[Q&A] Clarification on Multivariate Time Series Setup in Chronos 2 (covariates vs multi-target design)

[YiqiaoXu-Joe]

Hi Chronos team,

I’m trying to better understand how Chronos 2 handles multivariate time series, especially regarding the distinction between dependent vs independent time series and the correct way to structure data when multiple related signals exist.

Below is my current understanding — please correct me if any of this is inaccurate:

---

My Understanding So Far

1. Dependent time series (single target, no cross-learning)

• There is one item_id and one forecasting target.
• Other time series are treated as covariates for this single task.
• The model does not learn cross-series patterns across multiple item_ids.

2. Independent time series (multiple items, cross-learning enabled)

• Multiple forecasting targets/tasks, each assigned a unique item_id (or group_id).
• When predict_batch_jointly = True:
  • All time series in the batch share the same group id, enabling cross-series learning.
  • Chronos treats these as part of the same “family” of related series.

---

Example Use Case

Suppose I want to forecast the 10-year treasury rate.
I have access to the following daily signals:

• Multiple stock prices
• 3-year treasury rate
• Weather information
• (Potentially) other macroeconomic indicators

---

My Question

What is the recommended way to structure this for Chronos 2?

Option A — Treat everything as covariates (single target task)

• item_id = only the 10-year bond
• All other series (stocks, 3-year bond, weather) become covariates
• This matches the “dependent time series” setup

Option B — Multitarget / multi-item setup

• item_id includes:
  • 10-year rate
  • 3-year rate
  • Possibly multiple stock tickers
• Weather and other contextual features become covariates
• predict_batch_jointly = True to enable cross-learning across related financial series

---

Which option better matches the intended Chronos 2 workflow?

My intuition is that Option B is more aligned with Chronos 2’s design — i.e., allowing the model to learn shared temporal patterns across multiple financial instruments. But I want to confirm whether:

• The model actually benefits from using multiple stock/bond series as separate item_ids
• Or whether it is better practice to keep the problem strictly single-target and treat all other signals as covariates
• Are there performance or stability trade-offs between:
  • High-dimensional covariates
  • vs multi-target shared learning?

---

Any clarification or recommended pattern for financial forecasting scenarios would be greatly appreciated!

Thanks in advance!