How does Timer-M1 learn primitives for multivariate time series forecasting?

Primitives as the shared unit across domains

The paper's premise is that time series from different domains share elementary patterns even when their overall dynamics differ. It calls these patterns primitives and splits them into two kinds. Temporal primitives describe regularities within a series: trend, periodicity, stochastic dependence, and regime changes. Relational primitives describe how variates relate: cross-variate correlation, delayed responses, cointegration, and event effects. Because these patterns are defined independently of any particular forecasting scenario, the authors argue they can be recombined and still remain useful when the surrounding signal or the available observations change.

This framing explains the design choice that follows. Rather than training only on whatever patterns real datasets happen to contain, the pipeline deliberately constructs samples in which the presence of a primitive is controlled. The paper motivates this by the complementarity of the two data sources: real series record complex, noisy dynamics but offer limited control over which patterns appear, while synthetic series offer control but may omit the irregularities of real systems.

Building samples from generated and real series

The synthesis pipeline first generates series that carry temporal primitives shared across domains. It then assembles real and generated series into multivariate samples using relational primitives, so that cross-variate structure is imposed by construction rather than merely observed. The result is a training corpus in which both the within-series shape and the between-series relation are specified.

Samples are then organized into episodes. Each episode assigns distinct channel roles: target variates to be predicted, past-only covariates available only in history, and known-future covariates available over the forecast horizon. This role assignment is what makes the objective multivariate rather than a set of independent univariate problems, because the model is optimized on predictable variates while using the exogenous information the episode makes available. The paper states that episodes vary what is observed and predicted while introducing noise and interference, which it links to generalization to new forecasting scenarios.

Gated blocks decide how much variates interact

On the architecture side, Timer-M1 adapts gated two-dimensional Transformer blocks that dynamically allocate cross-variate attention across layers. In plain terms, the model can vary how strongly it mixes information between variates depending on depth, instead of applying one fixed interaction strength everywhere.

The ablation isolates this choice. Learned layer-wise coefficients reach FEV SQL skill 0.4887, compared with 0.4836 for a fixed gate set to 0.202, the mean value from the stable checkpoint, and 0.4814 for a fixed gate of 1.0 that applies unattenuated cross-variate updates. The learned gates improve three of five task types, and the paper reports that the learned coefficients generally increase with depth, which it reads as stronger cross-variate integration in deeper layers. The comparison against the mean-matched fixed gate matters because it holds overall interaction strength roughly constant while removing only the layer-wise variation.

Joint generation adds more than source diversity

The data ablation separates four recipes: real data only, synthetic data only, both sources without joint generation, and the full recipe that composes them through shared relation operators. The full recipe reaches FEV skill 0.4887, versus 0.4431 for real only, 0.4710 for synthetic only, and 0.4725 for no-joint. It also gives the lowest GIFT-Eval MASE (0.6806) and CRPS (0.4688) and the lowest TIME MASE (0.6372), though TIME CRPS is marginally higher than the no-joint variant (0.5363 versus 0.5359).

The informative comparison is full recipe against no-joint, since both use real and synthetic sources. The FEV skill gap of 0.0162 is the paper's evidence that composing the two sources through shared relation operators contributes beyond merely pooling them. That supports the authors' claim for a unified construction pipeline rather than two independent training collections, and it is the result to weigh if the question is whether joint generation is worth the added complexity.

First on FEV and TIME, second on GIFT-Eval

Across three large-scale forecasting benchmarks, Timer-M1 ranks first on FEV and TIME and second on GIFT-Eval among recent time series foundation models. On GIFT-Eval, across 97 matched configurations, it ranks second among 12 compared pretrained models on both MASE and CRPS, with MASE 0.681 and CRPS 0.469. These are aggregate rankings over matched configurations, not a claim of best performance on every individual task.

The benchmarks do not report the same quantity, and the paper is explicit about this. FEV reports skill, computed as one minus a relative aggregate error against Seasonal Naive, with each relative error clipped to [0.01, 100] before aggregation, and reports MASE and SQL skill separately. TIME and GIFT-Eval report normalized MASE and quantile-based CRPS, where higher skill and lower normalized error indicate better forecasts. A first-place FEV skill score and a second-place GIFT-Eval MASE therefore answer different questions and should not be read as one combined measure of superiority.

Controlled tests probe each primitive separately

Beyond the benchmarks, the paper runs a primitive evaluation that fixes a generated trajectory and forecast origin and specifies which observations each model receives. Temporal extrapolation uses 256 historical observations and 128 forecast steps and covers periodicity, trend, and shift groups. Relational forecasting uses the same window and constructs multiple responses from shared sources, including signed, delayed, phase-shifted, scale-varying, cointegrated, and shared-regime relations. Input reliability uses 512 historical observations and 128 forecast steps and varies missingness, corruption, irrelevant inputs, event indicators, and variate order.

The reported aggregates are narrow and should be read with their denominators. Timer-M1 reaches NMAE 0.02432 across 13 periodic cases and NMAE 0.02160 across six joint-target tasks. On input reliability, it reduces target-missingness NMAE by 32.5% relative to Chronos-2. The paper also states the limits of these tests: the unannounced boundary changes are interpreted as stress tests, not as evidence that a model can anticipate an unobserved intervention, and the cross-model relational comparisons are not by themselves estimates of the gain from joint rather than independent prediction. Missingness results concern forecasting after the specified common preprocessing, not a separate imputation task.

What the primitive evidence does not establish

The controlled evaluation groups are defined by mechanism rather than by which model attains the smallest error, and unsupported input interfaces are left unranked rather than assigned an artificial penalty. Group scores are arithmetic means over scoring records, with one record per evaluated target and condition, so the count is not necessarily the number of independent generated trajectories. MSE and NMAE are not interchangeable, and the paper infers no cross-capability composite score from the selected groups.

The authors also set the scope for future work: expanding real-data coverage and primitive compositions to better represent irregular dynamics and partially observed dependencies in complex real-world systems, and exploring input-dependent cross-variate interaction and more efficient long-context modeling. These are stated as directions, not demonstrated results, so the current evidence supports primitive-based pretraining as a route to general forecasting across the tested domains and task settings, not a guarantee for dynamics outside the constructed primitive set.

The priority is joint construction, not more data

If one result should guide a reader's judgment, it is the gap between the full recipe and the no-joint variant: 0.4887 versus 0.4725 FEV skill, with both using real and synthetic sources. That comparison isolates how the two sources are combined, and it is the paper's strongest internal argument that the value of the pipeline lies in composing observed dynamics with controlled relational primitives rather than in adding either source alone. The gate ablation points in the same direction at the architecture level, where layer-wise allocation beats a mean-matched fixed gate by 0.0051 FEV skill.

The practical reading is therefore conditional. Primitive-based pretraining is supported where the target behavior can be expressed through the temporal and relational primitives the pipeline constructs, and the benchmark ranks and controlled tests are the evidence for that support. For behavior that depends on irregular real-world dynamics or partially observed dependencies beyond those compositions, the paper itself defers the answer to future work.

Sources used in this answer