This feature is in official beta.
As the name suggests, the Forecast feature is used to forecast the duration of the corresponding initiative. To see it in action, open an initiative anywhere in the account and go to the tab called “Forecast”.
The forecasting feature is going to be as good as your historical data is. If you have a stable, predictable process, your forecasts will be very accurate.
However, if your cycle times fluctuate too much, if you allow your work to age indefinitely, if you don’t have WIP limits and continuously start new work without having finished other work first, your forecasts will be off.
In order to obtain a forecast, you need to have at least one child card (living in a Cards workflow) linked to the initiative. When the simulation starts, it will find all child cards of the initiative that live on the lowest level (in a cards workflow), it will find similar cards that were finished in the past; then it will run thousands of samples, trying to project the duration of the whole initiative. The result is visualized on a horizontal bar chart.
The graph provides four estimations: Optimistic, Realistic, Pessimistic, and Very Pessimistic, all measured in days.
If the initiative is not started, the forecast should be read as:
“Once the initiative is started, it will take between X and Y days to be finished”.
If the initiative is already started, the forecast should be read as:
“This initiative will take between X and Y days to be finished”.
If the initiative is finished, the forecast can be used to calibrate your forecasting skills. In other words, we provide a forecast that you can compare with the actual duration and reflect on.
Understanding the Settings
The Forecast feature provides four primary settings:
- Primary Dimension
- Secondary Dimension
This parameter instructs the algorithm to use as input data the historical cycle times up to the “Certainty” percentile. The cycle time values above that threshold will be ignored.
The lower the percentage is, the shorter the estimated duration is going to be, however, the likelihood of getting it is also going to be lower.
To learn more about cycle time percentiles, please review the Cycle Time Scatterplot article.
This is one of the most important parameters, as it accommodates many factors that can influence the forecast. The lower the value of the parameter, the longer the duration.
A value of 1 assumes the following:
- all child cards are “normal”, meaning that they are likely to take the same time as your historical data
- The child cards are worked one after the other (sequentially)
- The next child card is started right after the previous has been completed
It is obvious that the assumptions above will not be fulfilled all the time. In that case, change the scale value to reflect your reality.
An example of a property that affects the scale parameter is “complexity”. If you know with a good certainty that the upcoming work is more complex than the typical task, you may want to set a low value (e.g. 0.5) to indicate that the initiative is likely to take longer. A scale parameter of 0.5 will increase the duration of the forecast by approximately 50%.
Another example: If you know that many people will work on the child cards simultaneously, you may want to increase the value of the scale parameter to reflect the desired level of parallelism. If you set the value to two, the forecasted duration will be approximately half the normal one. One caveat to that, though! If ten people work on the initiative, but they are all allocated only 10% of their time, the scale parameter should still be 1 and not 10.
With respect to sequencing, if you expect to have gaps between finishing one child card and starting the next, set a lower value for the scale parameter.
Quite often, all of these factors will take effect to some extent and that’s the “art” of forecasting - you need to develop a good intuition about the “scale” parameter. This can only happen if you know your work and process really well. The rest is trial and error.
(3) Primary Dimension
It was mentioned above that the algorithm will find similar cards that have been completed and will build the forecast based on this data set. Use the primary dimension to instruct the algorithm on what the most significant factor to search by is.
For example, if you know that work type A inherently takes longer than work type B, set “type” as a primary dimension.
(4) Secondary Dimension
The secondary dimension is the very same thing as the primary dimension, it just adds another slice to the data.
For example, if you know that user A will be much faster than user B, set “owner” as the secondary dimension.
Understanding the Advanced Settings
The Forecast feature provides seven advanced settings:
- Additional cards
- Remove values <
- Remove values >
- Historical data
- Keep outliers
- Forecast all children
By default, the simulation will run 1000 samples (it’s a Monte Carlo simulation). Change this parameter if you need a different number of samples.
(2) Additional Cards
Sometimes scope gets added to an initiative/project at a later stage. If you want to account for such additional work, enter how many child cards you expect to be attached to this initiative in the future.
The additional cards will be distributed proportionally with respect to the primary and secondary dimensions.
(3) Remove Values <
If you know that all child cards will take X days or more, you can exclude historical data with smaller cycle times. E.g. if you set this value to 2, historical cycle times smaller than two days will be excluded from the data set. This will usually result in a longer duration forecasted.
(4) Remove Values >
If you know that all child cards will take X days or less, you can exclude historical data with greater cycle times. E.g. if you set this value to 10, historical cycle times greater than ten days will be excluded from the data set. This will usually result in a shorter duration forecasted.
(5) Historical Data
By default, the simulation will get the last 100 finished cards and will use them to build the forecast. If you want to go farther back in time, you can increase this value. However, this is rarely recommended, as the distant past is rarely similar to the present.
The forecasts are heavily dependent on the Primary and Secondary dimension parameters. When the historical data is collected, it determines the corresponding data sets based on these two parameters.
If you've configured the primary dimension to be Owner and the secondary to be Type, it will generate a list of all possible combinations (based on what child cards are linked to the initiative) and then you will get historical data for each pair separately. If you have two child cards with owners A1 and A2 and types T1 and T2 respectively, the forecast calculation would need to collect the following groups of historical data:
- Cards that are finished and had the A1/T1 parameters (data set 1)
- Cards that are finished and had the A2/T2 parameters (data set 2)
If you have four cards with the combinations A1/T1, A1/T2, A2/T1, A2/T2, then the forecast calculation would need to extract four data sets of historical data. If the cards are more and the parameters are more, the number of historical data sets might grow significantly.
We have the lowest possible limit of 5 entries per one of these data sets. This means that the forecast will ignore data sets with less than 5 historical items, as they would be statistically invalid.
(6) Keep Outliers
By default, the algorithm will clean up the outliers from your data set. If you need all data points in the data set, check this option.
(7) Forecast Children Only
This option is useful if you want to forecast an initiative without considering all other cards on the board. This can be used for expedited requests, as it will assume that no other work blocks the execution of the child cards. Another scenario is when the people working on the initiative have no other assignments and will not be affected by the existing work.
(8) Forecast All Children
This option is useful if you want to calibrate your forecasting skills. The parameter makes sense when you want to compare the forecast with the actual outcome. In fact, if you open the forecast tab of an already finished initiative, this option is selected and cannot be unchecked.
* Return Scale
This is another option to be used for calibrating your forecasting skills. It is active only for finished initiatives and its purpose is to show you what Scale parameter would have resulted in a correct forecast. Use this option to teach yourself what scale values to use in the future.
Does the size of the cards matter?
Generally speaking no. What matters is how much time the cards actually take on the board. If two cards take the same time to finish (cycle time), they will mean the same to the algorithm, even if their sizes are very different.
Does the algorithm take into account the sequence of the initiatives on a timeline?
In the scenario where you have a parent card with two or more child initiatives on a timeline, the algorithm will not account for the sequencing. This is a known limitation, which will be addressed in future releases.