Probabilistic time series forecasting helps data scientists and analysts predict future values with uncertainty estimates. To leverage pre-trained language models for accurate time series predictions, use Chronos.

The power of Chronos lies in its ability to generate accurate forecasts right out of the box, eliminating the need for extensive model training or fine-tuning in many cases.

Here’s a quick example of using Chronos for forecasting:

import pandas as pd
import torch
from chronos import ChronosPipeline

pipeline = ChronosPipeline.from_pretrained(
    "amazon/chronos-t5-small", torch_dtype=torch.bfloat16
)

df = pd.read_csv(
    "https://raw.githubusercontent.com/AileenNielsen/TimeSeriesAnalysisWithPython/master/data/AirPassengers.csv"
)

# forecast shape: [num_series, num_samples, prediction_length]
forecast = pipeline.predict(
    context=torch.tensor(df["#Passengers"]), prediction_length=12, num_samples=20
)

This code loads the air passenger data, uses the pretrained Chronos model to generate 20 possible forecasts for the next 12 months, and then calculates the median forecast along with a 90% prediction interval.

The Chronos model internally transforms the numerical time series into tokens, processes them through a language model architecture, and then converts the output back into numerical forecasts. This approach allows it to capture complex patterns and dependencies in the data, potentially outperforming traditional forecasting methods, especially in zero-shot scenarios where the model hasn’t been specifically trained on the target time series.

Link to Chronos.