This feature is part of Assembled's Pro and Enterprise plan. Please see our Plans page for additional details about our Assembled plans and associated features, and please contact the Assembled team at support@assembled.com if you’re interested in using this!

Overview

This document describes the various forecast models offered by Assembled. Forecasts can be generated for each channel-queue pair. Furthermore, forecasts can either keep updating up until the relevant date (eg. October 4th forecast will populate 4 weeks out and continue to adjust until October 3rd) or can be locked N weeks out.

In addition to forecasts, Assembled lets you add adjustments and flag outliers.

Forecast model options:

N-Week Average

N-Week Average with Momentum

Seasonal

Prophet

N-Week Average

Overview

One of the most popular forecasting models, the N-week average simply uses the average from the past N weeks. It’s our recommended default forecast because it’s generally very accurate, and is the easiest to understand and audit. The N-week average model generates a forecast each night for 1 week in the future and then repeats that forecast into the future.

How it Works

You tell us what number to input for N (generally range from 2 to 12 weeks). Our standard suggestion to most customers is 8 weeks.

If we’re forecasting for the 9:00AM to 9:15AM interval next Monday for example, and we’re using the average from the past 8 weeks, Assembled will calculate the average number of contacts from 9:00AM to 9:15AM for Mondays over the past 8 weeks and apply that number as the forecast.

Who Should Use This Model

This model is very easy to understand and explain, generally very accurate for short term forecasting (~1 to 4 weeks) but decreases rapidly in accuracy when looking more than a few weeks or months out. If the past N weeks are fairly indicative of the following few weeks and you need to be able to easily explain the forecast to justify headcount, this is a good model for you. This model is also well suited for companies with relatively little data.

Although you can go back a few weeks to get the average, this model will not pick up on general upward or downward trends. If your overall contact volume has been steadily increasing, that will not be captured in this model. In other words, the forecasted number will never be higher or lower than the past N week highest or lowest numbers respectively, for that forecasted time interval.

To take a simple example, if the past data for the given time interval is: 2, 4, 6 and 8, this model would forecast that you will receive 5 contacts, even though the patterns would suggest that the forecast should perhaps be 10.

Furthermore, if you see large cyclical patterns or spikes on specific dates (eg. holidays), since this model only goes back a few weeks, it will not pick up on those. However, adjustments can easily be applied and are usually the way these expected surges are handled.

N-Week Average With Momentum

Overview

This model is very similar to the “N-week Average”, however, it accounts for general upward or downward trends over the past few weeks. The N-week average with momentum generates a different forecast every week. 

How it Works

Assembled will look back at N weeks of data and apply a line of best fit to the contact volume. Then, for a specific interval (eg. 9AM to 9:15AM on Monday), we look at how much 9AM on Mondays typically deviates from the other intervals. From there we project out the contact volume for a specific interval using the line of best fit and then we apply an increase or decrease for the specific interval based on how much it deviates from all other intervals over that period of time.

For example, if based on the line of best fit we would expect 9AM on a Monday to have 1,000 contacts but at that specific time of day the volume is typically 20% higher than average, we would forecast 1,000 x 1.20 = 1,200 contacts for that interval.

Who Should Use This Model

This model is best suited for organizations that are experiencing very strong, systemic growth that can be observed even within the past weeks. It also works if growth isn’t as strong, however, recent surges or dips that are indicative of a broader pattern may be incorrectly picked up.

Seasonal

Overview

This model is the most complex Assembled forecast model and takes into account over a year’s worth of data and captures seasonal forecast patterns. It’s well suited for organization that want to trade off simplicity and ease of explanation for higher forecast accuracy when prone to seasonal patterns. This model is also much more reliable for medium and long term forecasting given the amount of data it considers. 

