Foundation Models

Vectrix provides optional wrappers for state-of-the-art pretrained forecasting models. These models perform zero-shot forecasting — no training required on your specific data.

Note: Foundation model wrappers require additional packages: pip install "vectrix[foundation]"

Amazon Chronos-2

Chronos is a family of pretrained probabilistic time series models by Amazon. They tokenize time series into bins and use transformer architectures for forecasting.

from vectrix import ChronosForecaster

model = ChronosForecaster(
    modelId="amazon/chronos-bolt-small",
    device="cpu",
)

model.fit(y)  # stores context only — no training
predictions, lower, upper = model.predict(steps=12)

Available Models

Quantile Forecasts

import numpy as np

quantiles = model.predictQuantiles(
    steps=12,
    quantileLevels=[0.1, 0.5, 0.9],
)
# shape: (3, 12) — one row per quantile level

Batch Prediction

Forecast multiple series at once

series = [y1, y2, y3]
results = model.predictBatch(series, steps=12)
# returns list of (predictions, lower, upper) tuples

Google TimesFM 2.5

TimesFM is Google’s foundation model for time series forecasting, supporting up to 2048 context length.

from vectrix import TimesFMForecaster

model = TimesFMForecaster(
    modelId="google/timesfm-2.5-200m-pytorch",
)

model.fit(y)
predictions, lower, upper = model.predict(steps=12)

With Covariates

TimesFM supports exogenous variables (covariates)

predictions, lower, upper = model.predictWithCovariates(
    steps=12,
    dynamicNumerical=future_numerical_features,  # (steps, n_features)
    dynamicCategorical=future_categorical_features,
)

Checking Availability

from vectrix import CHRONOS_AVAILABLE, TIMESFM_AVAILABLE

if CHRONOS_AVAILABLE:
    model = ChronosForecaster()
else:
    print("Install: pip install 'vectrix[foundation]'")

When to Use Foundation Models