Zero-Shot Forecasting
Zero-shot forecasting refers to the ability of an AI model to make predictions for data series it has never seen during training — without any additional fine-tuning. The method is particularly relevant when companies need quick forecasts for new products, markets, or situations for which no sufficient data history yet exists.
Imagine you have to forecast the sales of a product that has only been on the market for three weeks. A classic forecasting model needs years of historical data — and simply fails here. Zero-shot forecasting solves this problem in a different way: a model trained on millions of time series from completely different domains is shown the few available data points and still delivers a usable forecast. In doing so, it transfers patterns it learned from data elsewhere to the new, unknown situation. “Zero shot” literally means zero training examples for the specific case — the model never saw this particular time series during training.
Why zero-shot forecasting makes a difference
Conventional time series forecasting methods — that is, methods for predicting values that develop over time — require many historical data points from exactly the domain they are supposed to predict. A model that forecasts electricity consumption cannot simply be applied to sales figures. This severely limits their use.
Zero-shot models are independent of this. They can be applied immediately to new data domains — without adaptation, without waiting time, without the effort of collecting training data. This is particularly valuable in business: new product lines, unfamiliar markets, or rare events by definition have no long data history. Anyone who needs to act quickly in such cases benefits from a model that still delivers a reliable estimate.
A common misunderstanding: “zero shot” does not mean the model knows nothing at all. It has indeed learned — just not on this particular case. The difference is comparable to an experienced meteorologist who can interpret the weather even in a city he has never been to, because he knows the fundamental patterns of clouds, wind, and seasons.
How a zero-shot model generates forecasts
The basic idea is to train a model on an enormous variety of time series — temperature data, stock prices, website visits, energy consumption, supply chain data. In doing so, the model does not learn specific facts, but structural patterns: What do trends look like? How do seasonal fluctuations behave? What shapes do outliers take? This abstract pattern knowledge can then be transferred to new time series.
Technically, current zero-shot forecasting models often rely on transformer architectures — the same basic structure underlying large language models like ChatGPT. Instead of words, they process time series points as input. The model receives the previous values of a series and outputs predicted future values. Such models are trained on sometimes more than a billion individual time series data points from public and commercial datasets.
A well-known example is the TimeGPT model from the company Nixtla, which performs on unknown datasets without any fine-tuning whatsoever. Google and Meta have presented similar approaches with research projects such as TimesFM and Lag-Llama. These models are often referred to as “foundation models for time series” — in reference to the large language models, which likewise serve as a universal basis for many tasks.
Zero-shot forecasting in products and headlines
In retail and logistics, time series forecasting is one of the most important applications of all. Companies like Amazon or Zalando plan their inventory based on such models. Zero-shot approaches become interesting there when new products are introduced or when data from new regions is missing — classic weak points of traditional systems.
In the finance industry, zero-shot forecasting comes up in discussions about AI-powered forecasting tools. The question is always how reliable such models are in completely unknown market situations — for example, after a stock market crash for which there are hardly any historical parallels.
The method is also being tested in climate research and healthcare — for example, in forecasting hospital occupancy. Here, data for rare events is often missing, exactly the scenario zero-shot forecasting is designed for. The technology is still young, but it is developing quickly: in 2023 and 2024, several such foundation models for time series were made publicly available for the first time.