alterlab-timesfm

TimesFM Forecasting

Overview

TimesFM (Time Series Foundation Model) is a pretrained decoder-only foundation model developed by Google Research for time-series forecasting. It works zero-shot — feed it a time series and it returns point forecasts with quantile prediction intervals, no training required. The timesfm package (current 3.0.2) ships two model APIs:

• TimesFM 2.5 (200M, Apache-2.0 weights) — timesfm.TimesFM_2p5_200M_torch with ForecastConfig; univariate, optional covariates via XReg. The default in this skill and the only choice for commercial or production use.
• TimesFM 3.0 (~330M, released Aug 2026) — timesfm3.TimesFM3Forecaster; univariate and multivariate forecasting with native past-only and past-and-future covariates. Its weights are under timesfm-non-commercial-license-v1.0, so use it only for non-commercial research and tell the user about the restriction.

This skill wraps TimesFM for safe, agent-friendly local inference. It includes a mandatory preflight system checker that verifies RAM, GPU memory, and disk space before the model is ever loaded so the agent never crashes a user's machine.

Key numbers: TimesFM 2.5 uses 200M parameters (0.93 GB safetensors); TimesFM 3.0 uses ~330M (1.3 GB). Run the system checker before the first load so an under-resourced machine fails fast instead of swapping or crashing.

When to Use This Skill

Use this skill when:

• Forecasting any univariate time series (sales, demand, sensor, vitals, price, weather)
• You need zero-shot forecasting without training a custom model
• You want probabilistic forecasts with calibrated prediction intervals (quantiles)
• You have time series of any length (the model handles 1–16,384 context points)
• You need to batch-forecast hundreds or thousands of series efficiently
• You want a foundation model approach instead of hand-tuning ARIMA/ETS parameters
• You have related channels or known future drivers (TimesFM 3.0 multivariate + covariates, or TimesFM 2.5 XReg)

Does NOT Trigger

Scenario	Use Instead
Classical models with interpretable coefficients (ARIMA/SARIMAX tables), VAR, or Granger causality tests	alterlab-statsmodels
Time-series classification, clustering, segmentation, or similarity search	alterlab-aeon
Exploring a time-series file's structure and quality before any forecasting	alterlab-eda
Tabular (non-temporal) prediction	alterlab-scikit-learn

Note on Anomaly Detection: TimesFM does not have built-in anomaly detection, but you can use the quantile forecasts as prediction intervals — values outside the 80% CI (q10–q90) are statistically unusual. See the examples/anomaly-detection/ directory for a full example.

Preflight: System Requirements Check

Run the system checker before loading a model for the first time on a machine: loading downloads ~1 GB of weights and allocates several GB of RAM, and the checker stops early with a clear message instead of letting the load crash or swap.

python scripts/check_system.py

This script checks:

1. Available RAM — warns if below 4 GB, blocks if below 2 GB
2. GPU availability — detects CUDA/MPS devices and VRAM
3. Disk space — verifies room for the ~800 MB model download
4. Python version — requires 3.10+
5. Existing installation — checks if timesfm and torch are installed

Note: Model weights are NOT stored in this repository. TimesFM weights (~800 MB) download on-demand from HuggingFace on first use and cache in ~/.cache/huggingface/. The preflight checker ensures sufficient resources before any download begins.

flowchart TD
    accTitle: Preflight System Check
    accDescr: Decision flowchart showing the system requirement checks that must pass before loading TimesFM.

    start["🚀 Run check_system.py"] --> ram{"RAM ≥ 4 GB?"}
    ram -->|"Yes"| gpu{"GPU available?"}
    ram -->|"No (2-4 GB)"| warn_ram["⚠️ Warning: tight RAM<br/>CPU-only, small batches"]
    ram -->|"No (< 2 GB)"| block["🛑 BLOCKED<br/>Insufficient memory"]
    warn_ram --> disk
    gpu -->|"CUDA / MPS"| vram{"VRAM ≥ 2 GB?"}
    gpu -->|"CPU only"| cpu_ok["✅ CPU mode<br/>Slower but works"]
    vram -->|"Yes"| gpu_ok["✅ GPU mode<br/>Fast inference"]
    vram -->|"No"| cpu_ok
    gpu_ok --> disk{"Disk ≥ 2 GB free?"}
    cpu_ok --> disk
    disk -->|"Yes"| ready["✅ READY<br/>Safe to load model"]
    disk -->|"No"| block_disk["🛑 BLOCKED<br/>Need space for weights"]

    classDef ok fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#14532d
    classDef warn fill:#fef9c3,stroke:#ca8a04,stroke-width:2px,color:#713f12
    classDef block fill:#fee2e2,stroke:#dc2626,stroke-width:2px,color:#7f1d1d
    classDef neutral fill:#f3f4f6,stroke:#6b7280,stroke-width:2px,color:#1f2937

    class ready,gpu_ok,cpu_ok ok
    class warn_ram warn
    class block,block_disk block
    class start,ram,gpu,vram,disk neutral

