01 DAILY RESEARCH SIGNALS / 2026-10-20
A curated stream of quantitative research, open-source projects, and AI engineering advances—screened, ranked, and structured for faster research decisions.
All sources checked. Pipeline operating normally.
02 CURATED INDEX
Browse by content type or search titles and tags.
INDEXED 15
SELECTED SIGNALS 03
This skill provides a multi-phase security audit workflow for coding agents, producing independently verifiable and machine-readable findings to enhance audit automation and trustworthiness.
vLLM appeared on GitHub Trending Daily and describes itself as a high-throughput, memory-efficient inference and serving engine for large language models. Its novelty lies in system-level optimization targeting both throughput and memory efficiency, which is directly relevant to developers deploying LLM services.
A project trending on GitHub that argues AI agents should think like the "laziest senior dev": the best code is code you never wrote. Its novelty is framing reduced code output as an agent goal, useful for those watching AI coding tool cost and complexity.
SELECTED SIGNALS 11
This study proposes RegimeRank, a model that formulates sector rotation as ranking sectors by expected future risk-adjusted active performance, and enhances ranking performance while maintaining interpretability through expert-defined temporal features, macro-regime conditioning, and cross-sector dependency modeling. Its novelty lies in integrating regime awareness with attention mechanisms to provide a diagnosable ranking tool for quantitative portfolio management.
TimeART proposes an agentic framework that integrates powerful tool-based analysis with LLM reasoning for time series question answering (TSQA). It also introduces a 100k expert trajectory corpus, TimeToolBench, and a four-stage training strategy to enhance tool-use and generalization of LLM-based time series reasoning models (TSRMs).
This paper introduces the branched signature model, generalizing classical geometric rough paths based on branched rough path theory, and establishes a universal approximation theorem. By constructing an extension map, branched signatures are realized as classical geometric signatures of extended paths, providing a computational scheme for data-driven modeling such as rough volatility calibration and optimal execution.
Proposes MOIRA, a framework that formulates pair trading as hierarchical reinforcement learning and uses large language models to parameterize both high-level pair selection and low-level execution, optimized via prompt updates and textual feedback without gradient fine-tuning.
The study proposes CEEMDAN-SE-SSA-VMGL to address high-frequency mode mixing and unmatched volatility heterogeneity in existing CEEMDAN hybrids, combining sample-entropy clustering, singular spectrum analysis denoising, and a differentiable GARCH layer that modulates the LSTM forget gate; it reports 9%–24% lower forecasting error than the best benchmark across five US and Chinese financial series.
This study reconstructs order-level data from Polymarket's thin order books and trains cooperative multi-agent market-making policies; its value lies in highlighting the critical role of reward design in decomposed market making and the evidentiary boundary of no positive portfolio value.
The study introduces a game-theoretic framework that lets large language models learn dynamic counter-strategies through in-context experience without fine-tuning, using Nash equilibrium as an evaluation benchmark. Its novelty lies in modeling scammer–baiter interactions as a two-player game and releasing a dialogue dataset with turn-level strategy labels, offering a quantifiable path for proactive anti-scam strategy optimization.
The study introduces LOBGrad, a conditional DDPM framework combining an autoregressive RNN encoder with a diffusion generator for probabilistic limit order book forecasting, establishes an excess-risk oracle inequality for the one-step DDPM noise-prediction objective, and shows improvements over an encoder-free ablation on BTC/USDT, GOOGL, and META.
NIOCIS compresses tool-use agent logs into a loop-collapsed action graph and uses a small model to emit a typed option distribution in one pass, replacing costly per-hop LLM reasoning with very low latency.
The study examines whether four AI trading agents followed their strategies and risk rules across 2,627 closed positions and offers a reusable agent evaluation toolkit to distinguish rule-following from luck before granting autonomy or capital.
This study evaluates the reliability of TimeGAN-generated scenarios in Rachev-ratio portfolio allocation using 12 ETFs, comparing historical, replacement, and hybrid policies; synthetic scenarios understate volatility and tail losses and distort dependence, with return differences not statistically significant.
SELECTED SIGNALS 01
The course addresses "context rot": as agents run longer, repeated, contradictory, or stale information accumulates in their context and model performance degrades. Its value lies in building agent harnesses from scratch in Python and treating retrieval, context management, and memory as explicit engineering modules rather than relying only on larger context windows.