This PR adds CAST, an ICDM 2026 accepted paper with code, to the **Market Microstructure and Regime Detection** section.

[FanBroWell]

What is this resource?

CAST is an ICDM 2026 accepted research paper with open-source code for cross-asset stock trading and drawdown control.

Why does it fit this list?

I added it to the Market Microstructure and Regime Detection section because CAST models latent market states and cross-asset interactions, and uses these estimated states for drawdown-aware trading. It has a research origin, an accepted conference paper, and publicly available code.

Paper: https://arxiv.org/abs/2609.14205
Code: https://github.com/FanBroWell/CAST

Affiliation disclosure (required)

I am the first author of CAST.

Checklist

• I have read CONTRIBUTING.md
• One resource per PR, added to the most relevant existing section
• Entry matches the section's format: - [Name](url) - Description ending with period.
• I searched the README — this is not a duplicate
• Affiliation disclosed above (or "None")

[leoncuhk]

Read the paper and the code. It's a fit: the method is different from anything on the list, and ICDM acceptance covers the traction gap for a month-old repo. Three things before I merge.

Section. This isn't microstructure, and it isn't regime detection the way the other two entries in that section use the term — no order book data, no discrete state switching. CoKF is online latent-state estimation on an IRW basis, and the decision layer is MPC. Rather than stretch that section, I'd open a new one under Research Papers:

### Forecasting and Trading Systems

Move the entry there. CONTRIBUTING tells contributors not to create sections on their own; this one's my call, so go ahead.

Description. The current wording drops the controller, which your own ablation says matters most: at λ=0 drawdown reaches ~80% and the account ends at $200. It also doesn't say what was tested. Suggested:

• CAST: A Cross-Asset State-Space Trading System for Drawdown Control in Stock Markets by Yu Peng, Matloob Khushi, and Josiah Poon - Cross-asset collaborative Kalman filter feeding a model-predictive controller that uses forecast uncertainty as an explicit risk penalty; four 30-stock panels (NASDAQ, CSI300, TPX100, a global basket) over a 2010-2025 frictionless backtest. ICDM 2026, code.

License. FanBroWell/CAST has no LICENSE file, so as it stands the code is all-rights-reserved and readers can't build on it. MIT or Apache-2.0, either works. This one blocks the merge.

Separately, not a blocker: §V-B says λ is fixed before evaluation, but experiments/m30_main_results.py sweeps λ over {0.05, 0.1, 0.3, 0.6} and keeps the best test-window Sharpe per (dataset, method), and the README calls that output the paper's main table (it says Table I; the comparison is Table II). If λ was picked on the pre-2010 window and the sweep is the Fig. 6 sensitivity study, worth saying so in the README — otherwise anyone reproducing it will read that as test-set selection.

Worth noting for readers who assume otherwise: execution is next close after the decision close, and ρ, σ_v, σ_w come only from pre-2010 data. The backtest loop is clean.

Push those and I'll merge.

[FanBroWell]

Thank you for the careful review of both the paper and the code. Your suggestions were very helpful.

I have pushed the requested updates: CAST has been moved to a new Forecasting and Trading Systems section under Research Papers, and the description follows your suggested wording.

I also updated the CAST repository based on your suggestions, including adding an MIT LICENSE file and clarifying the README around the lambda sweep, execution timing, and parameter calibration.

Thank you again for your time and constructive feedback.

[leoncuhk]

Apache-2.0 works, and the README now says what the best-λ CSV is and isn't. I tidied the entry to match the section format (hyphen separator, single space, plain code link) and noted the licence. Merging.

[FanBroWell]

Thanks for the tidy-up and the merge!

Wishing you all the best with your research.