chart of activity over time. If you do tweets per 2 minutes, just add them up to estimate tweets per hour.