Hardware Requirements by Model Version

Model	Parameters	RAM (CPU)	VRAM (GPU)	Disk	Context
TimesFM 3.0 (non-commercial weights)	~330M	≥ 6 GB	≥ 4 GB	~1.3 GB	up to 15,360
TimesFM 2.5 (default)	200M	≥ 4 GB	≥ 2 GB	~0.9 GB	up to 16,384
TimesFM 2.0 (archived)	500M	≥ 16 GB	≥ 8 GB	~2 GB	up to 2,048
TimesFM 1.0 (archived)	200M	≥ 8 GB	≥ 4 GB	~800 MB	up to 2,048

Recommendation: Use TimesFM 2.5 by default (smallest, Apache-2.0, full ForecastConfig control). Use TimesFM 3.0 for multivariate targets or native covariates when the non-commercial license fits the project. The 1.0/2.0 checkpoints need timesfm==1.3.0 and are only worth it for reproducing old results. Measured peak CPU memory for 32 series × 1,024 context (timesfm 3.0.2): ~2.0 GB for 2.5 and ~2.7 GB for 3.0; the thresholds above leave headroom. Check 3.0 with python scripts/check_system.py --model v3.0.

🔧 Installation

Step 1: Verify System (always first)

python scripts/check_system.py

Step 2: Install TimesFM

uv pip install "timesfm[torch]"          # TimesFM 2.5 + 3.0, PyTorch backend
uv pip install "timesfm[torch,xreg]"     # + XReg covariates for TimesFM 2.5 (adds JAX, scikit-learn)
uv pip install "timesfm[mlx]"            # TimesFM 3.0 on Apple silicon without PyTorch
uv pip install "timesfm[flax]"           # TimesFM 2.5 JAX/Flax backend

Step 3: Install PyTorch for Your Hardware

# CPU-only wheels (small download, no CUDA libraries)
uv pip install torch --index-url https://download.pytorch.org/whl/cpu

# NVIDIA GPU: take the CUDA-specific index URL from https://pytorch.org/get-started/locally/
# Apple Silicon: the default PyPI wheel already includes MPS support
uv pip install torch

Step 4: Verify Installation

from importlib.metadata import version
import timesfm  # noqa: F401  (import check)
print(f"TimesFM version: {version('timesfm')}")  # the package defines no __version__
print("Installation OK")

🎯 Quick Start

Minimal Example (5 Lines)

import torch, numpy as np, timesfm

torch.set_float32_matmul_precision("high")

model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
    "google/timesfm-2.5-200m-pytorch"
)
model.compile(timesfm.ForecastConfig(
    max_context=1024, max_horizon=256, normalize_inputs=True,
    use_continuous_quantile_head=True, force_flip_invariance=True,
    infer_is_positive=True, fix_quantile_crossing=True,
))

point, quantiles = model.forecast(horizon=24, inputs=[
    np.sin(np.linspace(0, 20, 200)),  # any 1-D array
])
# point.shape == (1, 24)        — median forecast
# quantiles.shape == (1, 24, 10) — 10th–90th percentile bands

TimesFM 3.0 (multivariate, native covariates)

import numpy as np
from timesfm3 import TimesFM3Forecaster  # non-commercial weights — see license note above

forecaster = TimesFM3Forecaster.from_pretrained("google/timesfm-3.0-pytorch", device="cpu")  # or "cuda"

out = forecaster.predict(np.sin(np.linspace(0, 40, 512)).astype(np.float32),
                         horizon=24, return_quantiles=True)
# out.forecast.shape == (24,)      — median forecast
# out.quantiles.shape == (24, 9)   — q10..q90; index 4 is the median (no mean column)

# Two target channels plus one known-future covariate (context 256, horizon 32)
target = np.stack([np.sin(np.linspace(0, 24, 256)), np.cos(np.linspace(0, 24, 256))]).astype(np.float32)
future_cov = np.sin(np.linspace(0, 30, 256 + 32))[None, :].astype(np.float32)
out = forecaster.predict(target, horizon=32, past_future_covariates=future_cov, return_quantiles=True)
# out.forecast.shape == (2, 32); out.quantiles.shape == (2, 32, 9)

Full parameter list (predict_batch, past_only_covariates, make_positive, …): references/api_reference.md.

