Time Series: Five Neural Networks That Forecast Better Than Analysts

While the world argues over whether GPT-5 will learn to reason, a quiet but very expensive revolution has taken place in the shadow of large language models. It concerns time series forecasting, the field that decides how much bread to deliver to a store tomorrow or when the shares of tech giants will fall. For a long time, classical statistics and gradient boosting ruled here, both requiring manual tuning for every task. But times have changed. Foundation models for time series have arrived, and they make the old methods look like an attempt to calculate a rocket's trajectory with a slide rule.

The first serious blow to conservatism was Amazon's Chronos. Its developers approached the problem with irony and elegance: they decided that numbers are just another language. Chronos quantizes the values of a time series and turns them into tokens, after which an ordinary Transformer architecture predicts the future as easily as it autocompletes your emails in Gmail. It sounds like an oversimplification, but in practice the model shows striking accuracy in zero-shot mode. That means the neural network sees the data of a specific warehouse or a currency quote for the first time in its life, yet produces a forecast more accurate than algorithms that were tuned for weeks.

Google did not stay on the sidelines and rolled out TimesFM. Unlike its competitors, the search giant fed its model 100 billion real data points from open sources and search trends. TimesFM uses a decoder-only architecture, which makes it remarkably fast. It handles very long planning horizons where ordinary neural networks start to "hallucinate" and draw straight lines off into infinity. The significance of this is hard to overstate: Google has effectively given the market a tool that scales the expertise of an entire team of data scientists down to a single API call.

Salesforce, for its part, introduced MOIRAI, a universal forecaster that can work with data of any frequency. This was the industry's main pain point: models built for hourly data are usually useless for monthly reports. MOIRAI solves the problem with flexible patches, adapting to the incoming stream on the fly. Alongside it comes Uni2TS, also from Salesforce, which tries to create a single framework for all tasks, from classification to filling gaps in data. These models no longer require a business to own terabytes of its own history for training; they arrive already "smart".

Lag-Llama also deserves a mention. As the name suggests, it is based on Meta's Llama architecture*, but adapted for probabilistic forecasting. This is critical for risk management. It is not enough for us to know that the oil price will be 80 dollars; we need to know the probability that it will fall to 40. Lag-Llama builds probability distributions so easily that it seems to have been made for exactly that. It is a clear example of how work from NLP (natural language processing) suddenly became a foundation for financial mathematics.

Why does this matter right now? We are moving from the era of "small models for every task" to the era of universal systems. Before, forecasting demand for sneakers and predicting load on power grids required different specialists and different approaches. Now one model can do both, and often does it better than narrowly specialized solutions. This radically lowers the barrier to entry for business. Even a small startup can now get Fortune 500-level analytics simply by plugging in the right library from Hugging Face.

The bottom line: the era of manual parameter tuning in statistical models is coming to an end. Foundation models for time series are the industry's new standard. Only one question remains: are you ready to entrust your budget planning to a neural network that was trained on internet texts and weather charts?

*Meta has been designated an extremist organization and is banned in Russia.