When is it Best to Model and Forecast by Segment?

Sometimes the data we want to forecast can be broken down by segment.  For example:

• Total births can be segmented by age group of mother;
• Airline passengers on domestic routes can be segmented by their origin - local or overseas.

Under what circumstances is it better to model individual segments rather than the aggregate market?  The answer depends to some degree what we mean by better - more accurate or more strategic, for instance.

In terms of accuracy, it is better to forecast by segments under certain conditions.  In particular, we want forecasting errors for each segment to compensate each other rather than reinforce each other.  This will often be the case when the segments are going in different directions, or are influenced by different explanatory variables rather than all behaving the same way.  As a case in point, age-specific birth rates in Australia have been increasing for women in their 30's but decreasing for women in their 20's and it is now more accurate to forecast by age group separately.  But in the 1970's they were going in the same direction and it would have been more accurate to forecast in aggregate.

Of course, we must have accurate measurement of demand by segment and accurate forecasts of future segment sizes.  This is often the case for demographics such as age group, but not for many segmentation bases.

Strategically, it is often valuable to forecast by segment.  This enables target setting, resource allocation, and identification of threats and opportunities.

For sensitivity analysis purposes it is often more accurate to model segments separately.  This is true when different segments are influenced by different factors - in which case the precision of regression estimates are higher if segments are separately modelled (assuming accurate measurement of demand by segment).  An example is segmenting demand on domestic routes on the basis of local or overseas origin.  Locally sourced traffic numbers will be influenced by local economic factors such as interest rates and internationally sourced traffic will be influenced by economic conditions in their home countries as well as currency exchange rates.

The potential of using segmentation to improve forecasting accuracy is illustrated by a forecast of Australian retail turnover growth in early 1995.  In the preceding five months, interest rates had risen by 2.75% in total.  Despite this, the forecast was that growth would accelerate until mid-1995 and then decline to very slow growth over the following year.  The model used recognised that two market segments react differently to interest rate rises:

• households with savings (mostly self-funded retirees) experience income growth and are able to spend more - and they react quickly;
• households with mortgages and other debt experience a reduction in disposable income but delay spending cuts by about 15 months - banks don't raise repayment levels immediately and these consumers try to maintain lifestyle for a period by running down savings and increasing debt.

While retail turnover data is not segmented, so these segments cannot be separately modelled, explicitly capturing the separate reactions in the aggregate model provided an accurate forecast.  Furthermore, the model correctly predicted the turning point - many models perform badly at this.