TimeMosaic: Adaptive Granularity Patch and Segment-Wise Decoding for Time Series Forecasting

Time series forecasting is rarely a one-size-fits-all problem. Real-world processes drift and shift at different tempos: some phases unfold with lightning-fast volatility, others settle into long, quiet trends. Fixed-granularity models often chase the wrong tempo, trading accuracy for stability or vice versa. TimeMosaic offers a solution by embracing temporal heterogeneity head-on. It weaves a mosaic of time patches with adaptive granularity and pairs each patch with a dedicated decoding strategy, yielding forecasts that are both responsive and robust.

Adaptive Granularity Patch: A mosaic view of time

The core idea is to partition a time series into patches, but unlike traditional windowing, TimeMosaic lets the granularity—the length and resolution of each patch—vary across the series. Patches are data-driven and responsive to changes in dynamics. In volatile intervals, patches become finer, capturing rapid shifts without diluting signal with overly long windows. In stable intervals, patches can be coarser, reducing noise and saving compute.

This adaptive patching rests on lightweight heterogeneity signals rather than fixed calendars. Techniques such as local variance, entropy, and change-point cues guide where to split and how granular each segment should be. The result is a temporal mosaic where each tile reflects the true tempo of its neighborhood, rather than forcing a global tempo onto everything.

• Responsiveness: fine granularity when quick changes are likely, enabling timely forecasts.
• Efficiency: coarser granularity in smooth regions reduces unnecessary detail and computation.
• Interpretability: patch boundaries align with shifts in dynamics, offering intuitive insights into when and where the series behaves differently.

Segment-Wise Decoding: Tailoring forecasts to patches

Once the mosaic is established, each patch is paired with its own decoding pathway. Segment-wise decoding means a dedicated predictive head or model component focuses on the unique patterns within a patch. This specialization lets the decoder leverage the most relevant temporal cues for that segment—be it short-term momentum, weekly seasonality, or long-range trend—without being distracted by distant, dissimilar contexts.

Crucially, these segment decoders aren’t isolated silos. They feed into a fusion mechanism that combines patch-level forecasts into a coherent global prediction horizon. Gating and alignment layers ensure continuity across patch boundaries, so the overall forecast remains smooth where appropriate and sharply responsive where needed. The result is a forecast that respects local peculiarities while preserving a consistent global view.

How TimeMosaic Works in Practice

The end-to-end workflow unfolds in a few clear steps. First, the time series undergoes heterogeneity-aware segmentation to produce a patch map with varying granularity. Next, each patch is encoded with features that capture its internal dynamics—noise levels, seasonality strength, and recent momentum. Then, a segment-wise decoder operates on each patch, producing local forecasts tailored to that segment’s tempo.

A final fusion step reconciles the patch forecasts into a unified future horizon. Training TimeMosaic involves losses that respect both local patch accuracy and global forecast coherence, with regularization to prevent overfitting to transient quirks. The approach naturally supports multi-horizon forecasts, since each patch can contribute differently across future steps depending on its temporal scale.

Why TimeMosaic Outperforms Fixed Granularity Models

• Adaptive fidelity ensures the model pays attention where it matters—rapid changes get precise scrutiny, while steadier intervals avoid overreaction.
• Local expertise segment decoders specialize on the idiosyncrasies of each patch, improving accuracy in heterogeneous regimes.
• Better calibration across the forecast horizon, because the mosaic structure preserves both short-term responsiveness and long-term stability.
• Enhanced interpretability through patch-level diagnostics that reveal where and when the series behaves differently.

Applications and Scenarios

TimeMosaic shines in domains where non-stationarity and irregular tempo are the norms. Financial time series, where volatility clusters and regime shifts are common, can benefit from patch-aware forecasting. Energy demand and supply chains, with seasonality bursts and abrupt disruptions, are natural fits for adaptive granularity. Climate and environmental data—think weather proxies that swing between calm spells and storms—also stand to gain from the mosaic approach. In any setting where the right tempo varies over time, TimeMosaic provides a principled framework to align modeling granularity with reality.

TimeMosaic treats time as a mosaic, where every tile has a right size, and every tile can be forecasted with a tailored lens.

Implementation Considerations

Practitioners should balance patch granularity choices with available data, computational budget, and the forecast horizon. Patch boundaries are most useful when they reflect genuine regime shifts rather than noise. A lightweight segmentation module helps keep training scalable, and a modular decoder design makes it easy to experiment with different architectures per patch. When evaluating, consider both patch-level metrics and the aggregated horizon metrics to gauge where gains come from and where they may miss subtle cross-patch interactions.

Beyond accuracy, TimeMosaic offers robustness to non-stationarity. By distributing the modeling burden across patches that reflect local dynamics, the approach tends to be more resilient to abrupt changes, seasonal quirks, and irregular sampling. For teams ready to rethink how they structure time, TimeMosaic provides a practical blueprint for forecasting that moves with the tempo of the data.

Dreams of faster, smarter forecasts without sacrificing stability are within reach when the time series is allowed to reveal its own tempo—and when the model responds with a patch-by-patch decoding strategy tuned to that tempo.