In this article we look at how to measure the accuracy of forecasts.

One of the most important and challenging jobs in the contact centre is forecasting demand. Forecasting models require constant refinement, so it’s up the planners to evaluate how accurate their predictions have been.

At face value, forecast accuracy might seem simple to calculate – were there as many contacts as predicted? On closer inspection, however, it’s clear that businesses also need to understand exactly when contacts were made, and whether the contact centre was prepared for them.

### What Is ‘Actual Volume’?

First, we need to understand what we mean when discussing contact volume. It’s another question that seems simple at first but that requires some thought.

If we take actual volume to mean ‘the number of calls answered’ we face a very big potential problem – we won’t count the calls that we didn’t answer.

When incoming volume is greater than call handling capacity, some calls will not be answered. If they’re not answered, and therefore not counted, we end up with an unrealistic idea of total volume.

For example, imagine a contact centre that forecasts 2,000 contacts on a day when 3,000 customers are going to call. If the centre runs at maximum occupancy all day and answers 2,400 calls, the forecast will only appear to be off by 400.

For this reason, most planners evaluate forecast accuracy based on calls offered rather than calls answered. Bear in mind that this approach is also imperfect, because it counts each repeat unsuccessful contact attempt separately, inflating total volume.

It is the ‘less bad’ approach, though, so we will assume that ‘calls offered’ is the measure of actual volume.

### Percent Difference or Percentage Error

One simple approach that many forecasters use to measure accuracy is Percent Difference or Percentage Error. This is simply the difference between the actual volume and the forecast volume expressed as a percentage.

\(Percentage Error = \Large{ \frac {(Actual – Forecast)}{ Actual}} \times 100 \)We can use this formula for the first interval in the example below, where actual volume is 105 and the forecast was 102.

\(Percentage Error = \Large{ \frac {(105 – 102)}{102}} \times 100 = 2.9\% \)

Interval | Calls Offered | Forecast | Error | % Difference |
---|---|---|---|---|

8:00 am – 8:30 am | 105 | 102 | 3 | 2.90% |

8:30 am – 9:00 am | 128 | 135 | -7 | -5.30% |

9:00 am – 9:30 am | 136 | 138 | -2 | -1.50% |

This is a useful and reliable way of measuring forecast error within an interval. However, forecasters rarely need to report on single intervals, focusing instead on accuracy over long periods.

One possible solution is to calculate the mean average of the percent difference for all intervals in the data set. Errors which are negative numbers (the result of over-forecasting) will be cancelled out by errors which are positive numbers (the result of under-forecasting).

To control for this, planners can use the ‘absolute error’. Absolute error is how far from actual volume the forecast was in either direction, regardless of whether it was positive or negative.

We can see that in the data set below, the average percent difference of the absolute errors is 7.09%. If this calculation had been based on actual errors, the percent difference would appear to be 0.06. That might be what contact centres want to see – but it does not represent the facts.

Interval | Calls Offered | Forecast | Absolute Error | % Difference |
---|---|---|---|---|

8:00 am – 8:30 am | 105 | 102 | 3 | 2.9 |

8:30 am – 9:00 am | 128 | 135 | 7 | 5.3 |

9:00 am – 9:30 am | 136 | 138 | 2 | 1.5 |

9:30 am – 10:00 am | 167 | 145 | 22 | 14.1 |

10:00 am – 10:30 am | 197 | 172 | 25 | 13.6 |

10:30 am – 11:00 am | 213 | 245 | 32 | 13.9 |

11:00 am – 11:30 am | 220 | 222 | 2 | 0.9 |

11:30 am – 12:00 am | 194 | 203 | 9 | 4.5 |

Total | 1360 | 1362 | 2 | 0.1 |

Average % Difference | 7.09 |

### Measuring at Intervals

An interval is a fixed unit of time, sometimes referred to as a reporting period. Forecasters can generate very different results based on the intervals that they measure.

Across the industry, intervals of fifteen minutes are generally seen as the most desirable because they represent the most granular data it is practical to measure. Intervals of thirty minutes or an hour are also common, especially in smaller contact centres that have more volatile contact patterns.

In the above data set, we saw that the period 8:00am to 12:00pm, measured at thirty-minute intervals, had an average percent difference of 7.09%.

