Rebuilding a planning function around the physical world

Every supply chain leader has lived this moment. A cold front pushes through a region, marginal batteries fail overnight, and the demand signal surfaces in the sales data a week later—by which point the shortage is three regions deep and the network is paying for emergency freight. The forecast that drove the plan was not wrong, exactly—it was looking in the wrong direction, reading history while demand was being written by the weather.

Vishal Singh

For most of a decade, I have led replacement-demand planning for low-voltage automotive batteries, where this failure mode is the default. The lesson generalizes well beyond batteries: for a large class of products, the most important thing a planning organization can do is stop forecasting its own sales history and start forecasting the physical world driving demand. That shift is not a model upgrade. It changes the operating posture of the entire function, and it is as much a leadership challenge as a technical one.

Why history is the wrong teacher

Conventional forecasting—exponential smoothing, ARIMA, and most of their machine-learning successors—shares one quiet assumption: that the best predictor of future demand is past demand. For stable consumer goods, that holds well enough. For products whose demand is triggered by external conditions, it fails in a specific and costly way. A battery does not fail because it was sold; it fails because heat degraded it over the summer and a cold snap exposed that degradation on the first hard-starting morning. By the time the purchase appears in the sales record, the event that caused it is already over. A history-based model is, by construction, always reporting yesterday’s news—and the categories that share this property are common: HVAC equipment, agricultural inputs, service parts, anything tied to an aging installed base.

How it works

The methodology I developed, Multi-Signal Causal Demand Sensing (MSCDS), inverts the usual approach. Rather than extrapolating a demand curve, it forecasts the causal drivers of demand and learns how they translate into replacement volume at a regional level. Five families of external signal feed the model: vehicle-in-operation data (the age, model, and climate exposure of the registered fleet, at ZIP-code granularity); weather; point-of-sale activity; retailer inventory; and pricing. In compact form, the regional forecast is a function of those signals:

where predicted demand for a region and period is driven by vehicle-in-operation (VIO), point-of-sale (POS), price and promotion changes, retailer inventory (I), weather (W), and order momentum (O). A linear backbone carries the first-order elasticities so the model stays transparent, while a non-linear term—a gradient-boosted ensemble complemented by a recurrent (LSTM-style) network—captures the interactions that matter most, such as an aging fleet and an incoming cold snap combining to produce a spike neither would generate alone. Each forecast is then decomposed into additive driver contributions, so a planner can read how much of a predicted surge comes from fleet age versus weather versus price. That decomposition is what turns a prediction into a decision a team will act on.

Deployed across a network supplying more than half of the U.S. aftermarket, MSCDS has delivered a high-single-digit improvement in forecast accuracy and, more consequentially, a substantial reduction in emergency freight and in working capital tied up in safety stock—because a sharper forecast hits the same service level with less buffer.

The harder change behind the model

The accuracy gain is not the part worth writing about. What had to change around the model is. Three shifts defined the transformation, and each was organizational before it was technical.

The unit of planning moved from the SKU to the installed base. Planners stopped asking “how many of this part did we sell last year” and started asking “how many vehicles of a given age, in a given climate, carry a component near the end of its life.” That reframing changes what the function measures and treats as a leading indicator.

The trigger moved from the sales report to the weather front. Replenishment began firing ahead of demand rather than in response to it. Inventory is positioned before failures occur, so the planning calendar is paced by what is about to happen, not what already did.

The decision moved from contested to shared. This is the change that made the rest stick. Because every forecast is decomposed into its drivers, planning, supply, and finance stopped arguing about whether to trust a black box and started discussing the physical world: Is a cold front actually coming? Is the fleet in that region actually old? Those are questions an operations team can answer and stand behind. Alignment that once consumed meetings now takes minutes.

Interpretability is a leadership requirement

Much of the conversation about AI in planning treats accuracy as the finish line. In a high-stakes environment, it is barely the starting line. A model that tells a team to move a meaningful amount of inventory will not be acted on unless the people accountable can interrogate it and put their name to the call. A forecast that cannot be questioned will not be trusted, and one that is not trusted changes nothing the company does. Accuracy earns a model a hearing; interpretability earns it a decision. Leaders evaluating AI in their own functions should weigh the second at least as heavily as the first, because the return on a forecast is realized only when a human chooses to act on it.

What transfers

Batteries are an extreme case, not a special one. For any leader whose demand is driven more by external conditions than by their own sales history, three lessons carry over. Find the true driver of demand and forecast that, then translate it into demand—rather than forecasting demand directly and hoping the driver is buried in the history. Treat interpretability as an operational requirement, not a technical luxury. And recognize that the scarce skill in modern planning is no longer building the model; it is turning its output into a decision that operations will execute and finance will fund.

There is a broader stake, too. A distribution network that absorbs weather-driven shocks in advance, instead of amplifying them afterward, is a more resilient piece of infrastructure—fewer stranded motorists, fewer grounded fleets, fewer critical vehicles that fail to start on the coldest morning of the year. The financial results show up on a balance sheet; the resilience does not, and it may be the more durable contribution. When the right part is on the right shelf on the worst morning of the year and a vehicle simply starts, no one notices the forecast that put it there. That quiet non-event is the entire point.

About the author

Vishal Singh leads demand planning for the U.S. and Canada at Clarios, the world’s largest manufacturer of low-voltage automotive batteries, and is the originator of the Multi-Signal Causal Demand Sensing (MSCDS) methodology described here. Over more than 15 years in supply chain and demand planning, including roles at East Penn Manufacturing and Apple, his work has spanned automotive, consumer electronics, and retail. He holds an MBA from Kellogg and a B.E. in Computer Science.

SC
MR