A comprehensive analysis of the time-series forecasting foundation model developed by Google Research
Introduction
TimesFM is a foundation model dedicated to time-series forecasting, developed by Google Research.
Traditionally, time-series forecasting required training a separate model for each domain. Retail demand forecasting, financial price forecasting, and manufacturing equipment anomaly detection each needed their own data and models. TimesFM changes this paradigm. Pre-trained on over 100 billion real-world time-series data points, this model can forecast new domains' time-series data instantly, without any additional training (zero-shot).
Just as ChatGPT did for the world of text, TimesFM is realizing the vision of "one model to forecast everything" for the world of time-series data.
It is based on the paper "A decoder-only foundation model for time-series forecasting," presented at ICML (International Conference on Machine Learning) 2024, and the current latest version is TimesFM 2.5 (September 2025).
Project Overview
| Item | Details |
|---|---|
| Developer | Google Research |
| Published | ICML 2024 |
| Latest version | TimesFM 2.5 (September 2025) |
| Parameter count | 200M (as of v2.5) |
| Pre-training data | Over 100 billion real-world time-series data points |
| License | Apache 2.0 open source (not an officially supported Google product) |
| Key integrations | BigQuery ML, Google Sheets, Vertex Model Garden |
| GitHub | https://github.com/google-research/timesfm |
| HuggingFace | google/timesfm-2.0-500m-pytorch |
TimesFM uses a patch-based decoder-only Transformer architecture. It tokenizes 32 consecutive timepoints into a single patch, structurally similar to how LLMs process text tokens.
Pros
1. Excellent Zero-Shot Performance
Provides immediately excellent forecasts on new time-series data with no additional training. Its performance has been validated across diverse domains — retail, finance, manufacturing, healthcare — and time granularities (seconds/minutes/hours/days/months/years).
2. Efficient Model Size
At 200M parameters, this is a very small size compared to modern LLMs. It's about 1/1000th the size of GPT-4 (hundreds of billions of parameters), yet delivers equal or superior time-series forecasting performance.
3. Long Context Length Support
TimesFM 2.5 supports up to a 16K context length. This is roughly 8x longer than the previous version (2,048), allowing years of daily data to be used as input in a single pass.
4. Probabilistic (Quantile) Forecasting Support
An optional 30M-parameter quantile head provides continuous quantile forecasts up to a 1K horizon. Rather than simple point forecasts, it enables probabilistic forecasting with confidence intervals.
5. Google Ecosystem Integration
- BigQuery ML: run large-scale time-series forecasting with a single line of SQL
- Google Sheets: use forecasting features directly from a spreadsheet
- Vertex Model Garden: integration with enterprise-grade MLOps pipelines
6. Multiple Backends and Extensibility
- Supports both PyTorch and Flax (JAX) backends
- Multi-GPU training and fine-tuning supported
- Efficient fine-tuning via HuggingFace Transformers + PEFT (LoRA)
7. XReg: Covariate Support
Starting with TimesFM 2.5, the XReg feature enables forecasting that incorporates external variables (weather, events, promotions, etc.).
8. Continuous Updates
Since 2025, various improvements have been made — a Flax version, XReg support, Agent Skills, and more — with active community pull requests being merged.
Cons
1. Not an Officially Supported Product
Although open source, it is not an officially supported Google product. Enterprise adopters shouldn't expect an SLA or official technical support.
2. Univariate-Centric
The publicly released checkpoints are optimized for univariate time-series forecasting. It's limited in multivariate forecasting scenarios where correlations among multiple variables matter.
3. Dependency and Installation Issues
Has specialized dependencies including JAX, Lingvo, and Praxis.
- Reported Python version compatibility issues
- Cases of
lingvopackage installation failures in Colab - Initial setup may require significant trial and error
4. Black-Box Model
Given its nature as a pre-trained foundation model, explaining the basis for its predictions is difficult. In domains like finance and healthcare where explainability matters, this poses regulatory risk.
5. Insufficient Validation in Specific Domains
Some research shows lower performance than Chronos-Bolt or Time-MoE in certain domains, such as electricity price forecasting. Validation in the target domain is essential before adoption.
6. Real-Time Processing Constraints
Real-time processing of large-scale time-series data requires building a separate serving infrastructure.
Potential Applications in Stock and Prediction Markets
TimesFM in Stock Market Forecasting
TimesFM has a structure directly applicable to stock price forecasting. However, the following considerations apply in practice.
