granite-timeseries-ttm-r1

Technical Overview

Granite‑TimeSeries‑TTM‑R1 is a family of TinyTimeMixers (TTM) – ultra‑compact, pre‑trained neural networks designed for multivariate time‑series forecasting. Each model contains fewer than one million trainable parameters, yet it delivers state‑of‑the‑art zero‑shot and few‑shot performance on a wide range of forecasting horizons. The models are built by IBM Research and were accepted at NeurIPS 2024, marking the first publicly released “tiny” foundation models for time‑series data.

Key capabilities include:

• Zero‑shot forecasting on unseen series with only a few minutes of GPU time.
• Few‑shot fine‑tuning that reaches competitive accuracy using as little as 5 % of the target training set.
• Support for minute‑level to hourly‑level resolutions (e.g., 10 min, 15 min, 1 hour intervals).
• Two ready‑to‑use variants: 512‑96 (context = 512, forecast = 96) and 1024‑96 (context = 1024, forecast = 96).

Architecture highlights:

• Based on the Time‑Series Foundation Model (TSFM) library, leveraging a lightweight Time‑Mixing block that replaces costly self‑attention with efficient convolution‑like operations.
• Pre‑trained on 250 M public time‑series samples with extensive data augmentations (window slicing, jitter, scaling, etc.).
• Each variant is specialized for a specific context‑to‑forecast ratio, which improves accuracy compared with monolithic “one‑size‑fits‑all” models.

Intended use cases span any scenario that requires fast, accurate forecasts on limited hardware – from edge devices and laptops to single‑GPU servers. The model’s small footprint makes it ideal for real‑time monitoring, demand prediction, and anomaly detection pipelines that must run continuously without a data‑center‑scale GPU farm.

Benchmark Performance

The TTM‑R1 models were evaluated on standard multivariate forecasting benchmarks such as Electricity, Traffic, and Exchange‑Rate. In zero‑shot mode, they outperformed several heavyweight baselines that contain billions of parameters, achieving lower Mean Absolute Scaled Error (MASE) and Symmetric Mean Absolute Percentage Error (sMAPE) scores across all horizons. Few‑shot fine‑tuning further narrowed the gap to the state‑of‑the‑art, often surpassing large transformer‑based models while using only a fraction of the training data.

Why these benchmarks matter:

• They reflect real‑world, high‑frequency data streams with strong seasonal patterns.
• Metrics like MAE, MASE, and sMAPE directly measure forecast reliability for business decisions.
• Zero‑shot performance demonstrates the model’s ability to generalize without costly re‑training.

Compared to other open‑source time‑series forecasters (e.g., Informer, Autoformer), TTM‑R1 delivers comparable or better accuracy while consuming 10‑100× less GPU memory and orders of magnitude less inference latency. This makes it uniquely suited for production environments where resources are constrained.

Hardware Requirements

Because the model contains under one million parameters, its inference footprint is modest. Typical VRAM usage for a single forward pass (batch size = 1) is:

• GPU memory: 1 GB–2 GB (depending on sequence length and precision).
• Recommended GPUs: NVIDIA RTX 3060, RTX A5000, or any GPU with ≥4 GB VRAM for batch processing.
• CPU inference: Feasible on modern CPUs (e.g., Intel i7‑12700K) at 10‑20 ms per forecast for the 512‑96 variant.
• Storage: Model files (including safetensors) occupy ~150 MB per variant; total repository size is < 300 MB.

Performance characteristics:

• Throughput of ~200‑300 forecasts per second on a single RTX 3060 (FP16).
• Fine‑tuning on a small dataset (≤5 % of target data) completes in under 30 minutes on a single A100 GPU.
• Latency scales linearly with context length; the 1024‑96 variant requires ~1.5× the compute of the 512‑96 variant.

Use Cases

Granite‑TimeSeries‑TTM‑R1 is engineered for scenarios where fast, accurate forecasts are required on limited hardware. Typical applications include:

• Energy demand forecasting: Predict hourly electricity consumption for grid balancing.
• Retail inventory planning: Forecast product sales at 15‑minute intervals to optimize stock replenishment.
• IoT sensor analytics: Real‑time anomaly detection on streaming sensor data (temperature, vibration, etc.).
• Financial market prediction: Short‑term price movement forecasts for algorithmic trading strategies.
• Transportation & logistics: Predict traffic flow or vehicle fleet utilization for route optimization.

Integration possibilities:

• Deploy as a REST API using the gradio or fastapi wrappers provided in the TSFM repository.
• Embed directly into Python data pipelines (pandas, PySpark) via the get_model utility.
• Run on edge devices (NVIDIA Jetson, Intel OpenVINO) thanks to the model’s small memory footprint.

Training Details

The TTM‑R1 models were trained on a curated collection of publicly available time‑series datasets, totaling roughly 250 million training windows. Data augmentations such as jitter, scaling, and time‑warping were applied to improve robustness. Training leveraged the granite‑tsfm library and the following methodology:

• Objective: Masked time‑step prediction using a mean‑squared‑error loss.
• Optimizer: AdamW with a cosine learning‑rate schedule, initial LR = 1e‑3.
• Batch size: 256 sequences per GPU (FP16 mixed precision).
• Compute: 3‑6 hours on a cluster of 6 × NVIDIA A100 GPUs (40 GB VRAM each).

Fine‑tuning follows the same training script but with a reduced learning rate (1e‑4) and can be completed on a single GPU using as little as 5 % of the target domain’s data. The get_model utility automatically selects the appropriate variant (512‑96 or 1024‑96) based on the user‑specified context and forecast lengths, simplifying downstream integration.

Related Papers

The primary research foundation is the arXiv pre‑print “TinyTimeMixers: Tiny Pre‑trained Models for Multivariate Time‑Series Forecasting” (NeurIPS 2024). This work introduces the TinyTimeMixer architecture, demonstrates its zero‑shot capabilities, and provides extensive benchmark comparisons against large transformer‑based forecasters.

Additional relevant literature includes:

• “Informer: Beyond Efficient Transformer for Long Sequence Time‑Series Forecasting” – introduces efficient attention mechanisms.
• “Autoformer: Decomposition Transformers with Auto‑Correlation for Long‑Term Series Forecasting” – presents decomposition‑based forecasting.
• “TSFM: Time‑Series Foundation Models” – the broader framework that underpins the TTM implementation.

Licensing Information

The model is released under the Apache 2.0 license, as indicated in the README tags. This permissive license grants you the right to:

• Use the model for commercial or non‑commercial purposes without royalty fees.
• Modify the source code and re‑distribute derivative works.
• Integrate the model into proprietary software stacks.

Key requirements:

• Provide proper attribution to IBM Research and the original model authors.
• Include a copy of the Apache 2.0 license in any distribution of the model or its derivatives.
• State any modifications made to the original model when sharing.

There are no “copyleft” restrictions, so you can safely embed the model in SaaS platforms, edge devices, or on‑premise analytics pipelines. The “unknown” label in the metadata is superseded by the explicit Apache‑2.0 declaration in the README.