Q: We are hiring an external SAP IBP Demand Planning Consultant. How do I know if he knows the subject.
A:
Ask which forecast models to use when and check for his clarity of responses. If he is not, then he is probably an SAP configurator. Not a consultant.
Constant Models – MVA, SES, – When the demand is stationary
Browns DES- When Demand does a course change (abrupt changes)
Holts Model – When Demand is Trended – up or down
Winters Model – When Demand has both Seasonality and Trend – up or down
ARIMA Models – When the data shows autocorrelation but no clear seasonality with a ‘manageable’ trend.
SARIMA – ARIMA + Seasonality
Regression Models – When some other variable with stable time series best explains the demand you need to predict.
Consensus. Forecast Agreed – When Evidence explains the accuracy of the forecast. E.g. Agreed Forecast = 0.4 x Stat Forecast + 0.2 x LYSP + 0.2 x Sales Forecast + 0.1 x Marketing Forecast + 0.1 x AOP. You can derive the coefficients using correlation between members and the actual sales. This works best in real world.
Bayesian Inference Models – When past belief / knowledge needs to be incorporated into your prediction. Unlike with Time Series where weighted or simple averages are used, Bayes model uses conditional probability. Time series does NOT include prior beliefs Time series models like SES and DES only product point forecast. Bayes model produces distribution of the predictions. Limited data is enough for Bayes.
Prophet Models – When time series contains multiple seasonality, trends, holidays, and missing data. Or when day or hour of the day is also a season. Useful in retail, e-commerce and QSR business
LSTM – Long Short-Term Memory – When the context of the input in known and important. When the noise in the data series is high. LOTS of transactions in per unit time. e.g. 10000 orders per minute. Probably useful for stock market punters and social media analytics.
Q: Ah is this all about AI / ML?
A: I do not know. Probably some of them call it such to sound cool or intimidate you with some mumbo jumbo.