Potential
- Processes OHLCV (open/high/low/close/volume) data as time series
- Zero-shot forecasting enables immediate prediction without per-stock training
- Quantile forecasts allow estimating price ranges and volatility
- Integration with BigQuery ML enables batch forecasting for large numbers of tickers
Limitations
- Stock markets are heavily influenced by unstructured data beyond time series — news, disclosures, sentiment
- May be vulnerable to market regime changes
- Additional layers (risk management, portfolio optimization) are needed to use it directly as a trading strategy
- May be better suited to medium/long-term trend forecasting than short-term (1-5 day) forecasting
Practical use cases
| Use case | Fit | Description |
|---|---|---|
| Stock trend direction forecasting | ★★★☆☆ | Directional forecasting is possible but accuracy is limited |
| Volatility forecasting | ★★★★☆ | Effective when using quantile forecasts |
| Demand/revenue forecasting (corporate analysis) | ★★★★★ | Useful for estimating revenue ahead of earnings releases |
| Macroeconomic indicator forecasting | ★★★★☆ | CPI, interest rates, and other economic indicators |
| Portfolio rebalancing trigger | ★★★☆☆ | Used for detecting trend reversal points |
Applications in Prediction Markets
Direct Applicability on Polymarket
Polymarket is a prediction market where users trade the YES/NO probability of specific events. Categories where TimesFM is directly useful:
High fit-Economic indicator markets: markets like "Will US CPI exceed 3% in 2025?" using historical CPI time series
- Financial events: forecasting rate paths such as "Number of Fed rate cuts in 2025"
- Commodity prices: markets related to oil and gold prices
Low fit
- Political events (election results, etc.) — polling data is more suitable than time series
- One-off events (sports results, etc.) — no time-series pattern exists
Polymarket usage strategy example
1. Select target market: "US unemployment rate above 4.5% in December 2025?"
2. Collect historical data: download monthly unemployment rate time series from the FRED API
3. Run TimesFM forecast: quantile forecast for the next 6 months
4. Convert to probability: calculate the probability of exceeding the threshold from the forecast distribution
5. Compare with market price: search for arbitrage opportunities against the current Polymarket price
Domestic (Korean) Prediction Markets
Korea does not yet have a full-fledged prediction market like Polymarket. Prediction markets and platforms like Polymarket are illegal in Korea.
- Informal prediction channels like KakaoTalk open-chat public sentiment aggregation-Securities firm research consensus — TimesFM can be used to forecast earnings surprises relative to consensus
Comparison with Competing Projects
Global Competing Models
| Model | Developer | Country | Parameters | Open Source | Zero-Shot | Multivariate | Probabilistic |
|---|---|---|---|---|---|---|---|
| TimesFM 2.5 | Google Research | US | 200M | Yes | Yes | Limited | Partial |
| TimeGPT | Nixtla | US | Not disclosed | No (commercial API) | Yes | Yes | Yes |
| Chronos-Bolt | Amazon | US | Various | Yes | Yes | Yes | Yes |
| Moirai | Salesforce | US | Various | Yes | Yes | Yes | Yes |
| Time-MoE | Beihang Univ. | China | Various | Yes | Yes | Partial | Limited |
| MOMENT | CMU | US | Various | Yes | Yes | Yes | Limited |
| Lag-Llama | ServiceNow | Canada | Various | Yes | Yes | No | Yes |
| UniTS | Fudan Univ. | China | Various | Yes | Yes | Yes | Limited |
| TimesNet | Tongji Univ. | China | Various | Yes | No | Yes | No |
Detailed Analysis of Chinese Competing Projects
1. Time-MoE (Time Mixture of Experts)
- Developer: Beihang University
- Feature: first application of a Mixture of Experts (MoE) architecture to time series
- Strength: outperforms TimesFM in Electricity Price Forecasting
- Weakness: limited probabilistic forecasting capability, relatively small community
- GitHub: https://github.com/Time-MoE/Time-MoE
2. UniTS (Universal Time Series)
- Developer: Fudan University
- Feature: a single model handles forecasting, classification, anomaly detection, imputation, and other tasks
- Strength: multi-task learning, multivariate support
- Weakness: possible performance gap versus specialized models on individual tasks
- Note: applies Chinese NLP research groups' LLM methodology to time series
3. TimesNet
- Developer: Tongji University
- Feature: an original approach that transforms 1D time series into 2D space for CNN processing
- Strength: covers five tasks — long-term forecasting, short-term forecasting, imputation, anomaly detection, classification
- Weakness: an architecture research effort rather than a foundation model; no zero-shot
- GitHub: https://github.com/thuml/Time-Series-Library
4. PatchTST
- Developer: Tsinghua University + IBM Research
- Feature: a patch-based approach similar to TimesFM, Transformer-based
- Note: can be seen as an architectural precursor to TimesFM
5. iTransformer
- Developer: Chinese Academy of Sciences
- Feature: applies the Transformer's attention mechanism to the variable dimension rather than the time dimension