The same data measured at hourly intervals has an entirely different result:

9:00 am – 10:00 am303283206.810:00 am – 11:00 am41047171.7

Interval | Calls Offered | Forecast | Absolute Error | % Difference |
---|---|---|---|---|

8:00 am – 9:00 am | 233 | 237 | 4 | 1.7 |

11:00 am – 12:00 am | 414 | 425 | 11 | 2.6 |

Total | 1360 | 1362 | 2 | 1.5 |

Average Difference | 3.2 |

The data is ‘smoothed’ because inaccurate intervals are diluted either by accurate intervals or by intervals that are inaccurate in the opposite direction.

Based on this, it seems obvious that the smallest possible interval is desirable, but this is not necessarily the case. We’ve already mentioned small contact centres, which tend to have difficulty with short intervals owing to the natural volatility in their occupancy rates.

There may also be organic variability in volume during the hour. Some research has outlined a calling pattern where 40% of traffic for the hour occurs within the first fifteen minutes. The suggestion is that customers are prone to scheduling calls for the start of an hour.

This will occur uniformly in every contact centre, but it is worth investigating when unexplained variances persist in forecasts.

Finally, there is the issue of overhang.

### What Is Overhang?

As call volume begins to increase, there is a gradual rise in occupancy. On the other side of this, when volume starts to decrease, there is a gradual fall in occupancy. Overhang refers to the calls which begin in one interval but which are still live into the next interval.

This means that contact centres should avoid measuring intervals that are less than double their AHT – otherwise they will consistently have too many agents engaged in calls from the previous interval to handle calls in the new interval.

However accurate the forecast may have been on paper, it will be clear that there are too many incoming calls for the available staff to handle.

So, while a fifteen-minute interval is preferable for some medium or large contact centres, that’s only the case when their AHT is below seven and a half minutes.

**Mean Absolute Percent Error **

The Mean Absolute Percent Error (MAPE) measures the error as a percentage of the actual value, which is calls offered.

To begin, we simply calculate the percent error of each interval.

We then calculate the mean average of the percent errors for the data set to get the MAPE.

Interval | Calls Offered | Forecast | Absolute Error | % Difference |
---|---|---|---|---|

8:00 am – 8:30 am | 105 | 102 | 3 | 2.9 |

8:30 am – 9:00 am | 128 | 135 | 7 | 5.3 |

9:00 am – 9:30 am | 136 | 138 | 2 | 1.5 |

9:30 am – 10:00 am | 167 | 145 | 22 | 14.1 |

10:00 am – 10:30 am | 197 | 172 | 25 | 13.6 |

10:30 am – 11:00 am | 213 | 245 | 32 | 13.9 |

11:00 am – 11:30 am | 220 | 222 | 2 | 0.9 |

11:30 am – 12:00 am | 194 | 203 | 9 | 4.5 |

Total | 1360 | 1362 | 2 | 0.1 |

MAPE | 7.03 |

MAPE is a useful way to communicate forecasting data across a business, because the result is expressed in percentage terms which are more likely to be meaningful than an actual volume figure.

Because it’s a percentage, percent error is often confused with percent difference, which we used in the first example. To complicate matters, both calculations will often return the same result when the absolute error is small enough.

Using our first interval as an example, both the percent difference and the percent error are 2.9%. Of course, this does not mean they are interchangeable, and the larger the error or the data set, the greater the variance.

MAPE can be problematic for small contact centres because it is very sensitive to scale. While this is unlikely to be an issue for medium or large contact centres, MAPE can return unreliable data for contact centres with limited forecasting data to feed in.

### Mean Absolute Deviation

Mean Absolute Deviation (MAD) is one possible alternative for small contact centres that have difficulty using MAPE. It’s simply the mean average of the errors (or ‘deviations’) for the data set. Notice that we are still using the absolute error here, which is important for returning meaningful results.

