# AlphaAllocator

Bounded news overlays for risk-aware asset allocation. A QRS@NTU research project; NLPCC 2026 Shared Task 4, Track 1: 1st of 10 teams.

- Overview: https://qrsntu.org/projects/alphaallocator/
- How it works: https://qrsntu.org/projects/alphaallocator/details/
- Paper (OpenReview): https://openreview.net/forum?id=u09PzTYNHk

Plain-text summary for AI agents and readers, generated from the evidence file behind the pages; the pages are the full account. This is an academic system, not investment advice.

## Design

AlphaAllocator does not hand the news to one large language model and ask it for the trade. It splits the job into the two questions news can answer, each with its own small model: a relevance model (a 4-layer BGE-family sentence encoder) decides which asset a news item is about, and a tone model (FinBERT) decides whether it is good or bad news. Their product, weighted for recency and certainty and centred across assets, gives one news score per asset. A risk-parity model decides the allocation (in Track 2, a blend of inverse-volatility, momentum and trend), and the news may only tilt it: the whole tilt is capped at 1.0% of the portfolio (1.2% in Track 2), so prices decide and news only nudges.

## Key results

- NLPCC 2026 Shared Task 4 (LLM-based Investment Advisor Agents for Asset Allocation in the Chinese Market), scored by the organizers on a hidden 2026 period.
- Track 1 (11 macro assets): 1st of 10 teams. Sharpe 1.1144, annualised return 13.52%, max drawdown -6.23%.
- Track 2 (16 sector assets): 6th of 13 teams. Sharpe 0.7566, annualised return 15.32%, max drawdown -7.49%.
- Public-2025 replay of Track 1 (243 daily decisions): Sharpe 2.270 with news against 2.198 for the same agent with the news zeroed (return 24.30% against 23.31%).
- ACM ICAIF 2026 Trading Agent Competition (30 U.S.-listed equities): results coming soon.

## Limitations

- The news effect is small and not causal: it is one period, and feeding each day the news from 5 or 21 dates later gives about the same Sharpe (2.271 and 2.271).
- There is no direct-LLM baseline: no agent that asks one large language model for the trade was run, so nothing is claimed about how one would do.
- The Track 2 news overlay did not help: 34.00% return and 12.53% drawdown with news, against 35.05% and 11.26% without.
- The 2025 replay is gross of costs (no orders, commission, slippage or rejected trades) and is not a held-out or preregistered test.

## Not claimed

- a causal effect of news on returns (news alpha)
- that the Track 2 text system beats its no-news reference
- that public 2025 is a held-out or preregistered test (first-inspection chronology is not recoverable)
- net-of-cost performance in the 2025 replay (it is a gross target-weight replay)
- that the configured 0.035 overlay turnover guard is active (it is not, in the submitted call path)
- the separate value of relevance, tone, recency, event similarity or uncertainty
- investment advice

## Status

- NLPCC 2026 Shared Task 4: poster and oral; the oral is at Evaluation Workshop I, 5 November 2026.
- ACM ICAIF 2026 Trading Agent Competition: coming soon.

## A question to start with

Explain the design to a newcomer, then give the three strongest criticisms and one experiment that would settle each.
