Call volume data exploratory analysis

Today I am performing exploratory data analysis of Call Volume data provided by US bank. The nice chart below shows the volume of calls for each day. It can be clearly seen from it that there is strong presence of inter-day pattern.

Call Center Volumes

This is weekly pattern with maximal volume at each Monday. The next image reveals it better. It is kind of “heat map”, where red color corresponds to higher call volumes.

Call Center weekly pattern

Using a smoothing on data reveals additional pattern at large scale – monthly pattern and season pattern (Spring, Summer, Autumn).

Monthly Call Volume pattern

One software, that can make forecasts for such data, have to have a way to reproduce such pattern on different time scales – weekly, monthly (known as “seasonality” in statistics). It have to use events such as holidays to properly adjust for days after (and before) them. There are trends that are not seasonal depending from economic factors (specific for the type of business that particular call center serves). This is very similar to price levels for some commodity products.

I read a lot of forecasting papers last few weeks. Most of them are focused on using past data for predicting future volume levels. Events can influence the call volumes and increase them in orders of magnitude and it is important to incorporate them in the forecast algorithm. My view is that there are following types of events:

Unpredictable events.
Predictable events not known to the customer. Example is promotion event.
Predictable events known to customer that can influence the call volume in advance. For example shopping before holidays.

Third case is important because it requires not only historical data but data from future (list of events) for forecasting algorithm.

Tags: call center, call center data, call center research, contact center, contact center research

Steady progress

I am doing steady progress on my algorithm development. It is statistical learning algorithm that can capture nonlinear behavior and long time dependency. My main concern at the moment is the speed of learning. It seems that most competitors use algorithms that forecast for the next period using from 1h to 1 day computational time, some even use a forecast by request via e-mail. My understanding of this business is that you need accurate and on time forecast and in some situations in “real time”.

I have done some research on what features are necessary for a call center forecasting software. Most platforms have a lot of features and are complex and complicated. My intentions are to develop simple to use (but sophisticated) program that initially will have the minimum of features to be useful. Here is my initial list:

Call Volume forecasting.
Import of CVS data.
Export of CVS forecasts.
Event tagging.
Handling 5min, 15min, 30min, 1h, 1day time intervals
Forecasts on periods from 1 week to 1 month (interday) and intraday forecasting.

Tags: call center, call center research, contact center, forecasting