Rossman_competiton

As a MATLAB user who transitioned to R for data science, one of the things I miss the most is fantastic interactive plotting. MATLAB lets you zoom and pan with ease, and multiple graphs can be stacked vertically with linked axes, so you can navigate multiple time series simultaneously. Recently I was pleasantly surprised to discover R’s dygraphs package, which lets you do almost everything you can do in MATLAB.

Let’s have a look at Rossmann sales data using dygraph:

## [1] "character"

## 
##  FALSE   TRUE 
## 858354 199943

##      store          dayOfWeek          date                sales      
##  Min.   :   1.0   Min.   :1.000   Min.   :2013-01-01   Min.   :    0  
##  1st Qu.: 280.0   1st Qu.:2.000   1st Qu.:2013-08-26   1st Qu.: 3727  
##  Median : 558.0   Median :4.000   Median :2014-04-20   Median : 5744  
##  Mean   : 558.3   Mean   :3.998   Mean   :2014-04-30   Mean   : 5774  
##  3rd Qu.: 837.0   3rd Qu.:6.000   3rd Qu.:2015-01-12   3rd Qu.: 7856  
##  Max.   :1115.0   Max.   :7.000   Max.   :2015-09-17   Max.   :41551  
##                                                        NA's   :41088  
##    Customers         open           promo         stateHoliday      
##  Min.   :   0.0   Mode :logical   Mode :logical   Length:1058297    
##  1st Qu.: 405.0   FALSE:178801    FALSE:653953    Class :character  
##  Median : 609.0   TRUE :879485    TRUE :404344    Mode  :character  
##  Mean   : 633.1   NA's :11        NA's :0                           
##  3rd Qu.: 837.0                                                     
##  Max.   :7388.0                                                     
##  NA's   :41088                                                      
##  schoolHoliday      split                 Id         
##  Mode :logical   Length:1058297     Min.   :    1    
##  FALSE:858354    Class :character   1st Qu.:10273    
##  TRUE :199943    Mode  :character   Median :20545    
##  NA's :0                            Mean   :20545    
##                                     3rd Qu.:30816    
##                                     Max.   :41088    
##                                     NA's   :1017209

This graph is initialized to show the first 6 months of 2014, but it contains the entire sales history of store 1081 along with the predictions of a simple model (median by store, dayOfWeek, and promo period). Promo periods are slightly shaded. State holidays are marked with dashed lines, color-coded to the value of StateHoliday.

Here’s how you interact with the graph:

• Pan: Drag the range selector at the bottom
• Zoom horizontally: click on the graph and drag left or right
• Zoom vertically: click on the graph and drag up or down
• Zoom out: double-click on graph window
• Moving average: enter a number of days in the lower left box

Try setting the moving average to 14 days: this removes the alternating promo/non-promo periods and shows slower trends, as well as the holiday behavior.