Many consumer product companies are showing keen interest in demand sensing. A few weeks ago, we published a blog describing why most CPG companies have hit a demand forecasting ceiling. This week Manoranjith Pathekkara, managing director of ToolsGroup's partner in India, LogicaMatrix, digs into how companies are employing demand sensing to break through that ceiling.
Demand sensing is a capability and technology for improving near-future forecasts using detailed short-term demand data. Near-future means hours or days, depending upon how dynamic your supply chain. Demand sensing reduces forecast error by up to 50%, increases inventory accuracy by up to 20%, and optimally deploys downstream (e.g., Distribution Centre) inventory.
In a demand sensing environment, downstream data such as customer, POS or channel data is employed to identify demand trends, provide advanced warning of problems, and remove the latency between plan and what is really happening in the supply chain. The quicker deviations can be identified, the quicker and more intelligently a company can respond.
Demand sensing can also use a much broader range of demand signals (including current data from the supply chain) and different mathematics to create a more accurate forecast that responds to real-world events such as market shifts, weather changes, natural disasters, consumer buying behaviour, etc.
Why have demand forecasts failed to improve in recent years despite advances in technology? One answer is that “aggregate level planning” using time-series methods often doesn’t achieve a Forecast Accuracy above 60% at item-location levels, even for fast moving items. Aggregation with slicing and dicing rules formulated on historical patterns completely black out the latest trends and patterns at an item-ship-to-daily level. It allows latency to creep-in into plans and masks the true demand signal. This is exactly the problem demand sensing overcomes.
What capabilities should a supply chain solution have in order to start sensing demand?
- The ability to model demand at the most atomic level, such as Item/Ship-to location/Daily - Ship-to locations can be key accounts, sell-in channels, geographical territory etc. This capability is very important because the existing demand models are iterated using the latest demand data at this atomic level to identify the statistical relevance of short-term spikes, outliers, trends and patterns.
- The ability to model demand variability - A demand confidence interval is needed to understand the latest data feeds and segregate noise from the demand signal. This is mandatory because noise has no statistical relevance, hence must be discarded, otherwise you will end up with a “nervous’ supply chain. And just using a normalized variability (say plus or minus 10%) isn’t enough. It doesn’t understand true deviations well enough, causing numerous false positives and false negatives.
- The ability to use downstream data - This could be ship-to data, VMI feeds, POS data, collaborative planning, etc.
- In advanced demand sensing, the impact of external variables like weather forecast, economic conditions, oil price or similar causal factors can be modelled into demand forecasting to predict short-term demand. For example prolonged sun may cause beer sales to go up. Prolonged rains may cause washing machine sales may go up. A manufacturer has to position sufficient inventory in locations where it’s likely to rain substantially to avoid lost sales to a competitor’s brand available at that place/time. Short-term weather feeds into forecasting system make quick repositioning of inventory possible.
- Supply chain planning platforms must scale to process high volumes of data associated with hundreds of thousands of item - location combinations every hour/day.
- To gain potential network benefits, the platform must support a seamless environment between planning and execution, as well as the ability to replenish the high frequency demand signals with optimized execution.
- Increased process automation is required to ensure that the resulting demand signal used to drive the execution environment does not require significant amounts of manual effort.
Do you have the data you need to start sensing demand? Most likely. Ship-to information on distribution, replenishment or sales order are key data feeds along with respective line-order details. Most manufacturing and distribution companies don’t have retail POS data, but sensing demand using sell-in data has been found to improve forecast accuracy by 30%.
What steps does Demand Sensing entail?
- Import short-term demand data on an hourly/daily rather than a weekly or monthly basis
- Immediately sense demand signal changes as compared to a detailed statistical demand pattern
- Evaluate the statistical significance of the change
- Analyse partial period actual demand and execute short-term forecast adjustments using automated routines
- Identify and rapidly react to replenishment issues or sudden changes in customer demand via advanced statistical analytics
Let me add a word of caution. There is no doubt that Demand Sensing with the right technology and processes in place will improve short term forecast accuracy. But forecast accuracy plays a big role in service levels only for fast moving items. For slow moving, highly volatile items, safety stock modelling becomes more important than forecast accuracy in meeting service level targets.
With the ever increasing ”long tail”, one needs to be clear what we are trying to improve; forecast accuracy or service levels (order fulfilment performance). For the latter, look for an integrated solution that can both model/sense demand and also model safety stock on the same platform based on service level oriented policies (stock to service). Just like Demand Sensing, modelling of safety stock has to be a dynamic ongoing process rather than a static quarterly process, because your long tail products are as important or becoming more important than your fast moving items.
Click below for a short whitepaper on moving to a either sell-in or sell-out demand sensing.