In briefInterpretable Equity Portfolio Allocation and Predictive Rebalancing, Tested on Real and Simulated Markets
SummaryThe paper presents one system for choosing stocks, deciding when to trade, and testing strategy robustness on 373 US equities, calibrated on 2014–2024 and traded from January 2025 to September 2026. It combines stock-selection rules, several weighting rules, and constrained share-count allocation, with a model-predictive controller that re-plans at set triggers. The authors also test the system on simulated markets built from hidden Markov generators to probe deeper declines.
In briefTensorLOB: A Structure-Preserving Tensor Model for Limit Order Book Forecasting
SummaryThe paper proposes TensorLOB, a model that treats limit order book data as a multiway tensor across time, side, price level, and order attribute. It extracts features with mode-wise tensor contractions and predicts with low-rank class-weight tensors, keeping each mode's role distinct. On the FI-2010 benchmark, TensorLOB achieves the highest accuracy, recall, and F1 score across all four evaluated horizons.
In briefHerding or Racing? How Orders from Different Wallets Move Together on Blockchain Exchanges
SummaryThe paper examines why trade directions are correlated across wallets on decentralized exchanges. Prior work attributed nearly all such cross-wallet correlation to herding, but the authors show this share mixes the chance that consecutive orders come from one account with a cross-account to same-account correlation ratio κ. On Uniswap v3 USDC-WETH, κ is only 0.05 to 0.10, while on an on-chain limit order book it is 0.72 to 0.76, with the co-movement concentrated in orders a few blocks apart. The authors attribute this to traders racing to act on centralized-exchange prices.
Yichong Bai (The Hong Kong Polytechnic University), Dingyue Kan (The Hong Kong Polytechnic University)
In briefBeyond Binary Detection: Type-Aware Hallucination Analysis in Financial QA
SummaryThis paper studies hallucination detection for question answering over SEC financial filings, using the PHANTOM long-context benchmark. The authors generate hallucinations in nine financially motivated categories and evaluate an LLM-based detector, which reaches 80.2% binary accuracy but only 44.7% subtype accuracy. They propose a two-stage hierarchical detector that first decides whether an answer is hallucinated and then assigns its subtype, raising binary accuracy to 95.6% and subtype accuracy to 56.1%.
In briefScoring party manifestos' environmental stances with a panel of language models
SummaryThe paper proposes a panel of three large language models from independent providers that scores each sentence of a party manifesto on environmental protection and climate change, then averages the scores to the manifesto level. It applies the method to 41 manifestos from the 2024 European Parliament election and validates it against the 2024 Chapel Hill Expert Survey for 25 parties. The authors present it as a low-cost, transparent complement to expert and hand-coded approaches, and release the prompt, code, and results.
