Can you really understand future demand, when it is nothing like the past?

Despite heightened awareness for the need to have an accurate demand signal to control inventory and working capital, CPG companies continue to struggle with the business performance on generating an accurate demand signal, both from the retailer / consumer  standpoint and their own need to replenish downstream customer channels like distributors / wholesales. Integrating consumer demand signals into shipment forecasts are becoming an essential survival best practice for CPG companies. So, what’s wrong with prevalent demand forecasting/planning approaches, demand forecasting processes, and technology solutions. Companies have struggled with data synchronization issues, lack of sufficient seasoned in-house talent for prescriptive analytics. Multi-enterprise demand sensing is essentially applying advanced analytics to the demand-supply matching needed 2-3 tiers upstream, where demand sensing may be producing an accurate signal for retail store shelf pick-up from the consumer. In essence, Multi-enterprise demand sensing blends sales/marketing strategies with supply chain replenishment and inventory management strategies to minimize the bullwhip effect and translate retail POS based demand signal to an accurate manufacturer shipment signal.

There has been a long entrenched and mistaken assumption that human overrides and judgments from sales/marketing/finance/executives on top of a statistically generated baseline forecast are the best way to go about the business of generating an unconstrained consensus demand signal. For a lot of companies, it’s still the 90s as they execute statistical forecasting in SAP APO or excel based on time series data and Holt/Winters trend/seasonality exponential smoothing type basic algorithms to essentially forecast their shipments in the next tier, not true consumer demand for their products.

Let’s examine a set of issues with prevalent demand forecasting/planning practices:

·         First, the assumption that past trends/seasonality cycles will repeat themselves forever in future is a wobbly concept in the best of times. In today’s demand pull driven networks and omnichannel consumer world with exploding choices and fickle loyalties, forecasting inherently unforecastable demand in terms of shipments from manufacturer to distributor (2-3 tiers upstream from retailer store) specially with solutions and forecasting concepts from last century is fool’s gold and a sheer waste of effort. In reality, future demand at the retailer SKU/store level is A. relatively less volatile and B. highly sensitive to demand shaping initiatives and variables like pricing, promotions, placement in store like end rack displays, overall consumer confidence, weather, social chatter regarding the product, and so on. The approach of forecasting based on extrapolating the past assumes any error to be noise instead of understanding the causal factors driving the high noise be it pricing/promotions you ran or your competitors did.

·         Second, forecast is often executed as middle-out where data aggregation along the demand hierarchy produces a stable ‘forecastable’ demand signal. This search of a forecastable level of aggregation is dangerous as eventually the forecast at the aggregated level needs to be disaggregated to the base product/distribution center/week level which is where inventory replenishment and shipment processes reside. As valuable granular trends are lost in the process of aggregation, the forecast disaggregation process is an example of math craziness, where robust sounding complex disaggregation approaches are deployed to ‘minimize’ the signal distortion, but it’s evident to one and all that the forecast at granular level (SKU/DC) is anything but accurate. Metrics gaming occurs fairly quickly by Demand Planning team measuring their forecast accuracy at aggregated levels although they know fully well there is no business value of the aggregated signal to the supply chain, no matter how accurate it may be.

·         Third, human overrides on top of a statistical baseline bring in personal/departmental bias and further hurt the end goal of producing an accurate demand signal that maximizes revenue and profitability and minimize the cost and working capital needed to get there. From Finance/Executives baking annual targets in the forecast to Sales/Marketing providing the rosiest demand scenario as their forecast to guarantee finished goods inventory oblivious to the working capital demands and inventory waste risk, it’s actually quite easy to prove the value dilution when organizations embrace Forecast Value Added FVA reporting and diligently capture and compare every human touch point against the naïve or system forecast to measure value added.

·         Fourth, most companies have inside-out demand forecasting processes where they are essentially forecasting their shipments to the next downstream tier as opposed to understanding demand downstream at point of consumption. Downstream ordering frequencies, order sizes, quarter end incentives, demand signal latency, etc. all combine to distort the relatively stable demand signal at retailer store to a highly volatile demand signal 3 tiers upstream. In supply constrained scenarios, shipments do not capture true unconstrained demand and neither do next tier customer orders, which may be inflated in an allocation scenario. In the absence of a multi-enterprise demand sensing signal that can link demand shaped signal at retailer store to corresponding distributor buys from the manufacturer with proper lag, demand sensing quickly degenerates to applying data science in the wrong place and for a short time frame which is within replenishment lead time and not terribly useful for the supply chain looking for a mid-term accurate demand signal.

·         Fifth, given the innovation centric SKU portfolio management approaches prevalent since last decade, there is tremendous SKU explosion and a longer than long tail for most organizations. This long tail has intermittent demand, is extremely hard to forecast, but given high service levels and resulting high inventory buffers, are a sure black hole for profitability. SKU proliferation also means your demand planners and data analysts are overwhelmed and cannot pay needed attention to generating robust predictive models.  

·         Sixth, even when companies have understood the need to better sense and shape the end point demand signal, the business teams have struggled with the complexity levels associated with advanced forecasting methods: multiple linear regression, ARIMA models, ARIMAX models, and so on. Throwing big data projects at the demand sensing problem inside-out without having an analytics innovation scalability & sustainability strategy is a costly misadventure.

There is a way out. Executive sponsors need to ensure seasoned data scientist talent is available and closely collaborating with business users / domain experts on a cognitive, visualization, and simulation platform that can ensure demand model accuracy on an ongoing basis and scale high signal quality across the full product portfolio. Given the non-linear relationships in highly volatile demand environments and the SKU explosion long tail challenge, the prescriptive demand model building and maintenance is simply outside the human capability of even large teams in the absence of a self-healing cognitive platform that can both accelerate initial model definition as well as ongoing maintenance.

Multi-enterprise demand sensing platform from GitaCloud is designed to provide business users a demand shaping what-if simulation platform or optimization abilities to forecast most effective promotion mix as well as link and translate the consumer demand to the replenishment demand. This helps to model the push/pull effects of the supply chain. The platform auto develops a consumer demand sensing model by integrating retailer POS sell-through data with all its explanatory variables: past consumption history, retail merchandising vehicles, pricing, advertising rating points, and promotions. A secondary model is then developed which can link the forecasted consumer demand to replenishment demand based on explanatory variables like demand latency between POS consumption and replenishment shipments into the channel, trade promotions, forward buying, order frequency, order sizes, wholesale price, etc.

GitaCloud orchestrates multi-enterprise end to end insights and accurate demand signals as a service with our:

·         cognitive learning platform that automatically generates demand models from the data and connects them across the value network

·         turnkey data / integration services to source / clean / blend data from downstream customer channels and other structured/unstructured sources outside the enterprise

·         seasoned demand model builders, and industry domain experts that can partner with your business people to build / maintain prescriptive demand models as a service with a low payback period.