One forecast model is seldom the best for every product and customer segment, yet that’s how most companies forecast. Leading analytical organizations apply different forecast models by segment to maximize accuracy and create their pricing strategy. To remain competitive, companies must implement this approach.
Simple exponential smoothing is one of the most basic models:
Though as the name says the model is simple, it is surprising how often it can be the best model to predict a segment with low and highly variable demand. Double, triple and seasonal exponential smoothing models can be tried to fit models with more complex autocorrelation patterns and seasonality.
Simple exponential smoothing belongs to the autoregressive, integrated moving average or ARIMA family of models. Notation for this type of models is usually given as ARIMA(p, d, q), where p is the number of autoregressive terms, d is the degree of first order differencing, and q is the order of the moving average. Simple exponential smoothing can be described as an ARIMA(0,1,1). For a complete explanation of ARIMA models see Forecasting Methods and Applications by Makridakis, Wheelwright and Hyndman.
The Fourier Decomposition model has shown itself to be a very reliable approach for estimating time series that exhibit strong seasonality.
In the equation above, P is the periodicity of the time series and K is the number of sin/cos terms. A trend component may also be added to this model. This approach is also often described as Spectral Analysis.
Models must be tested on a holdout set to ensure that they are not ‘over-fitting’ the historical data. Metrics, like MAPE, MASE and bias can be used to select the best model. Forecast accuracy metrics are critical and essential for any organization. For more on that topic, see my previous blog post (click here). Revenue Analytics has developed its own proprietary Pick Best Forecasting Module that tests scores of different models and parameter settings to select the best forecast for each product or segment automatically.