Time-300B: Multi-Domain Time Series Dataset

• Time-300B is a comprehensive, multi-domain time series dataset featuring over 300 billion time-stamped observations across diverse fields.
• It applies sophisticated preprocessing, sliding-window segmentation, filtering, and normalization across multiple resolutions to ensure balanced and high-quality training windows.
• The dataset underpins advancements in zero-shot and cross-domain forecasting, serving as the backbone for models like Cisco Time Series Model and Time-MoE.

Time-300B is a large-scale, multi-domain time series dataset developed to enable the pre-training of foundation models for forecasting tasks across diverse modalities, resolutions, and real-world environments. Designed explicitly to support zero-shot and multi-domain generalization, Time-300B serves as the backbone for major advancements in time series foundation models, including the Cisco Time Series Model and Time-MoE, and is characterized by its scale, heterogeneous sources, multiresolution context structure, sophisticated preprocessing, and open-access ethos (Gou et al., 25 Nov 2025, Shi et al., 2024).

1. Composition and Scale

Time-300B consists of over 300 billion unique time-stamped observations spanning millions of individual time series collected or synthesized across several domains. Its construction synthesizes three principal sources:

• Proprietary observability time series:

Approximately 400 million distinct time series collected over 13 months at 1-minute (majority) and 5-minute resolutions, yielding an initial pool of ∼3.1 × 10¹² raw minute-level samples before windowing. After filtering and windowing, these contribute ∼35% (1-minute) and 16.5% (5-minute) of the final points (Gou et al., 25 Nov 2025).

• Public multi-domain benchmarks:
  • GIFT-Eval: 4.5 million series, 230 billion points (∼29.5% of final pool); domains include traffic, finance, energy, and web (Gou et al., 25 Nov 2025).
  • Chronos: 0.9 million series, 85 billion points (∼4.5% of pool).
  • Additional datasets summarized in nine domain categories: Energy, Finance, Healthcare, Nature, Sales, Synthetic, Transport, Web, and Other. For instance, the Nature domain (with sources like ERA5, CMIP6, and WeatherBench) represents 31,621,183 series and 90.5% of the sequence count in one configuration, amounting to 279.724B observations (Shi et al., 2024).
• Synthetic data (KernelSynth):

Generated using mixtures of periodic and trend kernels to produce high-frequency, operationally realistic sequences, contributing ∼14.5% of total datapoints (Gou et al., 25 Nov 2025, Shi et al., 2024).

A strictly sliding-window approach is used: series are segmented into overlapping context-horizon windows (context length = 512 points per resolution; horizon length = 128 points) with stride length tuned to avoid domain or source dominance.

Domain and Series Statistics

Domain	# Series	# Observations (B)	% Obs
Nature	31,621,183	279.724	90.50
Synthetic	11,968,625	9.222	2.98
Energy	2,875,335	15.981	5.17
Web	972,158	1.804	0.58
Transport	622,414	2.130	0.69
Sales	110,210	0.026	0.008
Other	40,265	0.020	0.006
Finance	1,715	0.0004	0.0001
Healthcare	1,752	0.0005	0.0001

(Data sourced from (Shi et al., 2024), Table 1. All numbers are approximate.)

2. Data Domains, Modalities, and Source Taxonomy

Time-300B encompasses nine primary domains:

• Observability: Infrastructure metrics (CPU, memory, disk), application/service telemetry (request rates, latency), real-user and component metrics.
• Energy: Electricity load, building/power consumption, renewables.
• Finance: Exchange rates, stock indices, macroeconomic indicators.
• Healthcare: Hospital visits, epidemic time series, mortality.
• Nature: High-volume environmental data (weather, air quality, sunspots).
• Sales: Point-of-sale, retail, hierarchical aggregates, forecasting competition series.
• Transport: Traffic, mobility, public transit, ridesharing.
• Web: Web traffic, cluster resource traces, Wikipedia pageviews.
• Synthetic: Constructed via TS-Mixup, KernelSynth, and similar methods to model diverse periodic and stochastic phenomena.

Sampling frequencies range from sub-minute to annual, depending on domain (e.g., 4s, 15min, 1h, 1d, monthly, and synthetic, with no inherent physical time granularity).

3. Multiresolution Structure and Temporal Context

A defining characteristic is the multiresolution example design (Gou et al., 25 Nov 2025):

• Training examples: Each consists of
  • Coarse-resolution context $x_c \in \mathbb{R}^{512}$ (e.g., 512 points of 1-hour aggregates).
  • Fine-resolution context $x_f \in \mathbb{R}^{512}$ (512 raw 1-minute points).
  • Fine-resolution horizon $y \in \mathbb{R}^{128}$ (128 future 1-minute points).
• Resolution ratio: $K=60$ (i.e., 60 fine-resolution points per coarse point).
• Training regime: Multiresolution transformer models fuse $x_c$ and $x_f$ for prediction; at inference, fine-resolution predictions are rolled up to update the coarse context via:

$$\left[ \frac{1}{K} \sum_{j=1}^{K} \hat{y}_{(k-1)K+j} \right]_{k=1}^{\lfloor L/K \rfloor}$$

• Both 1-minute/1-hour and 5-minute/5-hour resolutions are included.