Interval | Calls Offered | Forecast | Absolute Error |
---|---|---|---|

8:00 am – 8:30 am | 105 | 102 | 3 |

8:30 am – 9:00 am | 128 | 135 | 7 |

9:00 am – 9:30 am | 136 | 138 | 2 |

9:30 am – 10:00 am | 167 | 145 | 22 |

10:00 am – 10:30 am | 197 | 172 | 25 |

10:30 am – 11:00 am | 213 | 245 | 32 |

11:00 am – 11:30 am | 220 | 222 | 2 |

11:30 am – 12:00 am | 194 | 203 | 9 |

Total | 1360 | 1362 | 2 |

MAD | 12.75 |

MAD is more robust than MAPE when it comes to very small data sets, although it returns data in terms of actual values rather than a percentage. This should be fine for the forecaster themselves, but it can make communicating results to other areas of business trickier.

### Standard Deviation

This measure is generally recognised as one of the most useful tools that forecasters have at their disposal. It represents the spread of the data, standardising the deviation (error) from the apparent norm.

Unlike the other measures we have looked at, Standard Deviation is a reasonably complex process to perform manually, making it prone to error. In most cases, Standard Deviation is calculated through WFM tools or an Excel spreadsheet. It is worth noting that there are variations on the standard deviation formula, each useful for different kinds of data sets. In forecasting accuracy we are most interested in ‘population standard deviation’.

This is the equation for population standard deviation:

*[You can look at how to perform this Excel function here.]*

There are five steps to calculating Standard Deviation:

- Find the mean of the data set
- Find the distance from each data point to the mean, and square the result
- Find the sum of those values
- Divide the sum by the number of data points
- Take the square root of that answer

Our data set is the errors rather than the absolute errors, meaning that we will be using positive and negative numbers.

First we find the mean of our data:

3 + -7 + -2 + 22 + 25 + -32 + -2 + -9 / 7 = -0.25

Then we find the distance from each data point to the mean and square it:

Error | Distance from Mean | \(x^2\) |
---|---|---|

3 | 3.25 | 10.56 |

-7 | 6.75 | 45.56 |

-2 | 1.75 | 3.06 |

22 | 22.25 | 495.06 |

25 | 25.25 | 637.56 |

-32 | 31.75 | 1008.06 |

-2 | 1.75 | 3.06 |

-9 | 8.75 | 76.56 |

Sum | 2279.48 |

Next, we find the sum of the squared values, which is 2279.48, and divide it by the number of data points, getting 284.94

Finally, we get our result by finding the square root of that value, which is 16.88. This is our Standard Deviation for the data set.

There is another useful application of Standard Deviation. Rather than using errors as the data set, forecasters can use the actual contact volumes. The Standard Deviation figure result will be a representation of the general level of volatility in call volume over time.

This can be used to inform the amount of variability the forecasters need to build into their scheduling. And, if volatility is high, this can also be the figure they use to explain why there are forecasting errors.

### Correlation Coefficient

When a forecast contains errors, it is important to establish what other variables were linked to the unexpected increase or decrease in contact volume. Calculating the correlation coefficient of data sets is an effective way to this.

The correlation coefficient indicates the degree to which the movement of one variable affects the movement of another variable. If *x* goes up, how does that affect *y*?

The result is a number between -1 and +1 indicating something between a strong positive correlation and a strong negative correlation. This is the formula for correlation coefficient:

This is another calculation that forecasters are very unlikely to perform manually due to a very large capacity for error. Here’s an overview of how to reach the correlation coefficient for any data sets:

- Define two data sets, in this case the contact volume for two separate days. We’ll call them
*x*and*y* - Find the mean of
*x*and the mean of*y* - Subtract the mean of
*x*from every*x*value (a), and the mean of*y*from every*y*value (b) - Calculate a × b, a² and b² for every value
- Sum up a × b, as well as a² and b²
- Divide the sum of a × b by the square root of [(sum of a²) × (sum of b²)]

*[Alternatively, you can read how to perform this function on Excel here.]*

**With thanks to:**

*The following ideas have been discussed with the industry experts pictured below: *

*Christine Stubbs, WFM Contact Centre Consultant at Maintel*

* Charles Adams, Customer Service Operations Manager at Ordnance Survey*

* Penny Reynolds, Founding Partner at the Call Center School*

Published On: 16th Aug 2017 - Last modified: 14th Nov 2018

Read more about - Customer Service Strategy, Forecasting, How to Calculate, Workforce Management (WFM)