Zero-Shot Time Series Forecasting using Chronos-2

---

Overview

This project demonstrates zero-shot forecasting, where a pre-trained time series foundation model (Chronos-2) is used to generate forecasts without any training or parameter tuning on the dataset.

Unlike traditional econometric or ML models, this approach directly leverages learned representations of time series patterns.

---

What is Zero-Shot Forecasting?

Zero-shot forecasting means:

The model does not learn from your dataset explicitly, but instead uses prior knowledge gained during large-scale pretraining.

In this project:

• Input → historical inflation data
• Output → future forecast (next 4 months)
• No model fitting step involved

---

About Chronos-2

Chronos-2 is a Transformer-based time series foundation model developed by Amazon.

• ~120M parameters
• Encoder-only architecture
• Supports:
  • Univariate forecasting
  • Multivariate forecasting
  • Covariate-aware predictions

The model exposes a Pandas-based API, making it easy to integrate into data workflows.

---

How Chronos Works Internally

Unlike ARIMA or traditional econometric models, Chronos does not model values directly.

Instead, it:

1. Normalizes the time series
2. Discretizes values into tokens
3. Converts the series into a sequence of tokens
4. Uses a Transformer to predict future tokens

Essentially, it treats time series like a language modeling problem, similar to how GPT predicts the next word.

---

Forecast Output

🔹 Historical vs Forecast Plot

Installation

pip install chronos-forecasting pandas matplotlib

Usage

from chronos import Chronos2Pipeline

pipeline = Chronos2Pipeline.from_pretrained("amazon/chronos-2")

pred_df = pipeline.predict_df(
    df,
    prediction_length=4
)

Dataset Used

Monthly WPI Inflation Data (attached in repository) (Taken from the database on Indian Economy RBI estimation) Period: March 2025 – February 2026