How to avoid demand forecasting traps

During supply chain improvement conversations, we’re often asked if we can do anything to improve demand forecasting accuracy. Developing a data-driven approach to this certainly interests us. But our conversations until today typically have resulted in improving supply chain performance not forecasting accuracy. So why does focus shift from forecasting accuracy to performance improvement?

We’ve already identified a few insights which might give you food for thought.

Where the real forecasting headaches occur
Longer term forecasting, using aggregated data for creating strategic direction or annual budgets, usually works out fine. No problems there.

The real headaches start when trying to forecast item-level demand within specific time-frames.
Aggregated data is of no use in this situation. Many methods and tools used for this define parameters based on the assumption of demand variation following a statistical normal probability distribution. But in reality, typically less than 15% of stock keeping units behave accordingly. A better approach for parameter setting is certainly required!

Production schedules and inventory plans are based on the assumption that item-level demand forecasts will prove accurate. Yet deviations from these forecasts are pretty much inevitable. And that presents a challenge to supply chains, because the pursuit of forecasted inventories is out of line with live demand.

Why is time-specific, item-level forecasting so difficult?

• Customers struggle to predict their future in sufficient detail.
• Sales representatives tend to forecast that they will meet their incentive objectives.
• Company focus might be on making the current quarter, rather than ensuring high quality input for forecasting next year’s quarters.

Acknowledging that item-level forecasting is difficult generates the perfect excuse for poor supply chain performance: “if forecasting would be significantly better, performance would significantly benefit”. That sounds appealing, but our data suggests that the hypothesized relation may not be there: there is a three step approach to improve supply chain performance despite forecasting challenges.

1. Fix the basics: standardize processes, parameter settings, procedures, roles and responsibilities and ensure that any hand-off of work between individuals (or functions) is well standardized.
2. Stabilize Output: reduce variation in daily manufacturing output or reduce variability around a target delivery date by learning from exceptions and include these learnings in the standardized processes.
3. Rightsize Inventories: implement Make to Order / Make to Stock policies and optimize inventory levels to meet variation in demand.

With the challenges around forecasting accuracy in mind, the question in this final step is “How do you maintain appropriate inventories when you can’t rely on forecasts?”

Planning without forecasting

Our investigations to-date show that the best predictor for (short term, item level) future demand is (short term) historical use. Base your replenishment process parameters on the most recent 12 to 18 months of data. While forecasting is excluded from these figures, any future facts (e.g. summer or Christmas breaks, new product launches or planned promotional campaigns) should of course be taken into consideration.

Are you still considering to improve forecasting accuracy?

Before moving ahead, how much of your current challenges are the result of poor process standardization? Have basics been fixed? Have you implemented continuous improvement cycles to stabilize processes by learning from (and taking actions on) deviations? Are you managing inventories based on an appropriate Make to Order / Make to Stock policies and are the latter leveraging historical data to set parameters?

Improvements in these areas typically resolve many of the performance issues that often are assumed to be linked solely to forecasting issues.

Jan Martens is Senior Business Process Excellence Consultant for R&G Global Consultants in The Netherlands