Skip down to the research record

STL settings for a monthly bookings trend: 4 checks

Quota Forecast Editorial Team · Published · 10 min read

Check four STL settings before reading a monthly bookings trend: data frequency, the seasonal default, the degree setting and the jump values. According to statsmodels, this monthly data (January 1959 to December 1987) has a clear trend and seasonality across the sample. According to statsmodels, the default of 7 for seasonal should be changed in most applications, and jump values should not normally be more than 10-20% of the size of seasonal, trend or low_pass.

Key takeaways

  • confirm that the series is monthly and shows a clear trend and seasonality before you read the line.
  • treat the seasonal default of 7 as a value to check, and change it in most applications.
  • set COMPONENT_deg to 0 when you want a constant only.
  • keep jump values at or under the 10-20% range for seasonal, trend or low_pass.
  • read a level with a Number visualization and a change from period to period with Metabase's Trend chart, or the equivalent in your tool.

Why a monthly trend line matters before the call

A monthly bookings trend gives you a line to read beside the number on the forecast call. The line is easier to discuss when you know which settings produced it, so this checklist covers four settings you can confirm before you open the chart. The goal is to confirm the lines behind the number, not to tune them toward a particular result.

The statsmodels example is a useful reference for the vocabulary, but its data is not yours. According to statsmodels, the example in Cleveland, Cleveland, McRae, and Terpenning (1990) uses CO2 data. A bookings series is not CO2 data, so treat the example as a worked case of the settings rather than as a benchmark for revenue.

On a forecast call, the trend line can sit beside the headline number, and a question about it is easier to answer when you know which settings produced the line. A reader who asks why the line bends at one point should get a plain answer from your notes rather than a guess. That is the practical reason to confirm the settings before the meeting.

Check that your data is monthly with trend and seasonality

Start with one series that has one value per month, in date order. Check for missing months and for months that were entered twice, and fix them before you run any decomposition. Plot the series and look for two things: a long-run direction and a pattern that repeats through the calendar year. If the line is flat and noisy, write that down before you change any setting.

Record the first and last month of the series in your notes, so the record matches the data you actually read. If the monthly totals come from closed deals, the bookings reconciliation checks for quarter close are a separate review to run first.

Before you open the decomposition tool, count the months in the series and compare that count with the period you intend to read. A chart that covers two years tells a different story from one that covers a decade, so label the period on the chart as well as in your notes. Store that label with the settings, so anyone who reopens the chart can see which months it covers.

If the series mixes several sales motions, such as new business and renewals, decide whether the trend should cover one motion or both. Write that choice in the notes before the run, so the reader on the call knows what the line measures.

Before you run: check the seasonal default of 7

Find the seasonal setting in the tool you use and note whether it still holds the default of 7 that statsmodels describes. Treat the default as a value you are checking, not a value you accept by leaving it alone. The checklist does not name a replacement value for your series, so choose one with the person who owns the bookings data and write the reason beside it.

A useful test is to run two seasonal values and compare the two trend lines side by side. That comparison shows how sensitive your chart is to the choice. It does not show that either value improves forecast accuracy, so do not present the result that way on the call.

Write the seasonal value you plan to use in your notes before the run, next to the default of 7 that the example reports. Keep the decision specific. Name the value, the person who chose it and the date, so a later review can see why the setting moved away from the default. If you leave the setting at 7, record that you checked it and kept it, so the record shows what was actually run.

Change one setting at a time. Changing the seasonal value, the degree and the jump values together makes it harder to trace which setting produced a given change in the trend, so the other settings can wait until the seasonal run is recorded.

While you run: set the degree and keep jumps small

The degree setting needs a careful read. According to statsmodels, the degree makes little difference in that example, apart from the trend around one sharp change. Read that as a prompt to inspect the trend around any sharp movement in your own series, and do not treat it as a rule for bookings.

For a constant alone, the documented route is to set COMPONENT_deg to 0, according to statsmodels. Keep the other components in your notes so the reader can see that the constant was a deliberate choice.

Jump values come last. Compare each jump value with the size of the component it applies to. Write each jump value down next to the seasonal, trend or low_pass size it was checked against.

Run the degree comparison on a copy of the series, so the main run stays as you recorded it. Compare the two trend lines, and note the months where they split. Those months are the ones to explain if the trend is questioned.

