Forecast models

Plans and permissions:

• Plan: Pro or Enterprise plan
• Edit Prophet parameters and change how far ahead forecasts are generated: Admin role, or a custom role with Adjust forecasts set to Can view and make adjustments to forecasts on the Forecast tab
• To edit Prophet parameters, a custom role also needs Review forecast models.
• To open Configure > Additional settings, a custom role also needs Configure staffing parameters.

Use this article to choose a forecast model. Each model turns your past volume into a forecast in a different way, so each suits a different pattern of demand and a different planning horizon. Assembled offers N-week average, N-week average with momentum, Seasonal, and Prophet models.

To compare models on your own data and choose one, go to Forecast > Manage forecasts. See Manage forecasts.

How forecasts are generated

• Each channel and queue has its own models, one for each metric you forecast, such as new cases and reopened cases for email, or contacts and average handle time for chat. A new queue starts with an 8-week average model.
• Forecasts are calculated for every 15-minute interval.
• Forecasts are regenerated every night, so a forecast for a future date keeps updating until that date arrives. To stop forecasts updating for the next few days or weeks, see Forecast lock.
• Forecasts are also regenerated when you save adjustments, outliers, or totals, or set a different model as current.
• Forecasts are generated from today through your company’s forecast window, 60 days by default. See Change how far ahead forecasts are generated.
• On top of any model, you can add adjustments for events you expect and mark outliers to leave out of the model’s history. See Forecast outliers and adjustments.

Compare model types

N-week average

The N-week average uses the average of the past N weeks. It’s usually accurate in the short term, and it’s the simplest model to understand and check. N usually ranges from 2 to 12 weeks, and 8 weeks suits most teams.

How it works

For each interval, Assembled averages the volume in that same interval on the same day of the week over the last N weeks. For example, with an 8-week average, the forecast for 9:00 to 9:15 AM next Monday is the average volume from 9:00 to 9:15 AM over the last 8 Mondays. Because every forecast comes from the same recent weeks, the forecast repeats the same weekly pattern into the future.

When to use it

Use it when the past few weeks are a good guide to the next few, and when you need a forecast that’s simple to explain, for example to justify headcount. It’s most accurate one to four weeks out, and works well when you don’t have much history.

Keep in mind:

• It doesn’t pick up trends: a forecast is never higher than the highest, or lower than the lowest, of the past N weeks for that interval.
• For example, if the last four weeks had 2, 4, 6, and 8 contacts in an interval, a 4-week average forecasts 5, even though the trend suggests 10.
• It doesn’t pick up yearly patterns or spikes on specific dates, such as holidays. Use adjustments for expected surges.

N-week average with momentum

This model works like the N-week average, but it also accounts for the upward or downward trend over the past N weeks.

How it works

Assembled fits a trend line to the last N weeks of volume and projects it forward. For each interval, such as 9:00 AM on Mondays, it then applies how much busier or quieter that interval usually is than average.

For example: the trend line projects 1,000 contacts for 9:00 AM next Monday, and 9:00 AM on Mondays is usually 20% busier than average. The forecast is 1,000 × 1.20 = 1,200 contacts.

When to use it

Use it when you’re seeing strong, steady growth or decline that shows up even over a few weeks. If growth is weaker, a recent surge or dip can be mistaken for a trend and carried forward.

Seasonal

The Seasonal model uses more than a year of history to capture seasonal patterns. It trades some simplicity for accuracy when your volume follows the time of year, and because it uses so much data, it’s more reliable for medium- and long-term forecasts.

How it works

The Seasonal model combines a growth trend with arrival patterns:

Assembled fits a trend line to the last year of volume and projects it forward, then adjusts each interval for the week of the year, the day of the week, and the time of day:

Forecast = projected volume from the trend line × week-of-year factor × day-of-week factor × time-of-day factor.

Each factor is how much busier or quieter that week, day, or time is than average. For example, the forecast for 9:00 to 9:15 AM on the Monday of week 5 uses the trend line’s projection, how busy week 5 usually is, how busy Mondays usually are, and how busy 9:00 to 9:15 AM usually is.

