Machine learningProject record · Sep 2026

Machine Learning & Model Research

A-share adaptive portfolio learning

Related DoubleAdapt/StockMixer, DeePM, and imitation-learning experiments in adapting models to A-share portfolio decisions.

PyTorchDoubleAdaptStockMixerDeePMImitation learningConformal calibrationLightGBM
Why it sits here. Placed by the scope of implementation, the available contribution evidence, and the distinct technical capability it demonstrates.
03 / Somewhere between data and understanding.STUDY IN SPACE

01 / IMPLEMENTATION & CONTRIBUTION

What the work involves

Donald adapted the upstream DeePM codebase for A-share research. Related local DoubleAdapt/StockMixer and imitation-learning implementations explore causal adaptation, feasible teachers, and reliability-based abstention.

Technical depth

Causal support/query adaptation, monthly retraining, market-specific execution constraints, utility labels, feasible teachers, DAgger scaffolding, conformal lower bounds, reliability abstention and walk-forward tests.

The project family

A_Share/doubleadapt_stockmixerA_Share/deepmA_Share/FSD

02 / RESULTS

What came out of it

Modular pipelines support walk-forward portfolio experiments. The documented DeePM study rejected the tested long-only selection thesis and separated market exposure from active selection value.

03 / SUPPORTING EVIDENCE

Follow the source

Implementation notes, project records, and supporting artifacts.

Source context & project scope

Original models and publications remain upstream. The negative DeePM report has universe/survivorship caveats and should be presented as its dated experiment, not a universal impossibility result. No headline financial metrics are promoted.

Monthly DoubleAdapt protocol, causal labels, simulator and promotion layers.

SOURCE · 2026-09-17

Typed PIT contracts, executable simulator, feasible teacher and reliability-gated imitation-learning scaffold.

SOURCE · 2026-09-17

Concrete simulation module, alongside teacher, DAgger, training and reliability modules.

SOURCE · 2026-09-17

Dated adverse result for a tested long-only selection variant.

SOURCE · 2026-09-17

52877f5 attributes the A-share cost-aware and neutralized experiments to donald7.

SOURCE · 2026-09-17
CONTINUE IN MACHINE LEARNING & MODEL RESEARCH

Project Genji: temporal model evaluation