The monthly Sales and Operations Planning review is a social process disguised as a mathematical one. A data science team generates a number, but a demand planner has to defend that number to a warehouse manager scheduling shifts or a fleet director buying trucks. If a new time series foundation model predicts a sharp drop in refrigerated freight volume, the commercial director will ask what is driving the drop. If the forecasting lead cannot point to a lost contract, a price change, or a macroeconomic indicator, and can only explain that the model recognized a decay pattern in the telemetry, the director will refuse to cut the labor budget. A forecast you cannot explain is a forecast you cannot ship.

For a decade, the standard approach to supply chain demand planning was to build a pipeline that trained a separate statistical model for every product in every location. A mid-sized logistics network might quietly execute hundreds of thousands of individual ARIMA or exponential smoothing models every weekend. Time series foundation models replace that sprawling infrastructure with a single, massive neural network. You no longer train a model on a route; you query a pre-trained model with the route’s recent activity.

This architecture shift solves the hardest part of forecasting: the cold start. Most logistics catalogs follow a power law. Ten percent of the routes, SKUs, or parts make up ninety percent of the volume. Classical statistical models and gradient-boosted trees handle this high-volume head perfectly because they have years of clean, continuous history and well-understood causal drivers. Foundation models belong on the long tail.

When a cross-border logistics provider opens a new regional distribution center, there is no localized historical data to train a traditional model. In the past, the data science team would map the new facility to an existing like-item facility to generate a baseline. With a foundation model operating in zero-shot mode, the team feeds the first three weeks of actual daily volume into the network, which outputs a 90-day forecast based entirely on the shape of that short sequence. The model does not know it is forecasting pallet volume for a new market. It just recognizes the geometry of a ramping trend.

Zero-shot means zero history on your series. It does not mean zero assumptions. Natural language foundation models treat text as sequences of word tokens. Time series foundation models treat numerical history as sequences of value tokens. They scale, normalize, and slice historical data points into discrete chunks. The model learns the structural grammar of how sequences rise, fall, plateau, and exhibit seasonality across billions of unrelated time series. But the pretraining corpus is somebody else’s catalog. Yours is not in it.

Foundation models solve the cold start while failing at the easiest part of logistics forecasting: knowing that volume goes up when prices go down. Classical machine learning models rely heavily on covariates. If you want to predict outbound shipments, you feed the model historical volume along with fuel prices, local weather, and holiday calendars. Current time series foundation models struggle to integrate these external, multivariate signals natively. They are designed primarily for univariate forecasting.

Promotions, price changes, distribution gains, and competitor actions are interventions, not patterns. A pretrained forecaster has no concept of a three-for-two mechanic unless you hand it a flag, and even then, it can only learn the average response from its pretraining data. Promotion lift is a causal question wearing a forecasting costume. Keep the causal driver model where it is and use the foundation model for the baseline.

Before running the model, you have to fix the data. Shipment history is not always demand history. A warehouse that shipped nothing during a stockout did not necessarily face zero demand. The same problem appears when carrier capacity limits bookings. If you feed those zeros into a foundation model as part of the context window, it will interpret them as a true drop in demand. The model will learn the constraint as if it were demand. This is true of classical models too, but foundation models can be more confident about it because they have seen millions of similar-looking dips. Lost sales are not demand. Unconstrain the data before inference.

The model is not the hard part. The reconciliation is. Bottom-up SKU-by-location forecasts from a foundation model do not automatically sum to the category plan finance approved. A forecast can be accurate against total daily volume at the distribution center while getting the mix wrong at the loading bays, assigning too much volume to ambient goods and too little to chilled goods. The aggregate score hides the staffing and space decision. Hierarchical reconciliation is a design decision you make before you pick a model, not after.

Do not rip and replace. Run the foundation model in parallel with the incumbent and score both on rolling-origin backtests over your own series. Stratify the backtest by velocity band, promo status, and intermittency. Report forecast value added over the incumbent, not raw accuracy in a vacuum.

If the challenger wins on the head and loses on the tail, you have a segmentation decision, not a model decision. Keep classical models for the high-volume head with clean history. Use the foundation model for the long tail, the intermittent spare parts, and the newly acquired catalogs with fragmented data. When a planner overrides a forecast because a customer has given notice of a temporary closure, the next evaluation should retain both the model forecast and the override. Otherwise, nobody can tell whether the model improved, the planner improved it, or both were rescued by an event recorded nowhere else. The fallback is part of the forecasting system, not an admission that the new model failed.

Foundation models predict distributions, giving you P10, P50, and P90 quantiles for free. This changes how safety stock is set and how risk is discussed with finance. A quantile forecast is a conversation about risk. A point forecast is an argument.

The infrastructure is moving from the many to the one, but the demand planner still has to explain the number to the operations director.

Take it with you

Takeaways

  • Foundation models handle cold starts and long-tail series better than classical models
  • Promotions and price changes are causal interventions that foundation models cannot learn from patterns alone
  • Unconstrain the data before inference, or the model will learn constraints as demand
  • Run parallel backtests stratified by velocity band, promo status, and intermittency before replacing anything

Asked and answered

Questions

Where do time series foundation models beat classical forecasting models?

They beat classical models on cold starts and the long tail: new facilities with no history, intermittent spare parts, and newly acquired catalogs with fragmented data. Classical models still win on high-volume series with years of clean history and well-understood causal drivers.

Can foundation models handle promotions and price changes?

Not natively. Promotions, price changes, distribution gains, and competitor actions are interventions, not patterns. A pretrained forecaster has no concept of a promotion mechanic unless you hand it a flag, and even then it can only learn the average response from its pretraining data. Keep the causal driver model where it is and use the foundation model for the baseline.

How should we evaluate a foundation model before deploying it?

Run it in parallel with the incumbent and score both on rolling-origin backtests over your own series. Stratify the backtest by velocity band, promo status, and intermittency. Report forecast value added over the incumbent, not raw accuracy in a vacuum.