|
Modelling Advertising and Promotion Effectiveness In this example, sales figures were analysed to see whether different kinds of advertising and promotions were paying for themselves. To isolate their real impact, any other factors influencing sales must also be taken into account. A 4Thought model of monthly sales was built, using season, TV advertising, radio advertising and promotional effort as model inputs. Establishing the inter-relationships between all the different inputs makes it possible to measure how each factor effects demand. Because 4Thought can find non-linear relationships, subtleties such as different media having a different kind of effect from one season to the next can also be measured. The model building process is an iterative one. Assuming at first there are no relationships between sales (the model output), and time, season, advertising and promotions (the model inputs), 4Thought gradually finds patterns in the data. It will then discover any linear relationships, followed by simple non-linear, and finally complex arbitrary non-linear forms. Each iteration 4Thought tests a sub-set of the data (the test set), using the remainder (the model set) to actually create the model. The model best predicting the test set is used for all the analysis and forecasting. Other, more complicated, and seemingly more accurate models can be found after many more iterations, but are misleading. While they might predict the model data better, they are a result of "over-fitting" where noise in the data is being modelled too, which results in an unrealistic model. Taking cross-sections through the model shows the relationships between sales and each model input, which reveals the relative impact of TV and radio advertising:
In the above graphic, the benefit of TV advertising is clear—a hundred units of TV advertising can push up sales by as much as 250%, while increasing the level of radio advertising has a relatively small effect. This information must be combined with knowledge about the costs of advertising and the variable profit per sale to work out the exact return on advertising investment in money terms. The same methodology may be applied to any other form of promotion activity:
Taking more cross-sections shows the quantified benefit of promotions, and the underlying seasonality. Using the most predictive, and hence realistic model to isolate the impact of each individual factor means that even if there is "missing" data, a clear view of each model input can be seen. Additional Cases
|
