Research question
How can language-model-derived event signals inform macro-asset allocation and sector rotation while keeping data visibility, allocation rules, and risk controls auditable?
Context
This project concerns NLPCC 2026 Shared Task 4, with an emphasis on explicit allocation controls and reproducible evaluation.
Method
The design separates text interpretation from portfolio control. News can be denoised, aggregated, and represented as structured event signals; portfolio weights still pass through explicit allocation, turnover, and risk-control logic.
Validation
The evidence ladder uses a tuning period followed by a frozen test period, no-news and no-LLM baselines, transaction checks, trace logs, and Sharpe, drawdown, turnover, and execution diagnostics.
Current result
The supported result is methodological: the system produces inspectable daily traces and comparable baseline runs. No public performance conclusion is stated.
Limitations
Language-model outputs can be unstable, event labels can be underspecified, and apparently strong allocation results may disappear after realistic costs or stricter data-visibility checks.
Public artifact
The validation and trace workflow describes baseline comparisons, held-out evaluation and allocation diagnostics.
Next question
Which parts of the allocation outcome remain after removing news, removing language-model assistance, and freezing every decision rule before evaluation?