Forecasting, the Agentic Way
TimeCopilot is open-source agentic forecasting with LLMs and the strongest time series models. We started TimeCopilot with a clear goal. We want to democratize time series forecasting and make it accessible not only to humans but to the next wave of automated agentic systems. Accurate, automated, and easy forecasts.
We are all grounded by temporal data
In 2005, Steve Jobs told the Stanford graduating class that we can only connect the dots looking backwards. At a fundamental level, he named one of the forces that shapes how we live. He explained how we move through the world without knowing what our actions will create, and why trusting those actions matters when the outcome is uncertain. For professionals who work with forecasts every day, this is not an abstract idea. Our work supports planning, risk, demand, supply, energy, finance, operations, and more. And as Steve said, the success of our work depends on how well we use past outcomes in the form of time series data to see what the future might look like through the lens of our own context.
Daily forecasting work is not glamorous. It is the nitty-gritty of managing pipelines, models, baselines, assumptions, overrides, and schedules. It means moving between scripts, dashboards, notebooks, and model families. Many forecasters know that explaining a forecast often takes more time than producing one.
Why go to so much trouble to understand the future? As a field we know that small shifts in starting conditions can create very different outcomes. And we know that uncertainty has structure. When we understand that structure, we make better decisions. There is satisfaction in taking something complex and turning it into a pattern we can recognize. In simple terms, we do it because we enjoy predicting the future. TimeCopilot is our rendition to the field. It reflects both technical progress and an ideal. We want to use all the technology we have to make better predictions.
We believe the future opens up with LLMs
Forecasting has moved through several eras. Classical statistics. Probabilistic approaches. Deep learning. And now time series foundation models trained on large and diverse datasets. Each method brings value, but none solves the full problem. Different inputs and assumptions need different tools. Every forecaster knows: no single model, no matter how powerful, can solve every problem. Forecasting fuel needs for a high tide water platform is not the same as forecasting train schedules for a small province in Germany. Our work as forecasters has always relied on understanding context and resetting the constraints that make specific pipelines work for specific use cases. But what if we had an automated helper that understood our particular needs? With the rise of agentic systems, we finally have an extra set of hands to take on the repetitive work we have to deal with every day.
As we see it, there’s a need to build a layer above all the foundation models and the systems that came before. A layer that knows when to use each method, how to combine them, and how to explain the reasoning behind a forecast.
TimeCopilot is that layer. It uses a wide set of time series foundation models and classical models and libraries. It applies LLMs to add context, constraints, and reasoning. It explains its assumptions in natural language. It updates forecasts as new information arrives. It lets people work through text while keeping the granularity that teams need of hand-tuning pipelines: a coordinated reasoning layer that uses many tools and adapts as context shifts.
With the dawn of large language models, it became evident that written word matters, and plain English may be the most natural interface for reasoning. Andrej Karpathy often notes that language is becoming the new interface for software. Forecasting fits that idea. The work is built on uncertainty, intent, and assumptions. Text gives us a simple way to express that. Text translates our needs and helps us define the guardrails for our context. TimeCopilot is the agent for forecasting in the same way Clay helps marketing teams move faster and Cursor helps engineering teams ship code with less friction.
And so we asked ourselves a simple question: can we use agents for time series forecasting?
TimeCopilot began as a first principles question about forecasting and context. Early prototypes led to a research effort. Our early project attracted collaborators, engineers, and contributors who cared about the same ideas. In a short time, it has become a meaningful open source time series project. We have assembled the largest unified collection of time series foundation models with more than 30 models across more than 7 families. We have shown practical examples of agentic forecasting. We reached the top position on the GIFT Eval benchmark above AWS, Salesforce, IBM, and top universities. We published at NeurIPS. We passed 12k downloads, 280 GitHub stars, and continue to grow a community around this work. Everything is open source and built with a scientific mindset.
We now have early support that helps us move quickly and work with organizations exploring where agentic forecasting can be useful.
We believe the next generation of forecasting will be grounded in data, expressed through text, orchestrated across many models, clear about uncertainty, and built in the open.
TimeCopilot is our next step towards that.
More soon.
Azul & Renée