Patch length

As explained in Time Series Forecasting Using Foundation Models

Input patch length is the number of time steps grouped into each input patch.

Patch length is the number of time steps grouped into each input patch. Each nonoverlapping patch becomes one token before embedding, reducing token count and computation while preserving longer input windows. section 6.1.1

Figure 6.2 The process of patching. Typically, each data point (circle) is a token that’s sent to the embedding layer. With patching, we group data points, and each patch becomes a token that is sent to the embedding layer.

Shows how grouping multiple data points into patches replaces individual tokens before the embedding layer. section 6.1.1

Moirai selects patch lengths heuristically from data frequency: higher-frequency series receive longer patches to control computation and retain long input windows. Monthly data uses three patch lengths, each with its own embedding layer. section 6.1.1

Frequency-dependent settings

Moirai uses length 8 for yearly or quarterly data; monthly data uses 8, 16, and 32; weekly or daily data uses 16 and 32; hourly data uses 32 and 64; every-minute data uses 32, 64, and 128; every-second data uses 64 and 128. section 6.1.1

Moirai-MoE fixes one patch length across frequencies, simplifying embedding through a single input projection. TimesFM uses an input patch length of 32 and forecasts output patches of 128 steps. section 6.2.1 section 7.1.1

How Time Series Forecasting Using Foundation Models covers it

How deep it goes

Explains token grouping, frequency heuristics, embedding consequences, and fixed-versus-variable designs without deriving model internals.

Examples

Uses frequency-specific Moirai settings, Moirai-MoE's fixed length, and TimesFM's 32-step input and 128-step output patches.

What you can do after

Evaluate patching designs by comparing token efficiency, embedding complexity, frequency specialization, and autoregressive forecasting steps.

How it connects

Connects tokenization, embedding layers, data frequency, input windows, autoregressive forecasting, and context length.

liveBookmanning.com

Read Time Series Forecasting Using Foundation Models on liveBook Get the book at manning.com

Related concepts

• Moiraiconfigured by · Moirai is a foundation model designed for time-series forecasting across varied datasets.
• Moirai-MoEconfigured by · Moirai-MoE is a Moirai-based universal time-series forecasting model that uses fixed-length, nonoverlapping patches across frequencies.

This explainer was generated by Manning's knowledge graph from the text of Time Series Forecasting Using Foundation Models. The section links point to the parts of the book it draws on.