Timesfm

TimesFM

Overview

TimesFM is a 200M-parameter foundation model by Google Research, pretrained on 100 billion real-world time points. It performs zero-shot forecasting across domains — no fine-tuning required. Feed it historical data, get predictions immediately.

Instructions

Installation

pip install timesfm

Basic Forecasting

import timesfm
import numpy as np

# Initialize model
tfm = timesfm.TimesFm(
    hparams=timesfm.TimesFmHparams(
        per_core_batch_size=32,
        horizon_len=30,
    ),
    checkpoint=timesfm.TimesFmCheckpoint(
        huggingface_repo_id="google/timesfm-2.0-200m-pytorch",
    ),
)

# Your historical data (e.g., daily sales for 1 year)
history = np.array([120, 135, 128, 142, 155, 148, 160, ...])

# Forecast next 30 days
forecasts = tfm.forecast([history], freq=[1])
predictions = forecasts[0]  # shape: (30,)

Frequency Parameter

Set freq to match your data granularity:

• 0: High frequency (seconds/minutes)
• 1: Daily
• 2: Weekly/Monthly

Multi-Series Forecasting

# Forecast multiple product categories at once
series = [sales_electronics, sales_clothing, sales_food]
forecasts = tfm.forecast(series, freq=[1, 1, 1])
# Returns list of 3 forecast arrays

Examples

Example 1: Demand forecasting

Input: 365 days of daily product sales data. Output: 30-day forecast with the model capturing weekly seasonality and growth trend automatically.

history = load_csv("daily_sales.csv")["quantity"].values
forecast = tfm.forecast([history], freq=[1])[0]
print(f"Next 7 days: {forecast[:7]}")
# Next 7 days: [182, 175, 190, 168, 195, 201, 178]

Example 2: Server metrics anomaly detection

Input: 720 hours (30 days) of CPU utilization. Output: Forecast next 24 hours. Flag if actual exceeds forecast by 2x standard deviation.

cpu_history = get_metrics("cpu_percent", days=30)
forecast = tfm.forecast([cpu_history], freq=[0])[0]
threshold = forecast.mean() + 2 * forecast.std()

Guidelines

• Provide at least 3x the forecast horizon as history (forecasting 30 days? give 90+ days history)
• TimesFM works best on data with clear patterns (seasonality, trends)
• For noisy data, smooth with rolling average before feeding to the model
• Compare against a naive baseline (last period's values) to validate improvement
• The model runs on CPU; GPU speeds up batch processing of many series