Manage forecasts

Use Manage forecasts to find the forecast model that fits your team best. On the Review tab, you can see how accurate each model would have been on your own past volume, compare models side by side, and make the best one your current model. Your current model drives the forecasts on the Staffing timeline and the required staffing calculated from them.

To open it, go to Forecast > Manage forecasts.

The page has two tabs. This article covers Review. On Adjust, you add outliers, adjustments, and totals: see Forecast outliers and adjustments. To upload your own forecast, select Upload totals: see Forecast totals. For how each model type works, see Forecast models.

Navigating the Review tab

Below these are the Accuracy comparison table and a chart of forecasts and actuals. See Read the Accuracy comparison table and Check the forecasts and actuals chart.

How forecast models work

• Each combination of channel, queue, and metric has its own current model. For email, for example, New cases, Reopened, and Case throughput each have a model, so choose one for each.
• A model set as current for one queue doesn’t apply to other queues in the channel. To use the same type of model for two queues, set it for each queue.
• The Review tab works on past dates. For each model, it shows what the model would have forecast for those dates and compares that with your actual volume.
• Outliers are left out of the accuracy calculations. To add or change them, use the Adjust tab.

How a simulated forecast lock works

Forecasts are regenerated every night, so a forecast for a date keeps changing until the date arrives, unless your company has a forecast lock. Simulate forecast lock shows how accurate each model would have been if its forecasts had been fixed a number of days ahead. Set Days in advance to how far ahead you build schedules: for example, 30 if you create schedules a month in advance, or 0 or 1 if you staff the day before.

Historical forecasts

The Historical model row and the Historical forecast line show what your forecasts actually were on the Staffing timeline before each date, whichever model was current then. Use them to compare what you forecast with what another model would have forecast.

Because they’re a record of what happened, they can’t be recalculated for a different lock. They appear only when Days in advance matches your company’s forecast lock, which is 0 days unless our team has set one. For example, with a 7-day lock, they appear when Days in advance is 7.

The historical forecast values match the ones in the Forecasted vs actual report. That report shows what was forecast against what happened in more detail, for each interval, with more metrics.

Set up the comparison

Choose settings that match how you plan. For example, if you schedule in hourly blocks a month ahead, compare models hourly, as they would have forecast 30 days ahead.

1. Go to Forecast > Manage forecasts and select the Review tab.
2. Choose a Channel. To evaluate one queue, choose a Queue too.
3. With the metric toggle, choose the metric to evaluate. The options depend on the channel:
   • Email: New cases, Reopened, and Case throughput.
   • Chat, phone, and SMS: Contacts received and Average handle time.
   • Social and back office: Contacts received and Case throughput.
4. Select the date range, and choose a preset such as Last 30 days, or Custom to enter Start and End dates.
5. Choose the interval: 15 minutes, 30 minutes, Hourly, Daily, Weekly, Monthly, or Quarterly. Shorter intervals are available only for shorter date ranges: 15 minutes for ranges under two weeks, 30 minutes under a month, and Hourly under three months.
6. Optional: to compare models as they would have forecast some days ahead, open More forecast actions (the arrow next to Upload totals) and select Simulate forecast lock. Enter Days in advance, from 0 to 90, and select Apply. See How a simulated forecast lock works.

The table and chart update as you change each setting.

For example: to find the best model for new emails in Tier 1 queue, where you build schedules 14 days ahead in 30-minute blocks, choose Email, Tier 1 queue, and New cases, set the date range to Last 14 days and the interval to 30 minutes, and simulate a forecast lock of 14 days.

Note: If your company uses the case lifecycle model, the metric options are named for cases or units of work, such as Case volume and Unit of work volume.

Add models to the comparison

1. Below the Accuracy comparison table, select Select a model for comparison.
2. Choose a model. The list shows the models set up for the channel, queue, and metric you’ve selected.

The model is added to the table and the chart. To take it out again, open its Model options menu (three dots) and select Remove comparison. This only removes it from the comparison; the model isn’t deleted.

Note: To add a model that isn’t in the list, such as another type or a different number of weeks, contact our team.

Read the Accuracy comparison table

• Each row is a model, and the model in use is marked Current model:.
• The Historical model row shows what your forecasts actually were at the time. See Historical forecasts.
• The columns are Mean absolute error, Over error, Under error, and Mean error. Each shows the error in the metric’s units, such as Contacts, and as a Percentage.
• To sort the table, select a column heading.
• The Most accurate tag marks the model with the lowest mean absolute error among your current model and the models you’re comparing.
• Below the table, a line says how many outliers were left out of the calculations. Select View outliers or Edit outliers to open them on the Adjust tab.

A value shows as - when there isn’t enough data to calculate it, for example when no volume came in during the selected range. Choose a different date range, queue, or channel.

Choose the model that fits your goals

The Most accurate tag is a good place to start, because mean absolute error measures how far off a model is on average, in either direction. Depending on how you staff, another measure may matter more:

• To avoid being understaffed, prefer a model with a low Under error and a positive Mean error. It forecasts at or above your actual volume more often.
• To avoid overstaffing, prefer a model with a low Over error.

Check the model’s line on the chart too, to see whether its errors cluster on certain days or times. As your volume changes, come back to this page from time to time to check that your current model is still the best fit.

Check the forecasts and actuals chart

The chart below the table, such as Contacts forecasts and actuals, shows the historical forecast, your current model, each model you’re comparing, and the actuals, over the date range and interval you’ve chosen. Outliers are marked.

To export the data, open More forecast actions (the arrow next to Upload totals):

• Download forecasts: a CSV of the forecast values for your current model and the models you’re comparing.
• Download forecast metrics: a CSV of the accuracy metrics in the table.

Set a model as current

Setting a model as current makes it the model your forecasts use for that channel, queue, and metric.

Important: Setting a model as current changes the forecasts on the Staffing timeline and on this page, and the required staffing calculated from them. Check the channel, queue, and metric before you confirm.

1. In the Accuracy comparison table, open the model’s Model options menu (three dots).

2. Select Set as current model. For the model tagged Most accurate, you can also select Set as current model under its name.

   

3. In Promote model, check the Channel, Queue, and Forecast on details, then select Confirm.

A message confirms that the model has been promoted, and it’s marked Current model: in the table. Forecasts are regenerated with the new model. To go back, set the previous model as current in the same way.

The Model options menu also has:

• Delete model: permanently deletes the model for your company, and can’t be undone. You can’t delete your current model.
• Edit forecast parameters: for Prophet models only. See Forecast models.

How accuracy is calculated

Each metric compares the forecast with the actual value in every interval of the date range, leaving out outliers.

For example: a model forecasts 10 contacts in each of four intervals, and the actuals are 8, 12, 10, and 6. The differences are +2, −2, 0, and +4.

• Mean absolute error = (2 + 2 + 0 + 4) ÷ 4 = 2 contacts, and its percentage = 8 ÷ 36 = 22%.
• Over error = (2 + 4) ÷ 2 = 3 contacts, from the two intervals where the forecast was too high.
• Under error = 2 ÷ 1 = 2 contacts, from the one interval where it was too low.
• Mean error = (2 − 2 + 0 + 4) ÷ 4 = +1 contact, so on average the model forecasts slightly high.