How Retailers Use TSFMs for Demand Planning

Retail demand forecasting is one of the hardest prediction problems in industry. A mid-size retailer might carry 50,000 SKUs across 500 stores, producing 25 million individual time series that need weekly forecasts. Most of these series are short, noisy, intermittent, and influenced by promotions, weather, holidays, and competitor actions. Traditional forecasting approaches buckle under this complexity. Time series foundation models offer a fundamentally different path.

Why Retail Demand Is Uniquely Difficult

Several characteristics make retail forecasting resistant to conventional methods.

Intermittent demand dominates the long tail. The majority of SKUs at any retailer are slow movers. A specialty condiment might sell zero units in a given store-week 60% of the time, then spike to five units during a holiday. Standard models like ARIMA or ETS assume continuous demand patterns and produce poor forecasts for these zero-inflated series. Specialized intermittent methods like Croston's exist but require manual configuration and still struggle with promotional spikes.

Promotional effects create sharp, temporary demand shifts. A buy-one-get-one promotion on laundry detergent can triple baseline demand for two weeks, then cause a post-promotion dip as customers draw down pantry stock. Capturing this requires models that understand promotional calendars, cannibalization across products, and pantry-loading dynamics.

New product launches present the cold-start problem in its purest form. When a retailer introduces a new private-label item, there is literally zero sales history. Traditional ML models cannot generate a forecast without historical data. Planners fall back on manual analoging, picking a similar existing product and using its history as a proxy, a process that is labor-intensive and error-prone.

Scale compounds everything. Training a separate ML model per SKU-store combination is computationally prohibitive and produces millions of models that are difficult to monitor and maintain.

How TSFMs Change the Equation

Time series foundation models address these challenges through their pretrained representations and zero-shot capabilities.

Zero-shot forecasting handles the long tail and new products. A pretrained TSFM can generate reasonable forecasts for a SKU it has never seen, using only a few weeks of sales history passed as context. For new product launches, even a single week of post-launch data can produce actionable demand estimates. This eliminates the need for manual analoging and dramatically reduces planner workload. For more on zero-shot versus fine-tuned approaches, see our comparison guide.

Prediction intervals enable inventory optimization. TSFMs that produce probabilistic forecasts (like Chronos or Moirai) output full predictive distributions, not just point estimates. The 95th percentile of the forecast distribution directly maps to safety stock levels. A retailer can set service level targets (e.g., 98% fill rate) and derive reorder points mathematically from the forecast quantiles, replacing the rules-of-thumb that many supply chain teams still rely on.

Transfer learning captures cross-product patterns. Because TSFMs are trained on diverse time series corpora, they bring knowledge of general demand patterns: weekly seasonality, holiday spikes, trend dynamics. This shared knowledge improves forecasts for sparse series where per-SKU models would overfit.

Specific Use Cases

Weekly replenishment forecasting is the core application. Retailers generate SKU-store-week demand forecasts that feed directly into automated replenishment systems. TSFMs produce these forecasts at scale with minimal per-SKU configuration. A typical workflow: aggregate point-of-sale data to the SKU-store-week level, pass 52 weeks of history as context, and request a 4-to-12-week forecast horizon.

Promotional uplift estimation uses TSFMs to establish baseline demand, then measures the delta between promoted and baseline periods. By forecasting what demand would have been without the promotion, retailers can isolate true promotional lift from underlying trends and seasonality. This informs promotional calendars, vendor funding negotiations, and markdown decisions.

Markdown pricing optimization relies on demand elasticity curves. TSFMs forecast demand at multiple candidate price points, allowing markdown optimization engines to select the price path that maximizes revenue recovery on end-of-season inventory. The probabilistic nature of TSFM forecasts is particularly valuable here, as markdown decisions need to account for downside risk.

Assortment planning for new stores uses forecasts from analogous existing stores as a starting point. A TSFM can ingest sales history from demographically similar stores and generate demand projections for a new location, informing initial inventory allocation and planogram decisions before any local sales data exists.

Practical Implementation

Data preparation is the largest effort in any retail forecasting deployment. POS data must be aggregated and cleaned: returns subtracted, stockout periods flagged (demand during out-of-stock periods is censored and must be corrected), and promotional calendars aligned. Choosing the right forecast granularity matters. SKU-store-week is standard for replenishment, but product-group-region-month may be more appropriate for strategic planning. Hierarchical reconciliation ensures that forecasts at different aggregation levels are consistent, so that the sum of store-level forecasts matches the regional forecast.

Accuracy benchmarks from retail datasets (the M5 competition on Walmart data is a common reference) show that TSFM-based approaches achieve weighted root mean squared scaled error (WRMSSE) within 5-10% of competition-winning ensemble methods, while requiring a fraction of the engineering effort to deploy and maintain. The energy demand forecasting case study shows similar results in a different vertical.

Connecting to Planning Systems

TSFM.ai's forecast API integrates into existing retail planning infrastructure through standard REST endpoints. Replenishment systems call the API with historical sales context and receive forecast distributions that map directly to reorder point calculations. The ability to batch thousands of SKUs in a single API call makes the integration practical even for large catalogs. For retailers already using planning platforms like Blue Yonder, Oracle, or Relex, TSFM.ai forecasts can serve as an input signal alongside existing statistical baselines, providing a powerful ensemble that combines domain-specific heuristics with learned temporal representations. Explore available models on our model catalog, try them interactively in the playground, or learn more about building production forecast pipelines.