Week of the year counts 7-day blocks from January 1, not calendar weeks. This is a simplified view: the model also handles cases such as the short last week of the year, and measures each week against a baseline weekly volume rather than as a plain share of the year.

When to use it

Use it when you have predictable seasonal patterns, or need to account for steady growth or decline. With the default settings, it picks up surges or dips that last days or weeks and come at the same time each year, such as higher demand for air conditioning from May to August.

Keep in mind:

• With the default settings, it may not pick up very sharp spikes that last less than a week, such as Black Friday, or patterns that don’t follow a 7-day cycle, such as one that repeats every 45 days. Its settings can be changed to account for these.
• It’s harder to explain than the averaging models. If explainability matters more than capturing seasonal patterns, use one of those.
• It works with less than a year of history, but performs best with at least 12 months.

Seasonal settings

The defaults suit most teams. To fit patterns specific to your operations, our team can change them for your Seasonal models:

• Week of year, day of week, and time of day: change how much history each pattern uses, for example to capture higher volume in summer, no volume at weekends, or a lunchtime peak.
• Growth: change the number of weeks the trend line uses (52 by default).
• Week of month and day of month: add these patterns, for example when volume is lower in the first week of every month.
• Holidays: add US or Canadian federal holidays, so the model accounts for changes in volume around them.

To change your Seasonal model’s settings, contact our team.

Prophet

Prophet is a machine learning model, based on the Prophet forecasting library from Meta, that handles complex seasonal patterns and trends. It’s flexible, and it can account for external factors such as holidays and special events. It works best for medium- to long-term forecasts. Compared with the Seasonal model, it’s harder to explain, because its forecast comes from a trained model. In the list of model types, it’s called Facebook Prophet.

How it works

Prophet splits your history into three parts, weighs them, and combines them into the forecast for each interval:

• Trend: the underlying growth or decline over time, which can change direction.
• Seasonality: repeating daily, weekly, and yearly patterns, such as Mondays or Fridays being busier.
• Holidays and events: dips or spikes around holidays.

For example, for 9:00 AM next Monday, Prophet considers the overall trend, how busy Mondays usually are, and whether a holiday affects that day.

When to use it

Use it when your volume has complex seasonal patterns and is affected by external events. It can capture weekly and yearly cycles at the same time, and consistent growth or decline.

Keep in mind:

• If you need a forecast that’s simple to explain, or your volume doesn’t have complex seasonal patterns or events, a simpler model is usually a better fit.
• It needs plenty of history. It can work with less, but it performs best with at least a year, especially to capture seasonal effects and holidays.

Prophet parameters

You can tune a Prophet model’s parameters. A good starting point depends on how changeable your data is:

The same dialog has Seasonality mode (Additive, the default, or Multiplicative), and Changepoint range, the share of your history in which the trend can change (0.8 by default, so the trend doesn’t change in the most recent 20%).

To change the parameters:

1. Go to Forecast > Manage forecasts.
2. Select the Review tab.
3. Choose the channel, queue, and metric.
4. If the Prophet model isn’t your current model, add it with Select a model for comparison.
5. Open the Prophet model’s Model options menu (three dots) and select Edit forecast parameters.
6. In Update model, change the values, then select Confirm.

A message confirms that the model’s settings have been edited.

To let Assembled find the best values instead, select Optimize parameters in Update model. The search usually takes 30 to 60 minutes, and the model is updated with the best values when it finishes.

Change how far ahead forecasts are generated

Your forecast window sets how far ahead forecasts are generated for every model in your company.

1. Go to Configure > Additional settings.
2. In the Forecasting card, choose How far ahead forecasts are generated: 14 days, 30 days, 60 days (the default), 90 days, 120 days, or 180 days.

The setting saves as soon as you change it, and existing forecasts are regenerated.

In the same card, Ignore pre-launch history controls how a new queue’s history is used. It’s No by default. Set it to Yes to leave out the empty period before a queue’s first contact, so models don’t treat it as zero volume. Changing it also regenerates forecasts.