- Strength: excellent performance in multivariate forecasting
Detailed Analysis of US/Canadian Competing Projects
1. Chronos & Chronos-Bolt (Amazon)
- Feature: T5-architecture based, tokenizes time-series values and processes them as a language model
- Strength: the strongest at probabilistic forecasting; outperforms TimesFM in electricity price forecasting
- Weakness: inference speed greatly improved in Chronos-Bolt but still relatively heavy
- Note: supports one-click deployment via AWS SageMaker JumpStart
2. TimeGPT (Nixtla)
- Feature: the first commercial time-series foundation model, offered as an API
- Strength: easiest-to-use interface, fine-tuning support, includes anomaly detection
- Weakness: not open source, incurs API costs, internal model structure undisclosed
- Pricing: $29-$299/month (as of 2025)
3. Moirai (Salesforce)
- Feature: patch-based masked encoder, handles various frequencies with a single model
- Strength: multivariate support, probabilistic forecasting, various model sizes (Small/Base/Large)
- Weakness: inference speed, some performance gaps versus TimesFM in specific domains
4. MOMENT (CMU)
- Feature: a large pre-trained model for time-series analysis, uses a masking approach
- Strength: broad task coverage — anomaly detection, classification, imputation
- Note: supervised-learning based; fine-tuning maximizes performance
Performance Comparison (Research Summary)
| Task | 1st place | 2nd place | TimesFM ranking |
|---|---|---|---|
| Long-horizon forecasting | TimesFM | Moirai | 1st |
| Electricity price forecasting | Chronos-Bolt | Time-MoE | 3rd |
| Short-term load forecasting | TimesFM | Chronos | 1st |
| Monthly precipitation forecasting | TimesFM | SARIMA | 1st (for some seasons) |
| Multivariate forecasting | Moirai | iTransformer | Lower ranks |
| Anomaly detection | MOMENT | UniTS | N/A |
Getting Started Guide
Installation
# PyTorch version (recommended)
pip install timesfm[torch]
# JAX/Flax version
pip install timesfm[jax]
Basic Forecasting Example (PyTorch)
import timesfm
import numpy as np
# Load model
tfm = timesfm.TimesFm(
hparams=timesfm.TimesFmHparams(
backend="torch",
per_core_batch_size=32,
horizon_len=128,
),
checkpoint=timesfm.TimesFmCheckpoint(
huggingface_repo_id="google/timesfm-2.0-200m-pytorch"
),
)
# Run forecast
forecast_input = [np.sin(np.linspace(0, 20, 100))] # example time series
point_forecast, quantile_forecast = tfm.forecast(forecast_input)
BigQuery ML Integration
-- Example of using TimesFM in BigQuery ML
SELECT *
FROM ML.FORECAST(
MODEL `project.dataset.timesfm_model`,
STRUCT(30 AS horizon, 0.9 AS confidence_level)
)
Adoption Roadmap (Investment Context)
Phase 1: Data Collection
├── FRED API: macroeconomic time series (GDP, CPI, unemployment rate)
├── Yahoo Finance / FinanceDataReader: stock price data
└── DART: corporate quarterly earnings data
Phase 2: Applying TimesFM
├── Measure baseline performance via zero-shot forecasting
├── LoRA fine-tuning with domain data
└── Generate confidence intervals via quantile forecasting
Phase 3: Signal Generation
├── Compare forecast vs. consensus
├── Compare Polymarket prices against TimesFM-derived probabilities
└── Compute risk-adjusted returns
Phase 4: Production Application
├── Portfolio rebalancing trigger
├── Options strategy (using volatility forecasts)
└── Prediction market arbitrage
Summary Comparison Table
| Feature | TimesFM | TimeGPT | Chronos | Moirai | Time-MoE |
|---|---|---|---|---|---|
| Developer | Nixtla | Amazon | Salesforce | Beihang Univ. | |
| Parameters | 200M | Not disclosed | Various | Various | Various |
| Open source | Yes | No | Yes | Yes | Yes |
| Zero-shot | Yes | Yes | Yes | Yes | Yes |
| Multivariate | Limited | Yes | Yes | Yes | Partial |
| Probabilistic forecasting | Yes | Yes | Yes | Yes | Limited |
| Google integration | Yes | No | No | No | No |
| Inference speed | ★★★★☆ | ★★★★☆ | ★★★☆☆ | ★★★☆☆ | ★★★★☆ |
| Community | ★★★★★ | ★★★☆☆ | ★★★★☆ | ★★★☆☆ | ★★★☆☆ |
Conclusion
TimesFM is a lightweight time-series foundation modeldeveloped by Google Research, with its main strengths being zero-shot forecasting performance and Google ecosystem integration. It delivers strong performance despite a compact 200M-parameter size, and can be adopted at enterprise scale via BigQuery ML.Application points in an investment/finance context- More suited tomacroeconomic indicators, corporate revenue, and demand forecasting than raw stock prices
- Usable for probability estimation in economic-indicator-related markets on prediction platforms like Polymarket
- Synergy is maximized when combined with LLM-based news sentiment analysis rather than used standalone
Selection criteria- BigQuery/GCP environment ->TimesFM- AWS environment ->Chronos-Bolt- API-only requirement ->TimeGPT- Multivariate + probabilistic forecasting ->Moirai- Power/energy domain ->Time-MoE
References
- GitHub repository
- Paper (arXiv)
- HuggingFace checkpoint
- Google Research blog
- BigQuery ML integration docs
Last updated: June 2025 | Written by: Vibe Investing Research