IS6400 Chap.7 Exponential Smoothing and Time Series
Exponential Smoothing and Time Series
Recent evidence receives more weight
Simple exponential smoothing blends the latest observation with the previous level. Expanding the update reveals geometrically declining influence from older observations.
A larger smoothing parameter reacts faster but carries more noise; a smaller one moves slowly and can lag a genuine shift.
Residual structure chooses the next model
The method assumes a changing level without persistent trend or seasonality. Repeating calendar errors or long one-sided runs should not be hidden by retuning alpha.
Add structure only when forward validation at the business horizon shows that it improves future decisions.
Interpret alpha as a trade-off
The newest observation receives weight alpha and the previous level receives one minus alpha, so the two visible weights must sum to one. Repeated substitution produces geometrically declining influence from older observations.
A large alpha reacts quickly after a genuine shift but also follows isolated noise. A small alpha filters noise but can lag, making the cost of its delay part of the parameter choice rather than a purely statistical preference.
Preserve the time unit and missing periods
Daily, weekly and monthly aggregation expose different amounts of noise, trend and seasonality.
Choose the unit from the operational decision before tuning. A zero observation is not the same as no observation, and collapsing missing periods changes the distance between events.
Record stock-outs, closures, promotions and reporting changes because they can create apparent level shifts that a smoothing update should not automatically absorb as ordinary demand.
Promote a more complex model only with forward evidence
Compare several alpha values across rolling forecast origins at the real horizon and retain a naive benchmark.
Report absolute error and signed bias by ordinary and stressed regimes. Repeating calendar residuals call for seasonal structure; a persistent slope calls for trend; event-specific misses may need explanatory inputs. Test the repair on later origins separate from those that suggested it, and keep the simpler smoother when extra state does not improve future decisions consistently.
What this chapter covers
- 01
Latest observation and prior level
- 02
Weights that sum to one
- 03
Geometrically decaying history
- 04
Responsiveness versus noise
- 05
Chronological backtesting
- 06
Seasonal and regime residual patterns
Worked example · free
Compare two smoothing speeds
- 3Compute the slow update. The marks shown are not an official University mark allocation.
- 3Compute the fast update.
- 2Explain the noise trade-off.
Key terms
- Exponential Smoothing
- A recursive update that gives recent observations more influence while retaining geometrically declining history.
- Alpha
- The smoothing weight placed directly on the newest observation in a simple level update.
- Level Shift
- A persistent change in the baseline around which observations fluctuate.
- Seasonality
- A recurring pattern tied to a regular calendar or cycle.
- Nonlinear Time Series
- A time process whose response changes with state, regime or the magnitude of earlier movement.
Exponential Smoothing and Time Series FAQ
Does a larger alpha always improve a forecast?
No. It responds quickly to genuine change and also follows random spikes more closely. Choose alpha on forward validation at the required horizon and inspect which events drive the largest misses.
Why do old observations still matter?
Repeated substitution shows their weights decline geometrically rather than becoming zero at a fixed cut-off. The prior level carries the accumulated influence of earlier observations into each new update.
What does a repeating residual pattern imply?
A stable calendar pattern suggests missing seasonality, while a regime-specific pattern may require explanatory features or a nonlinear model. Retuning one level weight cannot represent every repeated structure.
Exam move
Compute the same update with three alpha values and rank them by responsiveness. Draw bars for the implied weights on the newest four observations. Then write what each residual pattern suggests: random scatter, repeating calendar movement, widening variation and persistent one-sided misses.
Always evaluate by rolling through time at the real horizon and report directional bias beside absolute error so inventory consequences remain visible. Expand one update recursively until the geometric decay is visible and confirm that current and prior-level weights sum to one. Repeat after changing from daily to weekly aggregation and explain why the optimal responsiveness can change.
On a rolling-origin table, compare each alpha with the last-observation baseline in ordinary and shock periods. Finish with the smallest structural repair justified by residuals and a separate later window on which that repair would be tested fairly. Add an oral drill: define alpha, calculate one update, describe lag, name the residual pattern that breaks the method and state the comparison baseline.