Aligned series
Targets
(T, C)Past-only covariates
(P, C)Past-future covariates
(F, C + H)Future values must truly be known. Guesses belong in scenario analysis, not in the observed-data contract.
Zero-shot multivariate forecasting
How TimesFM‑3 sees the future.
It treats forecasting as masked completion on a two-dimensional grid: time runs across, related series run down, and known future signals stay visible while the model fills every target horizon in one pass.
Parameters330Mpublic checkpoint
Pretraining>1Ttime points
Input patch32contiguous steps
Output1 + 9point + quantiles
01 · EXPERIENCE IT
Change the evidence. Change the forecast.
Use a retail scenario to see why multivariate context matters. The chart is an explanatory simulation: it mirrors the mechanism described by Google, but it does not run the 330M-parameter model in your browser.
INTERACTIVE FORECAST LAB
DETERMINISTIC SIMULATIONIce cream demand · next 28 days
Sales forecast with planned promotions
Historical context on the left; 28-day forecast horizon on the right.
+19.8%on promotion days
observed targetunivariate projectionselected-input forecastknown future event
OBSERVABLE EFFECTFuture promotion tokens stay visible in the masked horizon, so the forecast can respond on the scheduled days.
02 · INPUT / MODEL / OUTPUT
Three contracts define the system.
Shapes are the cleanest way to see what each boundary owns. T is targets, C is context length, P/F are covariate channels, and H is the requested horizon.
→
Masked grid completion
Normalize
per variatePatch
32 stepsMix
time ↔ seriesDecode
single passTwenty transformer layers in the public checkpoint alternate the two relationship types.
→
Forecast distribution
Point forecast
(T, H)Quantiles
(T, H, 9)Levels
0.1 … 0.9Quantiles describe modeled uncertainty. Local calibration still needs to be verified with held-out data.
03 · TRACE THE FLOW
One forward pass, seven transformations.
Every step below changes the data representation. Select a stage or advance the trace to see its responsibility and hand-off.
Validate & align
STEP 1 / 7INPUT GATE
Check the forecasting contract
WHY IT MATTERS
04 · COMPONENT ANALYSIS
Each component owns one kind of uncertainty.
Inspect the architecture by responsibility. The tabs keep model behavior, tensor contract, and purpose tied together.
05 · EVIDENCE & LIMITS
What the launch establishes—and what it doesn’t.
Strong model explanations keep architecture facts, benchmark claims, and practical judgment in separate boxes.
Supported by official sources
Architecture & inference path
Patch size, lookahead tokens, alternating attention, masked-horizon decode, and nine quantiles.
GOOGLE RESEARCH BLOG ↗Checkpoint & public contract
20 layers, model dimension 1280, 16 heads, tensor shapes, data summary, and model license.
MODEL CARD ↗Decoder-only lineage
The original paper explains patch tokenization, residual projection, causal attention, and zero-shot pretraining.
ORIGINAL TIMESFM PAPER ↗Contiguous Patch Masking lineage
The launch credits TiRex’s CPM strategy for the single-pass masked-horizon approach.
TIREX / CPM PAPER ↗Do not over-read the result
01
No dedicated TimesFM‑3 paper yetThe model card cites the 2024 TimesFM paper. Version‑3 ablations and full training details are not exposed there.
02
Average rank is not universal dominanceTop rank across three suites does not guarantee the best model for every series, horizon, metric, or latency budget.
03
Association is not causationVariate attention can use correlations; it does not prove that one channel causes another.
04
Future covariates can leak or driftOnly information truly available at forecast time belongs in the future window.
05
Weights have a usage constraintThe current TimesFM‑3 weights are licensed for non-commercial, non-production use.
Trust the architecture enough to test it—not enough to skip the test.
Run a leak-free rolling-origin backtest against seasonal naive and strong task-specific baselines. Check point error, quantile calibration, latency, memory, and license fit. Reject deployment if any critical requirement fails.