0
$\begingroup$

I am trying to predict sales values using time series approach. Below graph is the sales for a store over a period of 942 days (sales will be 0 when the store is closed and are not plotted in first 2 graphs for sake for clarity in graph):

Complete time series

Stacking yearly sales give us a graph like below:

Sales per year stacked over each other

It looked like it may be worth investing in time series, so I tried drawing the acf and pacf graphs which are shown below:

ACF plot

PACF

The strong auto correlation in the curve suggest the presence of time series. So I applied, auto.arima() function from r to just look at how it performs (salesModel <- auto.arima(saleDataSeries)). Below is the residuals plot from arima. I tested the prediction on Kaggle Leaderboard and the approach did not perform that well(competition is over now).

Residual

Model Parameters:

Series: saleDataSeries 
ARIMA(5,1,4)                    

Coefficients:
          ar1      ar2      ar3      ar4      ar5      ma1     ma2     ma3
      -0.0885  -0.1938  -0.5965  -0.3084  -0.3058  -0.8985  0.0040  0.7063
s.e.   0.1036   0.0634   0.0451   0.0329   0.0690   0.1202  0.0704  0.0909
          ma4
      -0.6996
s.e.   0.0296

sigma^2 estimated as 4621003:  log likelihood=-8555.57
AIC=17115.2   AICc=17115.43   BIC=17163.67

Currently I am also looking at multiple seasonality using tbats, but I believe that there can be certain modifications that can be made to simple ARIMA model as the ACF/PACF graph still have some spikes in the residual component. I am unsure why this is happening and any insights would be helpful

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

ARIMA modelling (by itself) is of little use with your kind of data . See my response to Time series analysis to quantify trend when seasonal amplitude is decreasing as a guide to what can be done. Note that when I answered the post I was unaware that it was an economic series. I specifically disabled the identification of the optimum lead and lag effects around each holiday . I also disabled the identification of particular days in the month and particular weeks in the month that might be of interest. I also disabled long-weekend effects, Monday after a Friday holiday effect, Friday before a Monday Holiday effect, month-end effects etc. as I had no idea that it was economic data. If I had enabled these economically oriented remedies , the results would have been even more amazing. If you wish you can post your data and I will demonstrate that for you.

By the way my "eye" tells me that you have a level shift not a trend but only good exploratory analysis can confirm that.

Improve this answer
$\endgroup$
5
  • $\begingroup$ The data is publicly available at kaggle.com/c/rossmann-store-sales/data (one would need to sign up though). The train.csv has all the data and I had grouped the data by store and analysing for each store. The data I have posted is for store 6. Let me know if you can't access it. $\endgroup$
    – Aman Deep Gautam
    Dec 15, 2015 at 19:48
  • $\begingroup$ Also could you explain, intuitively what assumptions in ARIMA fails one this data? $\endgroup$
    – Aman Deep Gautam
    Dec 15, 2015 at 20:11
  • $\begingroup$ Essentially ARIMA is all about memory i.e. leaning on prior values explicitely. In terms of retail daily data one needs to incorporate fixed effects like day-of-the-week , month-of-the-year etc .... Adding level shifts and determinstic trends (like the one you assumed) should be done when needed not by assumption. Combining these effects with ARIMA can often ( read:nearly always) lead to a powerful/efficient model with or without user-specified predictors like weather/price/promotion/#of sales clerks etc. This class of model is called a Transfer Function or a Dynamic Regression. $\endgroup$
    – IrishStat
    Dec 15, 2015 at 20:41
  • $\begingroup$ LIke all models assumptions regarding stability of coefficients and constancy of error variance along with an error series free of information are important to verify. $\endgroup$
    – IrishStat
    Dec 15, 2015 at 20:43
  • $\begingroup$ If you wish to form the data as an excel file and email it to me I will be glad to look at it. $\endgroup$
    – IrishStat
    Dec 15, 2015 at 21:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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