Why the Simple Method Usually Wins
Complex methods beat simple ones in-sample and rarely out of sample. The three reasons, and the conditions under which complexity does pay.
A simple method usually beats a complex one on out-of-sample workforce demand, and the reason is not that complexity is bad. It is that most of the variation in a demand series is noise, that noise cannot be forecast, and that the extra flexibility of a complex model is spent fitting it.
The method in “Why the Simple Method Usually Wins” is only as useful as the definitions and source records behind it. A team exploring view the complete workflow description for managing freelance project time can add timely work and project context, with transparent settings, limited access and a documented correction route around every consequential use.
This is well established and it is persistently resisted, because a simple method is hard to be proud of.
Additional context for “Why the Simple Method Usually Wins” is available from Azure Machine Learning resources. External material is most useful when it prompts a specific validation check rather than replacing local judgement.
The three reasons
Noise gets fitted. Parameters get estimated from too few observations. And a complex model degrades faster when the world changes, because it has more assumptions to be wrong about.
- Noise gets fitted and will not repeat.
- Parameters get estimated from too few observations.
- A complex model degrades faster when the world changes.
All three get worse as the horizon lengthens and as the series gets shorter.
In-sample and out-of-sample
A model with more parameters will always fit history better. That is arithmetic, not evidence.
The only measurement that means anything is performance on periods the model did not see, forecast with only the information that would have been available. Anything else is a description of the past.
The honest test
Hold back the last several periods. Fit on the rest. Forecast the held back periods using only prior information. Compare to a weighted average built the same way. Repeat at several hold-back points.
A morning's work, and it settles most method arguments permanently.
What the test usually shows
The simple method within a point or two of the complex one at the near horizon, ahead of it at the far horizon, and far ahead of it in the periods following any change.
Where the complex method wins, it usually wins because of a driver, which can be added to the simple one.
Complexity that does pay
A named, knowable, large driver. Multiple interacting seasonal cycles in a long, clean series. Genuinely high-frequency data with real structure. Hierarchies that need reconciling.
Four conditions, and most operations meet none of them.
The maintenance cost nobody prices
A complex model needs re-fitting, monitoring, and somebody who understands it. When that person leaves, the organisation has a model it cannot change and will not retire.
That cost is real, recurring and rarely in the business case.
The explanation cost
A decision maker who asks why the number is 126 deserves an answer. "It is the average of comparable recent periods, raised four per cent for the campaign on the ninth" is an answer.
"The model says so" is not, and it is the answer that eventually gets the forecast ignored.
Where complexity hides
Not usually in the headline method but in the preprocessing: the outlier rules, the smoothing, the automatic seasonal adjustment, the hierarchy reconciliation. Each is a decision nobody made deliberately.
Simplicity means the whole chain is inspectable, not just the last step.
The hybrid that works
A simple baseline, explicit named adjustments for known events, and a stated band. Every term visible, every term attributable, every term re-estimable on its own.
This is what most good workforce forecasting looks like, and it is rarely what gets specified in a project.
When a complex model is already in place
Do not remove it. Run the simple one alongside for two quarters and compare, in public. The comparison either justifies the complex model, which is useful, or it makes the case for change without an argument.
Either outcome is better than a debate about preferences.
What simplicity is not
An excuse for not looking at the data, not building an event calendar, not testing drivers. Most of the work in good forecasting is upstream of the method entirely, and a simple method does not reduce it.
The simplicity is in the arithmetic, not in the thinking.
Simplicity and the handover
The practical argument that usually lands. A simple method can be handed to a new planner in an afternoon; a complex one cannot be handed over at all, only inherited.
In a function with normal turnover, that difference decides whether the forecast is maintained or merely run.
What gets lost when the complex model stays
Not accuracy, usually. The ability to change anything. A model nobody understands cannot have a driver added, a break handled, or a horizon altered, so it stops responding to the operation.
Within two years that is a much larger cost than the point of accuracy it was adopted for.
The practical rule
Start with the simplest thing that could work, measure it honestly, and add complexity only where the measurement shows it pays more than it costs to maintain.
Very few forecasting functions have ever run that test, which is why so many of them maintain models they cannot explain and cannot retire.