Forecast from CSV

import pandas as pd, numpy as np

df = pd.read_csv("monthly_sales.csv", parse_dates=["date"], index_col="date")

# Convert each column to a list of arrays
inputs = [df[col].dropna().values.astype(np.float32) for col in df.columns]

point, quantiles = model.forecast(horizon=12, inputs=inputs)

# Build a results DataFrame
for i, col in enumerate(df.columns):
    last_date = df[col].dropna().index[-1]
    future_dates = pd.date_range(last_date, periods=13, freq="MS")[1:]
    forecast_df = pd.DataFrame({
        "date": future_dates,
        "forecast": point[i],
        "lower_80": quantiles[i, :, 1],  # q10 — lower bound of 80% PI
        "upper_80": quantiles[i, :, 9],  # q90 — upper bound of 80% PI
    })
    print(f"\n--- {col} ---")
    print(forecast_df.to_string(index=False))

Forecast with Covariates (XReg)

TimesFM 2.5 supports exogenous variables through forecast_with_covariates(). It requires timesfm[xreg] and a model compiled with return_backcast=True (otherwise it raises ValueError). TimesFM 3.0 takes covariates directly in predict() (see above).

# Requires: uv pip install "timesfm[torch,xreg]"
model.compile(timesfm.ForecastConfig(
    max_context=1024, max_horizon=256, normalize_inputs=True,
    use_continuous_quantile_head=True, fix_quantile_crossing=True,
    return_backcast=True,
))
point, quantiles = model.forecast_with_covariates(
    inputs=inputs,
    dynamic_numerical_covariates={"price": price_arrays},
    dynamic_categorical_covariates={"holiday": holiday_arrays},
    static_categorical_covariates={"region": region_labels},
    xreg_mode="xreg + timesfm",  # or "timesfm + xreg"
)
# point / quantiles: one array per series — (horizon,) and (horizon, 10)

Covariate Type	Description	Example
dynamic_numerical	Time-varying numeric	price, temperature, promotion spend
dynamic_categorical	Time-varying categorical	holiday flag, day of week
static_numerical	Per-series numeric	store size, account age
static_categorical	Per-series categorical	store type, region, product category

XReg Modes:

• "xreg + timesfm" (default): fit an in-context linear regression on the covariates first, then TimesFM forecasts the regression residuals
• "timesfm + xreg": TimesFM forecasts first, then a linear regression on the covariates fits TimesFM's residuals

See examples/covariates-forecasting/ for a complete example with synthetic retail data.

Anomaly Detection (via Quantile Intervals)

TimesFM does not have built-in anomaly detection, but the quantile forecasts naturally provide prediction intervals that can detect anomalies:

point, q = model.forecast(horizon=H, inputs=[values])

# 80% prediction interval
lower_80 = q[0, :, 1]  # 10th percentile
upper_80 = q[0, :, 9]  # 90th percentile

# Detect anomalies: values outside the 80% CI
actual = test_values  # your holdout data
anomalies = (actual < lower_80) | (actual > upper_80)

# Severity levels
is_warning = (actual < q[0, :, 2]) | (actual > q[0, :, 8])  # outside 60% CI
is_critical = anomalies  # outside 80% CI

Severity	Condition	Interpretation
Normal	Inside 60% CI	Expected behavior
Warning	Outside 60% CI	Unusual but possible
Critical	Outside 80% CI	Statistically rare (< 20% probability)

See examples/anomaly-detection/ for a complete example with visualization.

📊 Output, Config & Workflows

The output structure and full ForecastConfig reference are in references/output_and_config.md.

Quantile layout (TimesFM 2.5): quantile_forecast has shape (batch, horizon, 10). Index 0 is the mean; q10 = index 1, q50 (median) = index 5, q90 = index 9, so the 80% PI is q[:,:,1]–q[:,:,9]. TimesFM 3.0 returns 9 columns (q10–q90) with the median at index 4 — re-check indices when switching models.

Copy-paste workflows (single-series, batch, accuracy evaluation), GPU/memory performance tuning, and integration with statsmodels / matplotlib / EDA are in references/workflows.md.

📚 Scripts

• scripts/check_system.py — mandatory preflight checker; run before first model load. Reports RAM/GPU/disk/Python/install status and a recommended per_core_batch_size.
• scripts/forecast_csv.py — end-to-end CSV forecasting with automatic system check:

  python scripts/forecast_csv.py input.csv --horizon 24 \
      --date-col date --value-cols sales,revenue --output forecasts.csv
  

📖 Reference Documentation

Detailed guides in references/:

File	Contents
references/output_and_config.md	Output shapes, quantile index map, full ForecastConfig parameter reference
references/workflows.md	Single/batch/eval workflows, GPU & memory tuning, statsmodels/matplotlib/EDA integration
references/pitfalls_and_validation.md	Common pitfalls, quality checklist, known mistakes, regression-baseline validation
references/system_requirements.md	Hardware tiers, GPU/CPU selection, memory estimation formulas
references/api_reference.md	Full from_pretrained options, API surface, output shapes
references/data_preparation.md	Input formats, NaN handling, CSV loading, covariate setup

Before declaring any task done, run the quality checklist and review the common pitfalls/mistakes in references/pitfalls_and_validation.md — especially the quantile index off-by-one and infer_is_positive for negative series.