In your own series, look at the months around any sharp movement and check them against your sales records. Items such as quarter closes, pricing changes or a change in the sales team are worth checking, and you can only confirm them from your own records.

Jump values are easiest to check in a table in your notes, with one row for each component, the jump value, the component size and the ratio between them. The ratio is the figure to hold against the 10-20% limit, which according to statsmodels should not normally be exceeded. If a jump value is above that range for a component, reduce it in the STL run and record the change in your notes, rather than leaving the value unexplained.

Illustrative example: jump values against a seasonal size of 20

Suppose the seasonal component of a series has a size of 20. A jump value of 2 is 10% of that size, and a jump value of 4 is 20%, so both sit inside the range and can be kept. A jump value of 6 is above 20% of 20, so reduce it in the STL run, record the change in your notes, and write down the seasonal size it was checked against. These numbers are illustrative only and come from no real series.

After you run: read the trend with a Number or Trend chart

Decide first whether the question is about the level or about the change. A level can be read from a single figure, while movement from one period to the next needs a chart that shows the change.

If your chart tool is Metabase, according to Metabase, if the “Number” visualization has not already been applied by default, change the visualization type by clicking Visualization in the bottom left and selecting “Number” from the sidebar.

If the chart is built in Power BI, a refresh has a defined sequence. According to Microsoft Learn, whenever you refresh data, Power BI must query the underlying data sources, possibly load the source data into a semantic model, and then update any visualizations in your reports or dashboards that rely on the updated semantic model.

The table below lists one quoted point from each publisher used in this checklist.

PublisherSource pageQuoted point
statsmodelsSeasonal-Trend decomposition using LOESS (STL) example"The default value for seasonal is 7, and so should also be changed in most applications."
MetabaseNumbers documentation"If you want to display the change in the number from period to period, you can use the Trend chart instead."
Microsoft LearnData refresh in Power BI"Whenever you refresh data, Power BI must query the underlying data sources, possibly load the source data into a semantic model, and then update any visualizations in your reports or dashboards that rely on the updated semantic model."

Record which chart you used for each figure, so the call can tell a level from a change. After each Power BI refresh, check the date on the visuals before you quote them, since the sequence above ends with the visuals being updated.

Keep a record of settings with the forecast

Keep the record in the same place as your forecast notes, so the call has one set of settings to refer to. Include the following in each entry:

  • the series name, with its first and last month
  • the seasonal value you used, and whether it differed from 7
  • the degree setting, including whether COMPONENT_deg was set to 0
  • each jump value and the seasonal, trend or low_pass size it was checked against
  • the chart you used to read the change

Update the record each time you change a setting. If your team already keeps notes for the weekly call, add these lines to the stage duration history checks before the weekly call so the two records sit together.

Keep the earlier entry rather than overwriting it. The earlier entry shows which settings produced the line you presented last time, which is useful when the trend shifts between calls.

Attach the chart export or screenshot to the entry, and label it with the chart type you used and the date you ran it. A labeled export lets the next reader check the figure without rerunning the decomposition.

Name one person as the owner of the record, and have a second person read the entry before the call. The owner keeps the entries current, and the second reader checks that each setting in the list has a value.

Next step before the forecast call

Open your monthly bookings series and check the seasonal setting against the default of 7 before the next forecast call, changing it only if the bookings data owner agrees on a value. Before that call, also review the quote checks before the forecast call for a related set of checks.

Related guides: Check 6 caption settings before the weekly sales call · Quarter-end bookings: 9 billing and reporting limits.

FAQ

What is the default for seasonal?

The default is 7, according to statsmodels, and the same page says it should also be changed in most applications. Compare your own series with the example before you decide whether 7 fits. Write the value you chose in your record, even when it is 7.

How large can jump values be?

According to statsmodels, jump values should not normally be more than 10-20% of the size of seasonal, trend or low_pass. Check each jump value against the component it applies to, and note the comparison in your record. A value above the range is a reason to review that setting before you rely on the line.

How do I keep only a constant?

According to statsmodels, you can change the setting to only include a constant by setting COMPONENT_deg to 0. Record that choice with the other settings, so the constant-only run is easy to identify later. Label it clearly when you show it, because it is a different output from your main trend line.

How do I show period-to-period change?

According to Metabase, the Trend chart is the option for displaying the change in the number from period to period. Record which chart you used, so the call can tell a level from a change.

Sources