TimeCopilot

Can a forecast that costs an API call beat, Wall Street best analysts? On Q3 2026 earnings, time series foundation models and the other models we serve through FoundationForecast often can. Public markets are open on paper. Quarterly results are free on EDGAR, but turning them into a forecast takes tools and training most people don't have. A single Bloomberg Terminal seat was priced at $31,980 a year for 2025, and about a quarter of U.S.-listed companies have no analyst coverage at all. Time series foundation models could put this kind of analysis within reach of far more people, and fill in the gaps where no analyst is looking. We think that only happens if the models are easy to use. FoundationForecast, our open-source unified API, ran time series foundation models and other models from our model garden, in this test through one interface, we locked the forecasts before each report, and on multiple report days, models from our model garden beat the Wall Street consensus. There's still a lot our industry needs to figure out. Rob Hyndman recently laid out the skeptical case: published comparisons often use weak baselines, benchmark datasets keep showing up in training data, and leaderboards can hide how a model does on the series that actually matter. When testing against classical methods, only 9 of the 24 foundation models beat AutoARIMA on report day.A company's next quarter can't leak, because it doesn't exist until it's reported. And GIFT-Eval rankings barely predicted who came closest. This is the second in our series of experiments at @timecopilot on what these models can and can't do yet. The coin is still in the air. Disney, Home Depot and Nvidia report on November 17, Walmart on November 19, and their forecasts are already locked, we will see how good or bad the models are on those stocks soon. If you want to dig deeper, the full article, interactive terminal, data and code here: write up experiments.timecopilot.dev/… explore the interactive terminal experiments.timecopilot.dev/…

founder life is tough, full of ups and downs. days filled with joy and uncertainty. if one thing keeps you going, it's seeing your product get adopted and knowing you're solving a real problem. so happy to share that @TimeCopilot ecosystem just crossed 60,000 downloads. @azulgarza_ and I are so stoked. that's 33% growth since our last milestone post, about one month ago. one thing has changed, the industry we're tackling is moving faster every week, and we can feel it accelerating. what hasn't changed is our belief that open source and verifiable outcomes are the way to build a company that can change forecasting and the wider world of structured data. if you haven't tried TimeCopilot or FoundationForecast yet, give them a shot. and thank you to everyone who has. you're the reason we keep building in the open.

so excited and happy to see that Databricks just integrated foundationforecast, our time series foundation model library. it feels surreal to have a $190 billion company using something you built. by adding us, they instantly got every model we support. 19 models at the time of integration, and 24 today.

Can time series foundation models predict next week’s AI token volume on OpenRouter? Yes, and they do well, but still have a long way to go. One big shortcoming is that standardized evaluation instruments really can’t tell us what’s the best model for real world problems. The leaders of the widely used time series benchmarks GIFT-Eval and fev-bench didn't come out on top in this experiment. Timestral a 209M parameter model, first on both, finished between 10th and 14th of 23 on total tokens. There's so much happening in the forecasting and machine learning communities right now, and time series foundation models are one of the biggest pieces of it. New models and benchmark results drop almost every week. For practitioners, the open question is still whether they actually work on real-world problems. At @TimeCopilot we are starting a new series of experiments to better understand the opportunities and shortcomings of these new architectures. The first one forecasts token usage for OpenRouter's top 50 LLMs, weekly and daily, using 21 foundation models plus seasonal naive and AutoARIMA. Every forecast ran through the hosted API of our newly launched library ✨𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝗙𝗼𝗿𝗲𝗰𝗮𝘀𝘁. If you want to dig deeper, the full article, data and code is here timecopilot.github.io/experi…