Forecasting

What is forecasting?

Contact center forecasting uses mathematics to calculate future workload for all relevant channels, e.g. calls, chat, email and social media, based on history. WFM is about having ‘the right number of people in the right places at the right times, doing the right things’ and forecasting is the first step in the process of determining ‘the right number of people’.

Sometimes, it simply isn’t possible to produce a forecast. Business process outsourcers (BPOs) are commonly provided by their clients with the forecast or the required staffing. In-house contact centers launching a new line of business will have no history with which to predict the future. Even in these cases, as soon as contact history begins to accumulate, the power of forecasting can be put to work.

Why does forecasting matter?

Forecasting is the foundation stone of workforce management. Without forecasting, it isn’t possible to balance the supply of staff with demand from customers, because the demand is unknown. If demand is unknown, it’s impossible to schedule the workforce efficiently.

How does forecasting work?

Forecasting can be broken down into 4 steps:

1. Collect and analyze historical data

The main input to the forecasting process is the number of contacts such as calls that arrived (or were “offered”) per interval (e.g. every 15, 30 or 60 minutes). Ideally, average handling time (AHT) will be subject to the same analysis. That’s because the number of staff needed to handle the workload is a function of both volume and AHT, and AHT typically varies over time. In many contact centers, handling times are longer in the evening than during the day, for example.

Long-term and short-term forecasting are both important. The more history that is available, the better you can detect seasonal patterns and special events in the data. It is important to collect data in short intervals like 15 minutes. That’s because the goal is to match supply and demand across each working day, observing the peaks and troughs and avoid under- and over- staffing. That isn’t possible If you only consider the total volume per day.

Typically, the data comes from your contact routing system, e.g. the automatic call distributor (ACD) that routes incoming calls to agents. You are dealing with huge volumes of data, so an automated integration with the ACD will save a lot of time and reduce the scope for errors.

Good forecasting practice includes analyzing the data to find anomalies such as gaps in history and one-off spikes in volume. These don’t belong in the forecast and must be removed. You should however keep a note of exceptions that you know will recur in future, such as billing runs, advertising campaigns, and public holidays. We’ll come back to that in step 3, below.

2. Predict future volume and AHT

Once you have a clean set of data, you are ready to generate your forecast down to interval level.

Multiple forecasting methods (or algorithms) exist, including:

• Moving weighted averages
• Triple exponential smoothing
• Auto-regressive integrated moving average
• Neural networks
• Multiple temporal aggregation

For the simpler algorithms, it’s possible to do forecasting with a spreadsheet. The more powerful algorithms require a professional WFM application. injixo, for example, deploys multiple algorithms and uses artificial intelligence to constantly select and configure the algorithm that gives the best results with your data.

3. Apply business intelligence

Human intelligence is at least as important as artificial intelligence when it comes to forecast accuracy. No contact center experiences the same volume and pattern of calls 365 days a year, and frequently the planner will be aware of upcoming events that didn’t occur in the past. You have to deal with:

• Marketing campaigns and promotions
• Operational changes (e.g. billing, logistics, sales)
• Organizational crises (e.g. bad PR, competitive pressure)
• Corporate strategy and tactics (e.g. market development and expansion, new product launches, price changes, changes in customer base)
• Public holidays, some of which don’t happen on fixed dates
• Natural events (e.g. weather)

The impact of these exceptions must be factored into the forecast. This will be a manual process if you are using a spreadsheet, but if you are using a WFM application there should be a forecast calendar feature. Some of these drivers will have been revealed during the analysis in step 1. The rest need constant vigilance and good collaboration with colleagues in departments such as marketing. As with any process with a human element, there are several pitfalls to avoid.

4. Calculate required staffing

The final step is to convert the forecast of volume and AHT into the required headcount at interval level. Staffing requirement is the key input to the scheduling process. You need to determine the number of staff needed in each interval to handle a forecast workload to a given grade of service. For example, taking next month’s forecast of volume and average handling time for sales calls in 15-minute intervals, and calculating the number of agents required in each 15-minute interval to answer 80% of calls within 20 seconds.

Modern contact centers handle customer interaction via multiple channels: phone, web chat, email, etc. There isn’t a one-size-fits-all method of staffing calculation. The best-known method is Erlang, named after the Danish mathematician who invented it. Erlang is proven for inbound calls, but it’s no use for web-chat, because it doesn’t consider the fact that agents can typically handle more than one chat at a time. It’s no good for emails either, because emails don’t hang up and the goal is typically to handle them in a timescale measured in hours not seconds.

There is a different calculation method for each contact channel:

• Inbound calls: Erlang. There is a family of Erlang methods, but Erlang C is most commonly used in contact centers. Erlang C takes into account the expected volume of calls, the expected average handling time, and the desired service level (SL). Some WFM applications enable grade of service to be expressed as average speed of answer (ASA) or abandonment rate (ABR) as well as SL.
• Web chats: Staffing is calculated using a derivative of the Erlang C method but adapted to take into account concurrency and context-switching overhead.
• Email, back office, social media: A linear calculation is used, based on completing the number of contacts within a given time frame, taking into account the AHT.
• Outbound calls: Such calls are made at the discretion of the contact center, not the customer, and multiple approaches are possible. The simplest method is a constant requirement, based on the number of agents the planner wishes to schedule. A linear calculation can also be used based on the number of contacts in the campaign and the desired timescale. Some WFM applications provide a proprietary calculation that takes into account parameters such as right- and wrong-party connect rates.
• Non-demand: Sometimes, the planner needs to create a constant staffing requirement across opening hours, without an underlying forecast. This can happen when launching a completely new service, or for activities with such a low or volatile number of contacts that it is impractical to forecast. Still, the need for a certain number of agents is known. This is sometimes the case with social media, during the early stages of handling that channel.

Agents with more than one skill will naturally spend less time waiting for a contact they are qualified to handle. The occupancy and utilization of multi-skilled agents will therefore tend to be higher than that of single-skilled agents. This in turn means that the greater the extent of multi-skilling, the fewer agents are required to handle the workload. This effect is known as ‘pooling efficiency’. Modeling this pooling efficiency in the forecasting, staffing and scheduling process is complex and requires algorithms that are only available in powerful WFM applications such as injixo.

Staffing calculations need to take shrinkage into account. That is the percentage of paid time that agents are not available to handle contacts. This includes unproductive at-work activities such as breaks, meetings, training sessions, and 1:1s, plus out-of-office time for vacations, sickness, lateness, and other unexplained absences. 30% shrinkage means that each agent contributes the effort of 70% of one full-time equivalent. That means that the staffing requirement must be inflated by (1 / 0.7 = +43%) to counteract the effect of shrinkage.

What impact does forecasting have?

Forecasting opens the door to the rest of the WFM process. It’s only when the forecast is in place that you can calculate the number of heads needed in each time interval and schedule your agents accordingly. Without forecasting, scheduling is reduced to simple rostering, without regard to the underlying demand.

No forecast is ever completely accurate and that’s why the workforce management process includes real-time management. Without a good forecast, ideally, one that is updated continuously as new data becomes available, the job of real-time management is considerably harder. Instead of taking infrequent, considered corrective actions, planners are constantly fire-fighting.