Every demand forecast carries a hidden assumption: that supplier lead times are roughly what they were last quarter. That assumption is baked into the model architecture, the reorder point calculation, and the safety-stock buffer. When lead times hold, the assumption goes unnoticed. When they shift, the forecast starts generating errors that look like demand volatility but are actually supply-side timing failures.
We see this pattern repeatedly when planning teams bring us their historical forecast error data. MAPE climbs during the same windows that lead times expanded. The planners attributed the error to demand unpredictability. In many cases, demand was flat. What shifted was how long it took stock to arrive.
What "Lead-Time Variance" Actually Means in Practice
Lead time, in most planning systems, is stored as a single number: a field in the ERP that says "14 days" or "21 days" per supplier-SKU combination. That number comes from somewhere, often a procurement negotiation or an onboarding entry that was accurate once and has not been touched since.
The reality in most manufacturing and distribution supply chains is that lead time is a distribution, not a point estimate. A supplier with a 14-day stated lead time might deliver in 10 days 30% of the time, 14-16 days 50% of the time, and 22+ days 20% of the time. When conditions tighten (port congestion, a supplier production delay, a carrier capacity crunch), that 20% tail expands.
Variance matters for forecasting because the reorder point formula depends on lead time directly. If your reorder point is set for a 14-day lead time and you're actually operating in a 20-22 day window, you're triggering replenishment orders late relative to when you'll need stock. The forecast says "you'll be fine for 3 more weeks." Actual stock arrives 8 days after you expected it. You run short.
Where Standard Forecasting Models Break Down
Most demand forecasting methods, from simple exponential smoothing to more structured statistical models, treat demand history as the primary input. They're designed to answer: given what demand looked like over the past 6-18 months, what will it look like over the next 90 days?
That question is useful but incomplete. The downstream procurement decision, the reorder trigger, depends not just on forecasted demand units but on when those units need to be ordered given lead time. If the model treats lead time as a constant when it is actually variable, the recommendation outputs a reorder point that is systematically miscalibrated during disruption periods.
Consider a scenario: a mid-size distributor in the Southeast US sources a component category from two Tier-1 suppliers, both with manufacturing capacity in Southeast Asia. Stated lead time is 18 days. In Q1 2025, port congestion at one of the West Coast hubs pushes effective lead times to 26-30 days for three weeks. The ERP-stored lead time does not update. The forecast model fires replenishment alerts based on an 18-day assumption. Orders are placed on schedule. Stock arrives 10 days late. The distributor runs low on 14 SKUs in a category where customers cancel orders after 5 days of stockout.
That is not a demand forecasting failure. The demand model was accurate. It is a lead-time-variance failure, and it is invisible to any system that treats lead time as a static input.
The Compounding Effect: Variance Inflates Apparent Demand Noise
Here is where things get worse. When a stockout occurs from a late shipment, the planning team often sees a spike in demand immediately after stock arrives. Customers who were waiting place catch-up orders. The model sees that spike and interprets it as a demand pattern: elevated demand in week N+1 following a quiet week N. On the next cycle, the model over-buys for that position in the planning horizon. Safety stock goes up. Working capital increases. The root cause, a 10-day lead-time extension from a port disruption that has since resolved, is long gone. The model is still responding to it.
This is why forecast error tends to persist well after a disruption ends. The original signal contaminates the demand history, and standard models incorporate that contaminated history into future forecasts. We call this lead-time echo: the ghost of a supply disruption showing up as demand noise in cycles 3-6 months later.
What a Lead-Time-Aware Forecast Looks Like
The fix is not complicated in concept, though it requires data infrastructure that most planning teams do not have by default. A lead-time-aware forecast treats lead time as an estimated distribution rather than a constant, and it updates that distribution continuously based on observed actual delivery performance and external signals.
The practical components are:
- Actual lead-time tracking per supplier and lane: not just stated lead time, but the delivery date versus PO placement date for every order. This data lives in the ERP but is rarely extracted and analyzed as a time-series.
- External signal feeds that predict lead-time shifts: port congestion data, carrier capacity indicators, weather event proximity to supplier regions. These lead the actual delivery delays by 1-3 weeks, which is enough time to adjust reorder timing before stock runs short.
- Dynamic reorder point recalculation: when the lead-time distribution shifts (say, mean moves from 18 to 24 days), the reorder point for affected SKUs updates automatically, not at the next quarterly planning review.
We built Supplyverde's forecast engine around this architecture. The demand model outputs a 90-day SKU-level forecast. Separately, the disruption risk layer reads external signals and current actual delivery performance to produce a lead-time distribution estimate per supplier. The reorder point for each SKU is calculated using the current distribution, not the static ERP field.
A Note on What This Does Not Solve
We are not saying that demand forecasting methods are wrong or that improving them is not valuable. Forecast accuracy matters. A better demand model reduces the buffer you need to carry. What we are saying is that demand-side accuracy improvements have a ceiling when the supply-side timing assumption is frozen. You can get MAPE from 18% down to 12% with a better demand model and still have the same stockout rate if lead-time variance is large and untracked.
Similarly, this approach does not eliminate the need for safety stock. It makes safety stock more precisely sized. A buffer calculated against a known lead-time distribution with current signal inputs is smaller than a buffer calculated against "whatever went wrong last time plus a margin." Smaller, better-targeted buffers free up working capital without increasing stockout risk.
Getting Practical: What to Measure First
If you want to understand whether lead-time variance is a meaningful driver of your forecast error, start here. Pull your last 12 months of purchase orders with PO date and actual receipt date. Calculate actual lead time per order. Compare to the stated lead time in your ERP. Look for periods where the gap between actual and stated is larger than 5 days. Then cross-reference those periods with your MAPE or stockout history.
In our experience, teams that run this analysis for the first time find that 30-40% of their forecast error events trace back to lead-time windows where actual outpaced stated by more than a week. That is a large fraction of error to be addressing on the demand side with no effect.
The data you need for this analysis is almost certainly already in your ERP. The question is whether your forecasting process uses it, or whether it stops at demand history and trusts that procurement will handle the rest.
Most planning teams are implicitly dividing the problem at the wrong seam. Demand forecasting and supplier lead-time management are treated as separate workstreams. When supply conditions are stable, that separation works fine. When they are not, the mismatch shows up as forecast error that no amount of demand-side tuning will fix.