Forecasting 101

Quantitative forecasting involves the use of algorithms which usually assume that history repeats itself. It begins with a process of decomposing time series data into four components:

The objective of algorithms is to remove noise, which can be thought of as random behaviour, seldom predictable. Level is an estimate of the mean or average; trend suggests an upward or downward trajectory, seasonality refers to periodic fluctuations that usually occur at the same frequency and within similar intensity; noise is merely an irregular behaviour not consistent with previous trends. Depicted below:

Selecting an algorithm for time series formulation can be challenging. I recall one of my first lessons in forecasting “All algorithms are wrong, some are useful”. There are a number of algorithms to choose from, broadly classified as State space, Regressive or Neural net. Below are a few examples:

Algorithms	Advantages	Disadvantages
Fourier	- Good for stable data - Fits linear trend - Has 3 terms (Level, Trend & Seasonality)	- Need 18 to 24 months data for reliable model - Mostly black box
MLR: Multiple linear regression	- An extension of Fourier, therefore good for stable linear trend	- Require causal| external data
Moving Average (MA)	- Simple & easy to apply - Need little history - Can be weighted to recent history	- No trends - Seasonality can be applied but is modelled outside.
AVS Graves: Adaptable variable smoothing	- Less sensitive to outliers - Can model cyclic patterns - Utilised for optimistic projections	- Produces flat line forecast - Based on smoothing, not suited for non-continuous demand
Crostons:	- Good for intermittent demand	- Non seasonal - Flat line forecast

The above methods are reasonably easy to understand and apply. Box Jenkins methods for Auto regressive applications, however, can be a little overwhelming for beginners.

Having a forecast is often not the issue. Anyone can call a number. Whether through SWAG (“scientific wild ass guess”) or a well thought out guesstimate. Super forecasters however, prefer a structured approach. It is advisable to set up a forecast practice and adopt a cyclic process that includes forecast generation, reconciliation and evaluation. Below are some guidelines:

• Begin with an understanding of the purpose of the forecast
• Assess the environment
• Determine what needs to be forecasted
• At what level
•  How often?
• Quantify factors affecting demand
• Identify data sources, availability, completeness & consistency
• Run a classification process – this would profile your forecast unit as: continuous, seasonal or non-seasonal.

Selecting the algorithm, building & evaluating the model requires specialist knowledge. Forecasting is about chance & change. To gain credibility, a sound framework will go further than a complex algorithm. The CPDF certification equips one with practical and a structured approach to forecasting. For more information visit the CPDF Training website: PEER Forecasting