8
$\begingroup$

I have a call center data (such as one below) that has call data collected every 15 minutes. For a day the periodicity is 96 and for a week the periodicity is (7 x 96 = 672). If I would like to forecast this data I can use models that handle multiple seasonality such as DSHW, TBATS and BATS in R forecast package. As you can imagine having a periodicity of 96 and 672 would require tremendous amount of computational resource and not very many techniques can handle such a long periodicity.

I recently came across an excellent article by Shen that uses Singular Value Decomposition to reduce the dimension/periodicity of the time series, so we could have a parsimonious model and the model would be more efficient/require less computational resource. I'm open to other methods such as discrete wavelet transformation (dwt), discrete Fourier transformation (dft) etc.,

    Time in mins                    
Date      0:15 0:20 0:45 1:00 1:15 1:30
1/1/2012    41  12  33  42  38  27
2/1/2012    50  12  20  14  27  46
3/1/2012    30  17  21  44  47  45
4/1/2012    26  15  10  13  50  38
5/1/2012    39  17  22  28  48  12
6/1/2012    30  10  25  30  21  21
7/1/2012    47  46  18  10  36  37
8/1/2012    26  37  41  36  37  12

Below are my questions:

  1. How to implement SVD/DWT/DFT based dimensional reduction for the aforementioned time series data using tools like R and matlab ?
  2. Are there any automatic procedures available for time series dimensional reduction ?
$\endgroup$
  • $\begingroup$ What do you want your data to tell you? What are you trying to estimate? You need to tailor your approach to what you want to know. $\endgroup$ – Placidia Jan 14 '14 at 22:43
  • $\begingroup$ @Placidia as noted in my question, I would like to Forecast this type of data which has complex multiple seasonality. $\endgroup$ – forecaster Jan 14 '14 at 22:44
  • $\begingroup$ Is your data showing frequency of calls for binned call-length by 5 minute intervals per day? Because if so, then I don't understand the comment about the data being collected every 15 minute. Do you get a set like that every 15 minutes? Because if so, the data sample is really misleading as the date field should say "%d/%m/%y %h:%m" where the last two entries should be hours and minutes. If that is the case, I still don't understand your question as the period of the seasonality does not render a time-series harder to estimate. Am I missing something? $\endgroup$ – means-to-meaning Jan 14 '14 at 23:18
  • $\begingroup$ @means-to-meaning I changed the labels in the data. thanks for pointing it out. $\endgroup$ – forecaster Jan 14 '14 at 23:21
  • 1
    $\begingroup$ Any luck with this problem? Can you provide a general overview on how you tackled it eventually? $\endgroup$ – Denys S. Sep 16 '16 at 22:37
5
$\begingroup$

You might want to consider forecastable component analysis (ForeCA), which is a dimension reduction technique for time series, specifically designed to obtaina lower dimensional space that is easier to forecast than the original time series.

Let's look at an example of monthly sunspot numbers and for computational efficiency let's just look at the 20th century.

yy <- window(sunspot.month, 1901, 2000)
plot(yy)

sunspots monthly

Sunspot numbers are only a univariate time series $y_t = (y_1, \ldots, y_T)$, but we can turn this into a multivariate time series by embedding it its lagged $(p+1)$-dimensional feature space $X_t = \left(y_t, y_{t-1}, \ldots, y_{t-p}\right)$. This is a common technique in non-linear time series analysis.

XX <- embed(yy, 24)
XX <- ts(XX, end = end(yy), freq = 12)
dim(XX)

## [1] 1166   24

In R you can use the ForeCA package to do the estimation. Note that this requires the multivariate spectrum of a $K$-dimensional time series with $T$ observations, which is stored in a $T \times K \times K$ array (one can use symmetry/Hermitian property to half the size). Hence it takes considerably longer to compute than iid dimension reduction techniques (such as PCA or ICA).

So here we take the 24-dimensional time series of embedded sunspot numbers and try to find a 6-dimensional subspace that has interesting patterns that can be easily forecasted.

library(ForeCA)
# this can take several seconds
mod.foreca <- foreca(XX, n.comp = 4, 
                     spectrum.control = list(method = "wosa"))

mod.foreca

## ForeCA found the top 4 ForeCs of 'XX' (24 time series).
## Out of the top 4 ForeCs, 0 are white noise.
## 
## Omega(ForeC 1) = 53% vs. maximum Omega(XX) = 43%.
## This is an absolute increase of 9.9 percentage points (relative: 23%) in forecastability.
## 
## * * * * * * * * * * 
## Use plot(), biplot(), and summary() for more details.


plot(mod.foreca)

foreca results from sunspots

The biplot shows that the first component all points in the same direction, which is telling us that this component will be the overall/average pattern. The barplots on the right show how the forecastable components (ForeCs) have indeed decreasing forecastability, and the first component is more forecastable than the original series. In this example, all original series have the same forecastability, as we used the embedding. For general multivariate time series, this is not the case.

Now what do these series look like?

mod.foreca$scores <- ts(mod.foreca$scores, start = start(XX), 
                        freq = frequency(XX))
plot(mod.foreca$scores)

ForeCs

Indeed, the first component is more forecastable than the original series, since it is less noisy. The remaining series also show very interesting patterns, that are not visible in the original series. Note that all ForeCs are orthogonal to each other, i.e., they are uncorrelated.

round(cor(mod.foreca$scores), 3)

##        ForeC1 ForeC2 ForeC3 ForeC4
## ForeC1      1      0      0      0
## ForeC2      0      1      0      0
## ForeC3      0      0      1      0
## ForeC4      0      0      0      1

The spectrum of each series also gives a good idea of the different ForeCs [sic!] in sunspot activity.

spec <- mvspectrum(mod.foreca$scores, "wosa")
plot(spec)

spectrum of ForeCs

$\endgroup$
1
$\begingroup$

We see data like this from a number of fast food franchises that want to predict demand in 15 minute intervals. You can construct a daily model that incorporates day-of-the-week effects, weekly effects, monthly effects , week of the-month effects , day-of-the month effects, weekend effects , holiday lead and lag effects , changes in level effects , changes in trends effects, change in day-of-the-week effects , autoregressive effects and pulse effects along with any user-specified predictor series. The daily forecasts can then be used to help forming 15 minute interval forecasts finalizing in a required reconciliation between daily forecasts and 15 minute forecasts. AUTOBOX (http://www.autobox.com) is a tool that can be helpful in this regard while other major software vendors come up short as do all the "free solutions" to this very interesting problem. In terms of full disclosure , I had a hand in developing some of these solutions for Automatic Forecasting Systems, the makers of AUTOBOX. You might try your hand at doing this in R , if you have enough time and patience to program the required steps.

$\endgroup$
  • $\begingroup$ thanks for your response. My question was about dimensionality reduction and forecasting directly to handle data that was collected every 15 minutes as opposed to a top down approach like the one you are suggestion. I'm sure you are aware that there is a tremendous ongoing research to use advanced techniques to do forecasting by minute or some instance by second for which dimensional reduction was recommended. $\endgroup$ – forecaster Jan 22 '14 at 3:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.