Reliable forecasting is a key component in all workforce management (WFM) processes – you need to know what to plan for. However, predicting the future is no easy task. If it were, we’d see a whole lot of fewer errors being made in the world. Even with massive information about the past and with all the evidence to believe that the future will look similar to that of today, accuracy still doesn’t come easily. The difficulty applies, for example, to predicting the call, chat and e-mail volumes of a contact center.
Teleopti remains at the forefront of the state-of-the art research carried out in this field, through its collaboration with Stockholm’s Royal Institute of Technology (KTH). This is where I come in the picture as a master candidate at KTH, as well as Göran Svensson, doctoral candidate in applied mathematics at KTH and employed at Teleopti.
Göran Svensson’s research involves stochastic processes in relation to queuing systems, both in classical and mean-field form. He also researches skill-based agent scheduling and develops Erlang-inspired models for chat-based communication. Finally, he studies models and algorithms for blending in multi-media and multi-skill environments. To what end, you may ask? To contribute to Teleopti’s understanding of the different aspects of WFM and his cooperation with lead developers there, in enhancing the product portfolio accordingly.
If you, like Teleopti, have many, many customers in 80+ countries, it’s a challenge to find a forecasting method that works for everyone. What seasonal variations are there for this particular customer and in this particular country? What bank holidays are there? How do all the various factors come into play?
I address these questions in my KTH master’s thesis, which offers a step-by-step forecasting method. It’s intended to serve as the basis for the next generation of forecasting for Teleopti WFM. The method examines historical data to determine how to best predict the future in specific cases. Briefly, the method consists of the following steps:
- Remove irrelevant data that differs too much from the rest of the data.
- Use the auto-correlation function to find out the seasonal variations that are present in the data.
- Estimate and remove the trend.
- Split the data into two parts. Use the first part of the data to fit different models. Compare the results from using the different models with the second part of the data. The model giving the result that best fits the second part of the least squares method (standard approach in regression analysis) is the one to be used to forecast future volumes.
- Estimate the chosen model again, using all the data and remove it from the full sample of data.
- Forecast the trend, using the Holt’s method.
- Combine the estimated trend with the estimated seasonal variations to carry out the forecast.
This method has been tested on countless real data and works very well, even if some instances require manual adjustments in order to more accurately identify patterns and variations. Teleopti is continuing to refine this method, making the production of solid forecasting even better for our customers. As such, the company remains in the lead, seeking answers and improved solutions from its collaboration with KTH.
If you’d like to know more or discuss forecasting model, please fill out the form below and we will contact you!
What are the biggest forecasting challenges your company faces today? What do you wish your forecasting tool could help you with? Please let us know!