Legacy ARIMA and exponential smoothing models were designed for stable, linear demand patterns. They degrade sharply during demand shocks, new product launches, and channel mix shifts. Foundation model time series forecasters trained across thousands of product histories generalize to new SKUs immediately, incorporate external signals automatically, and cut forecast error by 30 to 50 percent relative to classical statistical baselines.
The standard demand forecasting stack at a large enterprise is a patchwork of ARIMA models, exponential smoothing variants, and spreadsheet overrides maintained by a planning team that spends 60 percent of its time adjusting forecasts manually. These models have a structural failure mode: they extrapolate from historical patterns and have no mechanism for incorporating signals that have not previously correlated with demand in the training window. When a demand shock arrives, a new competitor enters, or a product launches into a new channel, the model is effectively blind and the planner's override is the only mitigation. This creates systematic over-stock in stable categories and chronic under-stock in high-velocity growth categories.
Amazon's Chronos (2024) and Nixtla's TimeGPT represent a new class of foundation models for time series forecasting. Pretrained on tens of millions of time series from diverse domains, these models generalize to new SKUs with zero historical data, incorporate promotional calendars and macroeconomic covariates natively, and produce calibrated probabilistic forecasts that let planners set service-level-optimized safety stock targets instead of rule-of-thumb buffers. Independent benchmarks on the Monash Forecasting Archive show Chronos achieving 30 to 50 percent MASE improvement over classical baselines across multiple demand categories.
Want to scope this solution for your organization? 15 minutes is enough to tell if this fits.
Schedule a 15-minute intro call →