Introducing CrossLearn: Extractor-First Reinforcement Learning with Chronos-2

After multiple iterations on time-series reinforcement learning setups, a pattern became clear:

Most of the complexity is misplaced.

We spend time tuning policies and algorithms, while the real bottleneck is often the quality of the representation. Poor features limit everything downstream.

Time-series data, especially noisy real-world signals, already contains structure:

• seasonality
• trends
• latent relationships

Ignoring this and forcing an RL agent to rediscover it from raw inputs is inefficient. It wastes both signal and compute.

CrossLearn is built around that observation.

It shifts the focus from

“Which RL algorithm should I use?”

to:

“How do I feed better representations into whatever RL algorithm I already trust?”

Core Idea: Extractor-First RL

Image by Author: CrossLearn

At a high level, CrossLearn separates:

• representation learning
• decision learning

Instead of feeding raw time-series directly into an RL agent, the pipeline becomes:

1. A pretrained time-series model extracts meaningful features
2. The RL agent operates on these features

This changes where the complexity lives.

The agent no longer needs to learn temporal structure from scratch. That burden is shifted to a model that was trained specifically for that purpose.

The result:

• Faster learning
• Better generalization
• More stable training

Why Chronos-2?

At the center of this approach is Chronos-2, a pretrained model designed to capture temporal patterns across domains.

Within CrossLearn, Chronos-2 acts as a feature encoder:

• converts raw sequences into dense embeddings
• encodes trend, seasonality, and temporal context
• reduces noise while preserving signal

These embeddings are structured and compressed representations of the original data.

Instead of learning from raw OHLCV windows or sensor streams, the agent learns from a representation that already encodes temporal structure.

Architecture Overview

CrossLearn provides plug-and-play feature extractors compatible with Stable-Baselines3 via the BaseFeaturesExtractor interface.

Adriano Marabuco de A. Lima et al.

The design is intentionally modular:

• extractors can be swapped without changing the RL pipeline
• multiple algorithms can use the same representation
• environments remain unchanged

There is no heavy abstraction layer. The goal is not to hide complexity, but to isolate it. You can experiment with representation quality independently of policy optimization.

What This Solves

1. Sample Inefficiency

RL typically needs large amounts of interaction data.

By starting with pretrained representations, CrossLearn reduces the learning burden. The agent does not need to “figure out” the structure of the data from scratch.

2. Instability in Training

Raw time-series is noisy. RL is sensitive to noise.

Extracted features are smoother and more consistent. This stabilizes gradients and improves convergence.

3. Weak Generalization

Models trained on raw signals often overfit to specific patterns.

Foundation model embeddings capture broader structure. This improves transfer across datasets and environments.

Use Cases

CrossLearn applies anywhere sequential data meets decision-making:

• financial trading strategies
• agricultural yield optimization
• demand forecasting with action policies
• energy load balancing
• industrial process control

If your system involves:

• time-series inputs
• decisions over time
• delayed rewards

This approach is relevant.

Why This Matters

Foundation models have already reshaped:

• natural language processing
• computer vision

Time-series modeling is following the same trajectory.

The next step is not just better predictions, but better decision-making systems built on top of those representations.

CrossLearn is a step in that direction. It connects pretrained temporal representations with reinforcement learning in a way that is modular and practical.

What’s Next

This is an early release with a focused scope.

Areas under exploration:

• additional extractors (vision models, alternative time-series foundations)
• multimodal inputs
• benchmarking across domains

Get Started

Explore the repository:

[embed]GitHub - cpohagwu/crosslearn: Reusable representation extractors for reinforcement learning across… Reusable representation extractors for reinforcement learning across vectors, images, and Chronos-backed time-series…github.com

The repository includes five Colab notebooks covering:

• classic control with REINFORCE
• Atari with CNN extractors
• trading with Chronos-2 (REINFORCE and PPO)

Try it. Break it. Extend it.

If you are working on time-series RL, this approach will likely change how you think about the problem.

Let me know what you think in the comments. If pretrained models already define the backbone of modern NLP and vision systems, will time-series foundation models become the default input layer for RL?

A message from our Founder

Hey, Sunil here. I wanted to take a moment to thank you for reading until the end and for being a part of this community. Did you know that our team run these publications as a volunteer effort to over 3.5m monthly readers? We don’t receive any funding, we do this to support the community.

If you want to show some love, please take a moment to follow me on LinkedIn, TikTok, Instagram. You can also subscribe to our weekly newsletter. And before you go, don’t forget to clap and follow the writer️!