Skip to main content

Research & Evidence

Explore the research AlphaIntel draws on. These papers are peer-reviewed; our implementation of them is ours, and we describe it openly.

Papers are shown in their language of publication, English. Titles, abstracts and keywords are not translated: a translated scientific abstract stops being quotable as-is.

Swipe to see all filters

Advanced Financial Reasoning at Scale: LLMs on CFA Level III

P. Shetty et al.·2025·LLM ReasoningEN

What this means for traders

Top AI models now score above the estimated passing threshold on CFA Level III mock exams, the final step of the professional analyst certification, though the authors caution this is not the same as passing the real exam.

Abstract EN

A comprehensive evaluation of 23 state-of-the-art LLMs on CFA Level III mock exam questions (multiple-choice and essays). OpenAI o4-mini achieved a composite score of 79.1%, and Gemini 2.5 Flash reached 77.3%, above the estimated 63% passing threshold, demonstrating strong financial reasoning capabilities.

LLMsCFAReasoningEvaluation

Sentiment Trading with Large Language Models

Kirtac, K. & Germano, G.·2024·Sentiment AnalysisEN

What this means for traders

In this backtest on U.S. news, the best LLM (OPT) produced a Sharpe ratio about 2.5 times that of the traditional dictionary method (3.05 vs 1.23); BERT and FinBERT also beat the dictionary, by a smaller margin.

Abstract EN

This study compares a dictionary-based method with LLMs for sentiment analysis of 965,375 U.S. financial news articles (2010-2023). In a backtest, a long-short strategy based on a fine-tuned OPT model (2.7B parameters) generated a Sharpe ratio of 3.05 after 10 bps transaction costs, significantly outperforming the traditional Loughran-McDonald dictionary (Sharpe 1.23).

Sentiment AnalysisLLMsTrading StrategiesAlpha Generation

FinChain: A Symbolic Benchmark for Verifiable Chain-of-Thought Financial Reasoning

Tang, Y. et al.·2025·LLM ReasoningEN

What this means for traders

AI models that show their reasoning step-by-step make far fewer logical errors in financial analysis.

Abstract EN

Introduces ChainEval to measure the quality of financial reasoning. Shows that Chain-of-Thought (CoT) prompting significantly reduces logic errors and that reasoning models correlate strongly with expert human judgment.

Chain-of-ThoughtBenchmarkReasoningVerification

TradingAgents: Multi-Agents LLM Financial Trading Framework

Xiao, Y. et al.·2024·Multi-Agent SystemsEN

What this means for traders

Having AI agents debate each other before trading beat simple rule-based strategies on returns and risk-adjusted returns in a short 3-month backtest; the paper does not measure errors or compare against a single AI model.

Abstract EN

Explores the debate mechanism between opposing agents (bullish and bearish researchers, plus a risk team). In a 3-month backtest (January to March 2024) on AAPL, GOOGL and AMZN, the framework beat five rule-based strategies (Buy and Hold, MACD, KDJ+RSI, ZMR, SMA) on cumulative return and Sharpe ratio, for example 26.62% cumulative return and a Sharpe of 8.21 on AAPL.

Multi-AgentDebateRisk ManagementTrading Framework

Single-agent or Multi-agent Systems? Why Not Both?

Gao, M. et al.·2025·System ArchitectureEN

What this means for traders

Multi-agent setups are not automatically more accurate: with today's stronger models a single agent often matches them at a fraction of the cost, so the best design escalates to several agents only when a single one falls short.

Abstract EN

Comparative analysis of multi-agent (MAS) and single-agent (SAS) LLM systems across 15 tasks. MAS can solve complex tasks a single agent fails, but its accuracy edge shrinks as frontier models improve, at a much higher cost. A hybrid design that tries a single agent first and escalates to MAS improves accuracy by 1.1% to 12% while cutting deployment costs by up to 20%.

Multi-AgentSingle-AgentArchitectureEfficiency

Deep Reinforcement Learning for Automated Stock Trading

Yang, H. et al.·2025·Machine LearningEN

What this means for traders

Switching between several reinforcement learning algorithms based on recent risk-adjusted performance gave a better Sharpe ratio and a smaller drawdown than any single algorithm in this backtest, though not the highest raw return.

Abstract EN

Demonstrates that an ensemble of RL algorithms (PPO, A2C, DDPG), which each quarter trades with the agent showing the best recent Sharpe ratio, adapts to changing market conditions. In a backtest on the 30 Dow Jones stocks (January 2016 to May 2020), it reached a Sharpe ratio of 1.30, above each individual algorithm (1.12 at best) and the Dow Jones index (0.47).

