Временные ряды: пять нейросетей, которые видят будущее лучше аналитиков

Time Series: Five Neural Networks That See the Future Better Than Analysts

While the world is passionately debating whether GPT-5 will learn to reason, in the shadow of large language models there has been a quiet but extraordinarily expensive revolution. We're talking about time series forecasting — the very field that determines how much bread needs to be delivered to a store tomorrow or when the stocks of tech giants will fall. For a long time, classical statistics and gradient boosting reigned here, requiring manual tuning for each task. But times have changed. Foundation models for time series have entered the scene, and they're making old methods look like an attempt to calculate a rocket's trajectory using a slide rule.

The first serious blow to conservatism came from Amazon's Chronos. The developers approached the problem with irony and elegance: they decided that numbers are just another language. Chronos quantizes time series values and transforms them into tokens, after which a standard Transformer architecture begins to predict the future as easily as it completes your emails in Gmail. This sounds like oversimplification, but in practice the model demonstrates stunning accuracy in zero-shot mode. This means the neural network sees the data of a specific warehouse or currency quotes for the first time in its life, but produces a forecast more accurate than algorithms that were tuned for weeks.

Google didn't stay on the sidelines and rolled out TimesFM. Unlike 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 incredibly fast. It handles enormous planning horizons where regular neural networks start to "hallucinate" and draw straight lines into infinity. The importance of this moment is hard to overstate: Google has essentially given the market a tool that scales the expertise of an entire department of data scientists down to a single API call.

Salesforce presented MOIRAI on its side — a universal forecaster that can work with data of any frequency. This was the industry's main pain point: models for hourly data are usually useless for monthly reports. MOIRAI solves this problem through flexible patches, adapting to the input stream on the fly. It's joined by Uni2TS, also from Salesforce, which attempts to create a unified framework for all tasks — from classification to filling gaps in data. These models no longer require businesses to have terabytes of their own history for training; they come already "smart."

We can't forget about Lag-Llama either. As the name suggests, it's based on Meta's Llama architecture but adapted for probabilistic forecasting. This is critical for risk management. It's not enough for us to know that oil prices will be $80 — we need to know the probability that they'll drop to $40. Lag-Llama builds probability distributions with such ease, as if it was always created for this purpose. This is a clear example of how achievements from NLP (natural language processing) suddenly became the foundation for financial mathematics.

Why does this matter right now? We're transitioning from an era of "small models for each task" to an era of universal systems. Previously, different specialists and different approaches were needed to forecast sneaker demand and predict electrical grid load. Now one model can do both, and often better than narrowly specialized solutions. This dramatically lowers the barrier to entry for business. Now even a small startup can get Fortune 500-level analytics by simply connecting 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 new industry standard. Only one question remains: are you ready to trust your budget planning to a neural network that was trained on internet text and weather charts?