•
Amazon Chronos-2 (pip install chronos-forecasting, HF amazon/chronos-2) is another zero-shot forecasting path for after-work notebooks—contrast with the Hub’s TimesFM-3 note. Not investment advice.
If you already skimmed our Hub note on TimesFM-3, you have one zero-shot time-series path from Google Research. This note is a different route for after-work notebook trials: Amazon’s Chronos-2, available as Hugging Face model amazon/chronos-2 with the Python package chronos-forecasting (pip install chronos-forecasting).
This is educational commentary for employed people studying after work. It is not investment advice. SoloQuant Academy does not host Chronos-2.
According to the Chronos-2 model card on Hugging Face:
Treat capability claims and any published benchmark rankings as the authors’ statements on the model card and paper—not as numbers we invent here.
Primary sources:
| Angle | TimesFM-3 (Hub note) | Chronos-2 (this note) |
|---|---|---|
| Vendor path | Google Research TimesFM family | Amazon Chronos family |
| Hub entry point | Introducing TimesFM-3 | This post |
| Local Python package (card / repo) | Follow Google / HF instructions for TimesFM | pip install chronos-forecasting then Chronos2Pipeline |
| HF id (card) | See TimesFM Hub note / Google HF | amazon/chronos-2 |
Both are foundation-model style zero-shot forecasting tools. Neither replaces clean timestamps, honest holdouts, or reading the license and compute requirements before you ship anything at work.
The model card’s local path is straightforward for a weeknight venv:
pip install "chronos-forecasting>=2.0"
Illustrative sketch adapted from the official Chronos-2 card (CPU or CUDA depending on your laptop):
# Verify against the live Chronos-2 model card on your machine
import pandas as pd
from chronos import Chronos2Pipeline
pipeline = Chronos2Pipeline.from_pretrained(
"amazon/chronos-2",
device_map="cpu", # use "cuda" if you have a GPU evening setup
)
# Tiny synthetic panel — replace with your own tidy series later
context_df = pd.DataFrame(
{
"id": ["demo"] * 48,
"timestamp": pd.date_range("2024-01-01", periods=48, freq="h"),
"target": [10 + (i % 12) * 0.5 for i in range(48)],
}
)
pred_df = pipeline.predict_df(
context_df,
prediction_length=12,
quantile_levels=[0.1, 0.5, 0.9],
id_column="id",
timestamp_column="timestamp",
target="target",
)
print(pred_df.head())
The card also shows covariate-aware predict_df with future_df on public electricity-price Parquet examples. Start smaller than production data on night one.
If Python packaging or DataFrames still feel shaky, fix that first—foundation models will not rescue a messy environment.
Chronos-2 is another zero-shot forecasting path you can try after work via chronos-forecasting and amazon/chronos-2, alongside the TimesFM-3 story already on the Hub. Compare routes; do not treat either card as a trading signal. Not investment advice.
Next step: strengthen Python and data hygiene with Intro to Python Programming—the only enrollment CTA in this note.