A Forecasting Case Study
All call centers are unique. This is certainly true when considering historical data. When it comes to using this data to produce forecasts, the ‘one size fits all’ approach is unlikely to work in all circumstances.
The Vantage Point forecaster is a versatile tool which can be tailored, using the appropriate directives, to produce the most accurate forecasts for any given call center. However, if the forecasting tool is to make the best predictions for an individual center, it is recommended that an analysis of the historical data is undertaken. Such an analysis was done for a Pipkins’ customer recently and the following case study is a summary of the key steps in the process with this first part focusing on the data analysis aspect.
2. Analysing the data
The first step is always to examine the historical data to look for certain factors, e.g. seasonality, growth, step changes, etc. The amount of available data is important when looking for patterns – ideally at least 2 year’s worth of data is necessary for such an analysis. In this case the customer had 3 years of data and all of this was used at some stage. The initial analysis looked at all queues. This can often give a good view of the overall nature of the call center although this view will be drilled down as the analysis progresses. Looking at the overall data from January 2009 on a monthly basis, this is the result:
There appears to be some underlying growth which seems quite steady. There are peaks in January of each year and also peaks in other months which are not as pronounced.
Looking at the weekly picture over the same period:
There are some significant weekly variations with some volumes changing by 20% or more week on week. Some of these will be due to obvious factors (Christmas and other public holidays) but it would be worth investigating further to try and discover other events which result in large changes in call volume.
To this end, we can look at an individual day, in this case Wednesday:
This shows a similar pattern to the weekly graph, which indicates that the fluctuations may not be limited to particular days, but more likely affect the whole week – this can be confirmed by comparing different days of the week – here is an example showing Tuesday and Wednesday. In the main they follow the same pattern
Looking at the monthly, weekly and daily patterns show some obvious repeated patterns as mentioned above (e.g. Christmas) but it is not easy to determine recurring (seasonal) patterns from these graphs. To help determine seasonal trends we can also look for a year on year correlation by comparing them:
It appears from this that there is quite a strong seasonal component to call volumes. Some events do recur at the same time every year, some may be offset by a week (this can depend on where the year starts), some show little or no pattern year on year.
Further analysis can be done by looking at the higher volume queues to see if the seasonal peaks are recorded for all queues or limited to certain ones:
This shows that at least some of the peaks are much more pronounced for certain queues although all show some seasonality around the same times.
The above would indicate that, when forecasting, the appropriate directives should be used to enable this to be taken into account and any special events identified which can be predicted.
This customer has several low volume queues (LVQs) – in this case these were queues which accounted for less than 100 calls per day. These were grouped by removing the higher volume queues identified earlier from the list of all queues.
Looking at the low volume queues (as a group) shows a different picture from that already described above:
The seasonality indentified in the higher volume queues is not as recognisable here but there appears to have been a significant change around September 2011. Further analysis of this shows an increase of calls between the end of August and late October. Volumes returned to normal until two new queues were opened in mid November which resulted in the overall increase.
This is illustrated below by showing the two new queues separately from the other LVQs:
Further analysis of the busy period (September – October) showed that the increase was limited to 4 out of the 25 LVQs with the final peak in the last week of October:
As this period is much more pronounced than in previous years the customer would need to determine if this was likely to recur and account for it in the forecast.
It is worth pointing out that although it is normally recommended that low volume queues be aggregated with higher ones, some events which may only affect some smaller queues for short periods may be ‘lost’ if the queues are aggregated. It is therefore worth doing the analysis initially to identify such cases.
It is important to treat the new queues separately when forecasting as the amount of historical data available is limited.
3.1 For the established higher volume queues there is a strong seasonal element and some underlying growth.
3.2 There are recently added queues which need to be forecasted separately.
3.3 There are some LVQs which may need to be forecasted separately if there is a evidence of periods of marked change in volumes.
Next we will consider how to use this information to tailor our forecasts.