top of page
Search
• Doug Casterton

# Contact Center Basic Forecasting Building Blocks

Updated: Feb 7, 2022

The basic building blocks of a systematic forecasting approach (Systematic: Components of the time series with consistency or recurrence and can be described and modeled) can be based on relatively simple mathematical methods.

Daily, Weekly, Monthly Volume forecasting involves thinking of a series as a combination of level, trend, seasonality, and noise components in addition to the Poisson fluctuations that you can expect. These components for a customer operational forecast have been depicted above.

Simply put, to predict the future, we need first to understand what has occurred previously and extrapolate that into the future. Thus, analyzing the historical volume and explaining what has happened using these factors above are perhaps the most critical part of any forecasting process. Analyzing the influence of these factors one by one is known as a decomposition method.

Making a graphical plot of the available data is often not that insightful due to the weekly cycle and special events that are likely to mask the identity of a slight but steady increase in contact volume. For example, we usually can quite clearly see the weekly cycle, with less traffic on Saturday and Sunday vs. a weekday - often with Monday as the highest volume day. However, it is often difficult to identify all outliers to derive a trend. For this reason, we have to identify the special events and holidays and then deseasonalize the data. This should be done for the weekly cycle but can also be completed for the yearly cycle if enough data is available.

In a perfect world following deserializing, we would get a straight line representing the long-term trend, with special days and the seasons filtered out. Of course, if the line represented a flat profile after this effort, then there is no trend. However, we rarely get such a flat profile, and there will still be unexplained variations. Some of these will be caused by normal Poisson fluctuations, but often they are more significant than we would expect just from a Poisson distribution. Once you have identified the trend, you can assemble the forecasting.

Once the trend is determined, we can now include the seasons (both intra-week and intra-year) and then the special days to produce the forecast. While deseasonalizing consisted of dividing by the average per day or week, seasonalizing consisted of multiplying the deseasonalized forecast by the seasonal factors. To correct the forecast for special days and events, we must quantify their effect. This can be achieved via various methods, but even simple averaging can be effective as a method if past events are likely to be similar to future ones.

So this concludes with a brief overview of a straightforward but effective forecasting method. It should be noted that many elements need to be factored in, and they will vary from operation to operation. For example, things that are likely to differ from operation to operation are; the amount of available data, the volatility of that data, and the extent to which special days influence.

Thus, the only way to decide what to include and what not to include is to continually test using a systematic approach of continuous improvement.