How it Works

Assembled takes into consideration:

• Overall yearly growth
  • Last 12 month contact volume growth (line of best fit over the last 12 months)
• Week of the year arrival pattern
  • Week of the year over the last 12 months
• Day of the week arrival pattern
  • Day of the week over the last 10 weeks 
• Time of day arrival pattern
  • 15 minute interval of the day over the last 2 weeks

Assembled will first find the line of best fit using the contact volume from the last 12 months. This is done in order to capture the general upward or downward trend in contact volume. From there it will project the volume for a given interval. We then apply a percent increase or decrease for that interval based on the week of year, day of week and hour of day. In a very simplified formula, this looks like:

Forecast for 9AM to 9:15AM on Monday of the 5th week of the year =
Percent of volume for the year that comes in on week 5 =
x Percent of volume for the week that comes in on Mondays
x Percent of volume for the day that comes in between 9AM and 9:15 AM
x Projected contact based on line of best fit

In this formula , the week of the year is defined at the Nth 7 day interval of the year starting on January 1st, not calendar weeks.

Note: this is a simplified formula because there are nuances on how this formula is applied and edge cases to consider such as shorter weeks on the last week of the year. For instance, when looking at the week of year volume, we don’t actually just look at the percent that that week represents of all volume for the year. We look at how much that week represents as a percent of all volume of the year relative to the baseline weekly volume.

This is a description of the default Seasonal model. This model is highly customizable (in addition to the possible forecast adjustments), and can be tuned to account for nuanced patterns specific to your operations. For more information, visit Seasonal model configuration.

Who Should Use This Model

This model is best suited for organizations who have predictable seasonal patterns and/or need to account for steady growth or decline. Out of the box, this model will pick up on surges or dips in volume that happens over the course of days or weeks on the same weeks of the year. For example if you sell air conditioning units and see a surge from May through August every year, this model will automatically pick this up. However, it’s worth noting that out of the box, this model may not pick up on sharp surges or dips that happens every year over a period of less than a week (eg. extremely sharp Black Friday surges) or patterns that don’t follow a 7 day cycle (eg. the pattern repeats every 45 days). That said, this model is highly tunable and can be adjusted to account for patterns specific to your organization.

Naturally, this model is more complex than other models, which means that it can be more difficult to understand the details of the forecast calculation. If explainability is more important than accuracy based on seasonal or cyclical patterns, other models may be a better fit for your organization.

Finally, this model performs best when it has sufficient amount of data to analyze. Although it can produce reliable results with less than a year’s worth of data, it performs best when data goes back at least 12 months.

Prophet

The Prophet model is a sophisticated forecasting model designed to handle complex seasonal patterns and trends in your data. It utilizes the Prophet machine learning library from Meta, integrated directly with your data in Assembled. It's highly flexible and can be customized to fit the unique characteristics of your data, making it ideal for organizations looking to balance accuracy with the ability to incorporate external factors such as holidays or special events.

Prophet is particularly effective for medium to long-term forecasting, where seasonal patterns and trends play a significant role. The key trade-off against the seasonal model is the reduced explainability that comes with the trained machine learning model.

How it Works

Prophet decomposes time series data into three main components:

• Trend: Represents the underlying growth or decline over time. Prophet can handle multiple growth models, allowing it to adapt to different types of data trends.
• Seasonality: Captures recurring patterns on a daily, weekly, or yearly basis. Prophet can automatically detect these patterns or allow for manual input to account for known seasonal effects.
• Holidays/Events: Account for holidays that can impact the forecast causing unexpected dips or spikes.

For example, if you're forecasting contact volume for a specific interval (e.g., Monday at 9:00 AM), Prophet will consider the overall trend in your data, the weekly seasonality (e.g. Mondays or Fridays tend to be busier), and any holidays that might affect that day's volume. These factors are then weighted and combined to produce the interval forecast.

Who Should Use This Model

The Prophet model is best suited for organizations with complex, seasonally driven patterns and external events that influence their data. It's an excellent choice if you need to account for:

• Complex seasonality: Prophet can capture multiple seasonal patterns simultaneously, making it ideal for data with both weekly and yearly cycles.
• Growth trends: Whether your data shows consistent growth or decline, Prophet can model these trends accurately over time.

However, Prophet's flexibility comes with complexity. Your trained model uses machine learning to identify the correct settings for your forecasts, but this makes explainability more difficult. If ease of understanding and explaining the forecast is critical, or your data doesn't have complex seasonal patterns or external events, a simpler model might be more appropriate.

An important thing to note is that this model requires sufficient historical data to produce reliable forecasts. While it can work with shorter time series, it performs best with at least a year's worth of data, especially when accounting for seasonal effects and holidays.