Thinking through the results of a modeling decision is a key part of ensuring good model performance; in other words, making sure the calculation engine isn’t overtaxed. This article highlights some ideas for how to lessen the load on the calculation engine.
Formulas should be simple; a formula that is nested, or uses multiple combinations, uses valuable processing time. Writing a long, involved formula makes the engine work hard. Seconds count when the user is staring at the screen. Simple is better. Breaking up formulas and using other options helps keep processing speeds fast.
You must keep a balance when using these techniques in your models, so the guidance is as follows:
- Break up the most commonly changed formula
- Break up the most complex formula
- Break up any formula you can’t explain the purpose of in one sentence
Formulas with many calculated components
The structure of a formula can have a significant bearing on the amount of calculation that happens when inputs in the model are changed. Consider the following example of a calculation for the Total Profit in an application. There are five elements that make up the calculation: Product Sales, Service Sales, Cost of Goods Sold (COGS), Operating Expenditure (Op EX), and Rent and Utilities. Each of the different elements is calculated in a separate module. A reporting module pulls the results together into the Total Profit line item, which is calculated using the formula shown below.
What happens when one of the components of COGS changes? Since all the source components are included in the formula, when anything within any of the components changes, this formula is recalculated. If there are a significant number of component expressions, this can put a larger overhead on the calculation engine than is necessary.
There is a simple way to structure the module to lessen the demand on the calculation engine. You can separate the input lines in the reporting module by creating a line item for each of the components and adding the Total Profit formula as a separate line item. This way, changes to the source data only cause the relevant line item to recalculate.
For example, a change in the Product Sales calculation only affects the Product Sales and the Total Profit line items in the Reporting module; Services Sales, Op EX, COGS and Rent & Utilities are unchanged. Similarly, a change in COGS only affects COGS and Total Profit in the Reporting module.
Keep the general guidelines in mind. It is not practical to have every downstream formula broken out into individual line items.
Plan to provide early exits from formulas
Conditional formulas (IF/THEN) present a challenge for the model builder in terms of what is the optimal construction for the formula, without making it overly complicated and difficult to read or understand. The basic principle is to avoid making the calculation engine do more work than necessary. Try to set up the formula to finish the calculations as soon as possible.
Always put first the condition that is most likely to occur. That way the calculation engine can quit the processing of the expression at the earliest opportunity.
Here is an example that evaluates Seasonal Marketing Promotions:
The summer promotion runs for three months and the winter promotion for two months.
There are more months when there is no promotion, so this formula is not optimal and will take longer to calculate.
This is better, as the formula will exit after the first condition more frequently.
There is an even better way to do this. Following the principles from above, add another line item for no promotion.
And then the formula can become:
This is even better because the calculation for No Promo has already been calculated, and Summer Promo occurs more frequently than Winter Promo.
It is not always clear which condition will occur more frequently than others, but here are a few more examples of how to optimize formulas:
FINDITEM formula
The Finditem element of a formula will work its way through the whole list looking for the text item, and if it does not find the referenced text, it will return blank. If the referenced text is blank, it will also return a blank. Inserting a conditional expression at the beginning of the formula keeps the calculation engine from being overtaxed.
- IF ISNOTBLANK(TEXT) THEN FINDITEM(LIST,TEXT) ELSE BLANK
Or
- IF BLANK(TEXT) THEN BLANK ELSE FINDITEM(LIST,TEXT)
Use the first expression if most of the referenced text contains data and the second expression if there are more blanks than data.
LAG, OFFSET, POST, etc.
If in some situations there is no need to lag or offset data, for example, if the lag or offset parameter is 0. The value of the calculation is the same as the period in question. Adding a conditional at the beginning of the formula will help eliminate unnecessary calculations:
- IF lag_parameter = 0 THEN 0 ELSE LAG(Lineitem, lag_parameter, 0)
Or
- IF lag_parameter <> 0 THEN LAG(Lineitem, lag_parameter, 0) ELSE 0
The use of formula a or b will depend on the most likely occurrence of 0s in the lag parameter.
Author David Smith.