🎯 What Is TimesFM-3
Google dropped TimesFM-3 — a foundation model that forecasts your sales, inventory, and revenue without ever training on your data. Surya, with 20 years in enterprise forecasting, was skeptical: "a model has to learn your business before it can predict your business." So he built an app and tried to make it fail. 0:00
The key upgrade from earlier versions: the original TimesFM could only read one signal. TimesFM-3 reads your whole business at once — past sales, weather, foot traffic, promotions — and does multivariate forecasting with zero training.
🧠 An LLM for Time Series
TimesFM has been around since 2024 — but unless you live in forecasting, you've never heard of it. 3:30
The same idea as Gemini/ChatGPT, inverted: LLMs read the internet's text and predict the next word; TimesFM reads the world's charts and predicts the next value. v1 trained on 100 billion time points (Google Trends, Wikipedia page views, traffic, weather, sales, synthetic data); the new one has read over a trillion.
"You hand it a chart it has never seen — your sales, your server traffic — and it predicts what comes next. The old TimesFM was blindfolded to promotions, weather, holidays; TimesFM-3 takes that blindfold off."
📱 The App He Built
Surya cloned the TimesFM repo and built a demand-planning app on top of it. 4:54
The setup: 330 million parameters (not a typo — tiny by LLM standards), 156 days of synthetic ice-cream data (ice cream, cone packs, syrups, promo flag, temperature, foot traffic), 128 days of history, 28-day horizon. Latency: 203 milliseconds for a forecast.
The evaluation metrics to watch: MAPE (mean absolute percentage error) and a revenue delta — how much incremental revenue the model attributes to each added signal.
🔬 Live Demo: Promotions, Weather, Foot Traffic
The test: forecast 28 days of ice-cream sales, layering in signals one at a time to see each one's effect. 6:36
| Signals used | Result |
|---|---|
| None (raw history only) | MAPE 4.7 — baseline forecast, misses the promo spikes |
| + Promotions | Captures the impact of promos, MAPE improves, shows incremental revenue |
| + Weather (Sep/Oct = cold) | Negative hit on ice-cream demand — as expected |
| + Foot traffic | Combined signal improves the forecast further |
The point: each added variable plays a measurable role — and "we did not have to train the model at all."
🥊 Head-to-Head vs Classical Models
TimesFM-3 against the models a data science team would actually build. 7:57
| Model | MAPE |
|---|---|
| TimesFM-3 | 4.73 |
| Holt-Winters (seasonal naive baseline) | 6.72 |
| Classical ML pipeline (ARIMA/XGBoost-style) | 8.89 |
TimesFM-3 wins by a lot — 4.73 vs 6.72 and 8.89. But the honest caveat: this is one test on one (synthetic) dataset, not a benchmark.
🔍 Open the Box: Under the Hood
The "open the box" section — how TimesFM-3 processes data without fine-tuning. 8:26
| Mechanism | What it does |
|---|---|
| Continuous patching | Reads the whole series at once instead of day-by-day — collapses 128 noisy days into 4 clean monthly snapshots of 32 days each. That's why it's fast and captures trend without single-day noise |
| Horizontal attention | Runs along time, each row looks strictly left at its own past — mathematically forbidden from peeking at the future. This is how it learns the organic baseline |
| Vertical attention | Within the same layer, flips to run across signals up/down the column — reads planned promo and forecast weather from the future rows |
"The future doesn't travel horizontally across time — it travels vertically up the column. The sales token glances down its September column, reads the planned promo and cooling temperature, and lifts demand right there. TimesFM-3 repeats this cycle 20 times through its transformer stack."
The analogy: a retail demand planner reading past sales charts left-to-right, then glancing down at the promotional calendar on their desk. That's the two-directional attention in one image.
⚖️ The Verdict
Did Google kill traditional forecasting? The transparent answer. 11:34
What it kills: the baseline grind — for a business with no forecasting team, this isn't an upgrade, it's "the difference between having forecasts and not having them."
What it does not kill: large retailers where half a percent of forecast accuracy is worth millions. Their teams will still train custom models because they know things the model cannot see — supply chain constraints, cannibalization, pricing strategy.
The license caveat: the model weights ship under a non-commercial license — free to research and evaluate, but for commercial production use the intended path is Google's platform (BigQuery integration).
💡 Key Takeaways
- TimesFM-3 forecasts with zero training — a foundation model that reads your whole business in one pass.
- It's an LLM for charts — predicts the next value the way an LLM predicts the next word, trained on a trillion time points.
- Multivariate is the big unlock — v1 read one signal; v3 reads sales + weather + promotions + foot traffic together.
- It beat classical models handily — MAPE 4.73 vs 6.72 (Holt-Winters) and 8.89 (classical ML) — on one synthetic test.
- Continuous patching + two-directional attention — reads months as patches, then alternates horizontal (past-only) and vertical (cross-signal) attention.
- 330M parameters, 203ms latency — tiny and fast by LLM standards.
- It kills the baseline grind, not custom models — for teams where half a percent is worth millions, custom training still wins.
- Non-commercial license — commercial use routes through Google's platform (BigQuery).
📍 Timestamp Index
0:00 Google drops TimesFM-3
1:39 Forecasting with zero training
2:41 Did Google kill traditional forecasting?
3:30 What is TimesFM (an LLM for time series)
4:54 Where to get the model + the app
6:36 Live demo: promos, weather, foot traffic
7:57 Head-to-head vs classical
8:26 Open the box: under the hood
11:34 The verdict
12:13 License & BigQuery