# Time Series Foundation Models A time series foundation model (TSFM) is a model pre-trained on a large mix of time series, so it can forecast a series it has never seen without being trained on it first. It's the [[AI Foundation Models|foundation model]] recipe (pre-train once, use everywhere) applied to numbers over time instead of text. Examples: Chronos (Amazon), TimesFM (Google), and [[Timestral]] (Mistral). The classic way to forecast is to fit one model per series: ARIMA, Holt-Winters, a gradient-boosted tree like XGBoost with hand-made features. That works well, and it's cheap. But it means someone has to build, tune and maintain each model. A TSFM skips that step. ## Where they make sense The best argument I've read for TSFMs came from the [[Hacker News]] thread about Timestral, and it's about scale, not accuracy: - **Thousands of series.** When you have 10,000 products or sensors, many series are too short or too noisy to fit a good model on their own. One generic model that has "seen" a lot of seasonality and trends does better than 10,000 badly fitted ones. One commenter runs about 400,000 models every 30 minutes. That's the kind of situation where a single pre-trained model starts to look attractive - **Good enough is fine.** Plenty of systems only need a rough forecast: autoscaling a load balancer, ordering spare parts, flagging an anomaly. A zero-shot forecast in 20 ms beats a perfect one nobody has time to build - **The middle of the market.** Between "cheap classical method" and "custom model built by a data scientist", there's a large gap. TSFMs live there ## Where they don't - **One important series.** If a single forecast drives real money (electricity prices, a treasury position), a tuned model built by someone who knows the domain will usually win. Several commenters said their TSFM tests "did not beat our AutoARIMA baselines" on noisy internal data, and that XGBoost per series was hard to beat - **Context the numbers don't contain.** The strongest objection in the thread: a small model has no idea what a "bank holiday" means. Covariates (extra series like a holiday calendar, the weather, a planned promotion) should be the center of the system, not an optional add-on. Long-term shifts are hard too, like solar panels and EVs slowly changing the shape of electricity demand - **When you need to understand the system.** A TSFM is a black box. It gives you forecasts and quantiles, not the dynamics behind them. Engineers who need a model of the physical system (failure prediction on engines, for example) are often better served by statistical or physics-based methods As one commenter put it, "time series data is fundamentally different from textual data". There's no grammar to learn, and two series that look alike can be driven by completely different causes. ## How to evaluate one without fooling yourself This was the most practical part of the discussion: 1. **Always include a naive baseline.** "Tomorrow looks like today" (persistence) or "this week looks like the same week last year" (seasonal naive). You'd be surprised how often a fancy model barely beats it 2. **Freeze forecasts when you make them.** Store what the model predicted at the time, then compare with what happened. Re-running a backtest later quietly leaks information 3. **Watch out for data leakage.** Public datasets may well be in the pre-training data. The cleanest test uses data collected after the model was released 4. **Distrust vendor benchmarks.** Everyone claims first place on GIFT-Eval or fev-bench. Your own data is the only benchmark that matters ([[Goodhart's Law]] applies to leaderboards too) ## My take TSFMs follow the same path as [[TabFM]] for tabular data and extraction models like [[Cloudflare Clairic]]: small, specialized models that return numbers with an uncertainty estimate, cheap enough to call everywhere. The pitch isn't "better than your data scientist". It's "good enough for the 95% of series your data scientist will never look at". Seen that way, they make a lot of sense. And one commenter suggested something I find more interesting: let an AI agent run the grid search over features and classical models for you. That might beat a generic model on the series that matter, and you keep a model you can actually inspect. ## References - Hacker News discussion about Timestral: https://news.ycombinator.com/item?id=50150434 - Timestral announcement (Mistral AI): https://mistral.ai/news/timestral/ ## Related - [[Timestral]] - [[AI Foundation Models]] - [[TabFM]] - [[Cloudflare Clairic]] - [[AI Model Calibration]] - [[Machine Learning (ML)]] - [[Goodhart's Law]] - [[Mistral AI]]