Several previous articles have: described the need to develop cost forecasting models, provided a step-by-step method for building cost models, shown how to combine data from many centers (or products) to develop aggregate cost forecasting models.
All of these earlier articles described the various aspects of cost modeling using multiple regression models. The implicit assumption was that the data was "stationary". However, real business data is hardly that. Most data that are needed to build cost models show a trend, seasonality and of course, noise or randomness. Readers who are familiar with forecasting in general are aware of this.
It is possible to combine time series analyses with the multiple regression models to develop robust cost forecasting models. The process involved consists of the following steps
Step 1: Identify the factors which are most relevant to objective. For example, the objective could be to predict weekly transportation costs or monthly production costs. In any case, the first step is to assemble all available data and determine the factors or variables which have the strongest influence on the objective. You can use something like the Principal Component Analysis or Mutual Information based key driver analysis for this.
Step 2: Build a multiple regression model to predict the objective using the short-listed variables identified in step 1 and test the quality of the models. This model establishes the functional form of the relationship between the objective and the independent input factors.
Step 3: For each factor identified in step 1, perform a time series analysis. The goal of this analysis is to identify trend, seasonality and cyclicality of the input factors.
Step 4: Using time series analysis, develop forecasts for the input factors. Techniques best suited for this would be Simple Exponential smoothing (if there is no trend or seasonality), Holt-Winters two parameter smoothing (if there is a trend but no seasonality), Holt-Winters three parameter smoothing (if there is trend and seasonality) or ARIMA.
Step 5: Combine the time series forecasts with the regression model developed in step 2. The schematic below captures the essence of this type of modeling. If only time series analysis was used to make forecasts - for example, simply using the objective (such as weekly spend) and developing a forecast as described in Step 4 - the output forecasts could be very noisy. As shown in the output, combining time series analyses with the regression model will help to produce robust cost forecasts.
In the next article we will describe how to use R to develop time series forecasts for each of the scenarios described in step 4.
Download our free whitepaper on cost modeling and cost forecasting.