Data massaging adds error, just forecast!


In a recent meeting, the CIO of a leading commercial automotive company’s shared his experience of high complexity in managing forecasting data. I was not surprised. Often demand planners complain about managing forecasting data. I can relate to where there are coming from. It’s due to the approach prescribedData masaging adds error, just forecast by their legacy planning vendor! A year earlier, another Vice President Supply Chain took great pain in explaining how diligently his team goes through each time-series. First they ‘clean historical data’ from the impact of price, promotion, outliers, etc. and manually add them back at later stage. I gave a patient hearing. When asked, what if all his pain could have been taken away by a statistical software and provide better accuracy - it was difficult for them to believe!

Indeed, generations of demand planners have followed concepts like ‘correct history prior to forecasting’, in order to create a baseline forecast. To me this approach defeats the purpose of non-judgmental forecasting! This fallacy can be traced back to a class of transactional systems which also claimed to have ‘forecasting’ capability. Statistical best practices don't recommend manual judgement to estimate ‘promotional uplifts’. In fact, it let the system automatically pick up the impact through advanced statistical algorithms – e.g.: proven theories like ARIMAX (Auto Regressive Moving Average )– where ‘X’ stands for exogenous variables in a time series analysis of. Solutions like SAS® Demand-Driven Planning and Optimization (DDPO), based on true statistical principles, uses algorithms, like ARIMAX and UCM (Unobserved Component Modeling) that require information on influencer variables like price, promotion, weather, etc., in as-is form and estimates their impact on the target variable, say sales. In further strengthening it’s ‘data speaks’ philosophy, forecasting software can in-corporate ad-hoc events at any level of granularity and automatically assign it back to the historical outliers. For example, if your brand sponsored the Super Bowl the last two years, you only need to supply information like dates of the past and future event for system to tell you the historical lift, as well as future impact. My colleague and forecasting thought leader, Charlie Chase, wrote a blog, Stop cleansing your historical shipment data!, where he delves in detail into these different statistical methods.

Hence, if your demand planning vendor asks if you wish to determine the impact of causal variable: pause a moment! Ask yourself: do you really wish to judge that? Think instead, why do we want that kind of demand planning system? Let the ‘data speak’! Go one step ahead - rather than centralizing, let your field executives simulate field level localized information! Enable them with web based simulation to estimate impact of promotions, competition promo for an informed target setting discussion. Let this important info from the market place seamlessly flow back into consensus planning. Bye-bye subtraction, bye-bye addition, welcome forecasting!


About Author

Nilmadhab Mandal

Principal Systems Engineer

Nilmadhab is presently working with SAS and has been providing techno-functional consulting for SAS Supply Chain Analytics. He is an MBA and completed Advanced Program in Supply Chain Management. He has been invited to present at several seminars like International Business Forecasting Seminar (IBF) Amsterdam, Institute for Supply Management India and International Conference on Advanced Data Analysis, Business Analytics and Intelligence, IIM Ahmedabad (IIMA).

Comments are closed.

Back to Top