This design explicitly enables the learning of temporal dependencies across multiple scales, critical for operational and long-horizon forecasting (Gou et al., 25 Nov 2025).

4. Preprocessing, Filtering, and Normalization

The Time-300B preprocessing pipeline implements extensive per-series, per-window, and per-batch procedures (Gou et al., 25 Nov 2025, Shi et al., 2024):

• Series-level filtering:
  • Exclude series with excessive missing values or extended flat intervals.
  • Fill short gaps by last-value extrapolation.
  • Apply first-difference transform to cumulative counters to mitigate artificial growth.
• Window-level filtering:
  • Reject windows with high unpredictability, quantified by the ratio:

  $$\frac{\max|\text{horizon deviation}|}{\max|\text{context deviation}|}$$

  above a fixed threshold. - Downsample windows with excessive spectral entropy. - Remove windows with too few unique values. - Adapt stride per dataset to avoid overrepresentation.

• Deduplication (fine windows):

  • SimHash clustering and sampling to ensure diversity; histogram-based sampling for large clusters.
• On-the-fly context normalization:
  • For each context $(x_c, x_f)$, calculate mean and std of the first 32 points.
  • Normalize:

  $$x_c \leftarrow \frac{x_c - \mu_c}{\sigma_c},\quad x_f \leftarrow \frac{x_f - \mu_f}{\sigma_f}$$ - Model predicts normalized $y$; outputs are un-normalized at inference.

• Synthetic augmentation:

  • Synthetic series from KernelSynth (Gaussian + periodic kernels), designed for high-frequency periodic structure.
  • No adversarial or noise-based augmentation beyond data selection and deduplication.

For the public Time-300B variant (Shi et al., 2024), preprocessing includes:

• Missing-value splitting: segmentation at $\mathrm{NaN}$/$\pm\infty$,
• Window-quality filtering: sliding window zero/first/second-difference flatness tests (thresholds at 0.2),
• Downsampling large domains,
• No explicit global normalization.

5. Data Splits, Batching, and Sampling Strategies

Partitioning and batching approaches are tailored for large-scale pretraining and robust evaluation (Gou et al., 25 Nov 2025, Shi et al., 2024):

• Splits (Cisco Time Series Model):
  • Sequence windows are sorted by timestamp, partitioned into strictly train, validation, and test sets (preventing temporal leakage).
  • Entire series are held out for zero-shot, out-of-domain testing.
  • Test windows originate from time intervals later than the latest in train/validation.
• Splits (Time-MoE):
  • No explicit corpus-wide train/val/test splits for pretraining. During pretraining, batches are sampled to maintain balanced domain and value distribution. Only benchmark datasets (e.g., ETT, Global Temp) use classical splits for downstream evaluation.
• Batching and packing:
  • Short sequences are concatenated (packed) to minimize padding.
  • Random cropping produces fixed-length subsequences per batch (max length 4,096).
  • Batches: 1,024 sequences per step (~4 million observations).
  • No MoE routing during data prep; routing occurs during model forward pass (Shi et al., 2024).

6. Storage, Format, and Access

Time-300B is distributed as sharded, binary-encoded records accompanied by detailed metadata (Gou et al., 25 Nov 2025, Shi et al., 2024):

• Object storage sharding:
  • Each shard contains serialized windows $(x_c, x_f, y)$ (Cisco) or subsequences (Time-MoE) and per-window metadata (e.g., series ID, source domain, resolution, cluster code, entropy score, timestamps).
  • Public Time-300B includes a metafile indexing by (dataset, start_offset, end_offset) to enable constant-memory streaming.
• Data parallel training:
  • Shards are statically mapped to compute resources; normalization and filtering are applied on-the-fly by data workers.
• Licensing and availability:
  • Public Time-300B consolidates datasets from UCI, Kaggle, Zenodo, and related sources, preserving upstream licensing. The unified dataset and loader utilities are open-source, released with documentation at https://github.com/Time-MoE/Time-MoE (Shi et al., 2024).

7. Significance and Applications

Time-300B represents a foundational advance in time series data infrastructure for large model pretraining (Gou et al., 25 Nov 2025, Shi et al., 2024):

• Its multi-domain, multiresolution pipeline supports both general-purpose and observability-specialized foundation models.
• The dataset demonstrates the practical application of scaling laws (tokens, model size) for time series, paralleling advances in NLP and vision.
• Time-300B supports univariate, zero-shot, and cross-domain forecasting without per-series fine-tuning, enabling benchmarking for both dense and Mixture-of-Experts (MoE) architectures.
• The inclusion of both proprietary and open-access variants enables reproducibility, benchmarking, and further research using modern GPU-scale learning pipelines.

The design principles of Time-300B—scale, diversity, multiresolution, and balance—emerged as critical properties for enabling the next generation of zero-shot time series forecasting and efficient, cross-domain transfer (Gou et al., 25 Nov 2025, Shi et al., 2024).