Model Versions

timeline
    accTitle: TimesFM Version History
    accDescr: Timeline of TimesFM model releases showing parameter counts and key improvements.

    section 2024
        TimesFM 1.0 : 200M params, 2K context, JAX only
        TimesFM 2.0 : 500M params, 2K context, PyTorch + JAX
    section 2025
        TimesFM 2.5 : 200M params, 16K context, quantile head, no frequency indicator
    section 2026
        TimesFM 3.0 : ~330M params, 15K context, multivariate + covariates, non-commercial weights

Version	Params	Context	Quantile Head	Frequency Flag	Status
3.0	~330M	15,360	✅ 9 deciles	❌	Latest (non-commercial weights)
2.5	200M	16,384	✅ Continuous (30M)	❌ Removed	Default (Apache-2.0)
2.0	500M	2,048	✅ Fixed buckets	✅ Required	Archived
1.0	200M	2,048	✅ Fixed buckets	✅ Required	Archived

Hugging Face checkpoints:

• google/timesfm-3.0-pytorch (TimesFM 3.0; non-commercial license)
• google/timesfm-2.5-200m-pytorch (default)
• google/timesfm-2.5-200m-flax
• google/timesfm-2.5-200m-transformers (🤗 Transformers port, TimesFm2_5ModelForPrediction)
• google/timesfm-2.0-500m-pytorch (archived)
• google/timesfm-1.0-200m-pytorch (archived)

Resources

• Paper: A Decoder-Only Foundation Model for Time-Series Forecasting (ICML 2024)
• Repository: https://github.com/google-research/timesfm
• Hugging Face: https://huggingface.co/collections/google/timesfm-release-66e4be5fdb56e960c1e482a6
• Google Blog: https://research.google/blog/a-decoder-only-foundation-model-for-time-series-forecasting/
• BigQuery Integration: https://cloud.google.com/bigquery/docs/timesfm-model

Examples

Three reference examples live in examples/; all use the TimesFM 2.5 API (outputs regenerated with timesfm 3.0.2 in 2026-09). Use them as ground truth for correct API usage and expected output shape.

Example	Directory	What It Demonstrates	When To Use It
Global Temperature Forecast	examples/global-temperature/	Basic model.forecast() call, CSV -> PNG -> GIF pipeline, 36-month NOAA context, 60%/80% prediction intervals	Starting point; copy-paste baseline for any univariate series
Anomaly Detection	examples/anomaly-detection/	Two-phase detection: linear detrend + Z-score on context, quantile PI on forecast; 2-panel viz	Any task requiring outlier detection on historical + forecasted data
Covariates (XReg)	examples/covariates-forecasting/	forecast_with_covariates() API (TimesFM 2.5), covariate decomposition, 2x2 shared-axis viz	Retail, energy, or any series with known exogenous drivers

Running the Examples

# Global temperature (TimesFM 2.5)
cd examples/global-temperature && python run_forecast.py && python visualize_forecast.py

# Anomaly detection (TimesFM 2.5)
cd examples/anomaly-detection && python detect_anomalies.py

# Covariates (data + API walkthrough; real inference needs timesfm[torch,xreg])
cd examples/covariates-forecasting && python demo_covariates.py

Expected Outputs

Example	Key output files	Acceptance criteria
global-temperature	output/forecast_output.json, output/forecast_visualization.png	point_forecast has 12 values; PNG shows context + forecast + PI bands
anomaly-detection	output/anomaly_detection.json, output/anomaly_detection.png	Sep 2023 flagged CRITICAL (z >= 3.0); >= 2 forecast CRITICAL from injected anomalies
covariates-forecasting	output/sales_with_covariates.csv, output/covariates_data.png	CSV has 108 rows (3 stores x 36 weeks); stores have distinct price arrays

Quality, Mistakes & Validation

Before declaring any task done, run the post-task quality checklist, review the known mistakes (quantile off-by-one, covariate-horizon coverage, residual-based anomaly detection, etc.), and run the regression-baseline verification snippets — all in references/pitfalls_and_validation.md.

Part of the AlterLab Academic Skills suite.