Reinforcement LearningEnsemble MethodsAutomated TradingDDPGPPO

Optimal profit-making strategies in stock market with algorithmic trading

Wang, H., Xie, D.·2024·Machine LearningEN

What this means for traders

Classical ML methods like SVMs can still produce excess returns: in a 2.5-year backtest on the CSI 300, an SVM strategy stayed positive while the index fell. The authors warn that the result depends on the backtest window chosen.

Abstract EN

A study of the CSI 300 index (data 2006-2023, models trained on 2006-2020) in which a Support Vector Machine (SVM) timing strategy beat the index by 60.52% cumulatively in a January 2021 to June 2023 backtest, while the index fell 30.45%. That is a 30.27% total return for the strategy (11.69% annualized).

SVMAlgorithmic TradingLong-term AnalysisRobustness

Reinforcement Learning for Deep Portfolio Optimization

Yan, R. et al.·2024·Risk ManagementEN

What this means for traders

Baking risk constraints directly into the AI reward function produces portfolios that maximize returns while respecting hard drawdown limits.

Abstract EN

Integrates Modern Portfolio Theory constraints directly into the RL reward function (Deep Portfolio Optimization). Maximizes portfolio value while strictly adhering to risk constraints.

Portfolio OptimizationReinforcement LearningRisk ManagementMPT

Modeling News Interactions and Influence for Financial Market Prediction

Wang, M., Cohen, S. B., Ma, T.·2024·NLPEN

What this means for traders

Adding news text to price data raised risk-adjusted returns over a price-only version of the same model in historical S&P 500 tests, but only when news items were weighted by influence. Naively averaging news did worse than using no news.

Abstract EN

Shows that fusing news text with price data (FININ model) lifts the daily Sharpe ratio by 0.429 over the best competing model on the S&P 500 (0.341 on the NASDAQ 100), in historical tests over 2003-2018.

MultimodalNLPMarket PredictionNews Analysis

Benchmarking LLMs for Target-Based Financial Sentiment Analysis

CLiC-it·2025·NLPEN

What this means for traders

General-purpose LLMs now outperform finance-specific models at reading market sentiment, with no fine-tuning needed.

Abstract EN

Research indicating that generative models (like GPT-4, DeepSeek) now outperform specialized older models (FinBERT) in zero-shot sentiment analysis tasks.

Sentiment AnalysisGenerative AILLMsBenchmarking

Dynamic Stop Loss Strategy with Deep Reinforcement Learning

Anders, M. et al.·2024·Risk ManagementEN

What this means for traders

AI-driven stop-losses that adapt to volatility outperform fixed percentage rules in both profit and drawdown control.

Abstract EN

Shows that RL agents can learn optimal dynamic stop-loss policies that adapt to market volatility, significantly improving PnL and reducing maximum drawdown compared to static rules.

Stop-LossReinforcement LearningRisk ControlDynamic Strategy

Alpha-GPT: Human-AI Interactive Alpha Mining for Quantitative Investment

Saizhuo Wang, Hang Yuan, Leon Zhou, Lionel M. Ni, Heung-Yeung Shum, Jian Guo·2023·Quantitative FinanceEN

What this means for traders

Combining human domain knowledge with LLM creativity produces better alpha signals than either approach alone.

Abstract EN

Introduces a new alpha mining paradigm by introducing human-AI interaction and a novel prompt engineering algorithmic framework leveraging large language models. Alpha-GPT provides a heuristic way to understand quant researchers' ideas and outputs creative, insightful, and effective alphas.

Alpha MiningHuman-AI InteractionQuantitative FinancePrompt Engineering

LLMFactor: Extracting Profitable Factors through Prompts for Explainable Stock Movement Prediction

Meiyun Wang, Kiyoshi Izumi, Hiroki Sakaji·2024·Explainable AIEN

What this means for traders

LLMs can extract the specific factors driving a stock move and explain why, not just output a signal.

Abstract EN

Introduces LLMFactor, a novel framework employing Sequential Knowledge-Guided Prompting (SKGP) to identify factors influencing stock movements using LLMs. Extracts factors more directly related to stock market dynamics, providing clear explanations for complex temporal changes.

Factor ExtractionStock PredictionExplainable AISequential Prompting

HedgeAgents: A Balanced-aware Multi-agent Financial Trading System

Xiangyu Li, Yawen Zeng, Xiaofen Xing, Jin Xu, Xiangmin Xu·2025·Risk ManagementEN

What this means for traders

A fund-manager agent coordinating specialist hedging agents reached a 70% annualized return in a three-year backtest (2021-2023) on Bitcoin, Dow Jones stocks and forex.

Abstract EN

Introduces HedgeAgents, a multi-agent system aimed at bolstering system robustness via hedging strategies. The framework features a central fund manager and three hedging experts (Bitcoin, Dow Jones stocks, forex), achieving a 70% annualized return and a 400% total return in a 2021-2023 backtest.

Hedging StrategiesRisk ManagementPortfolio DiversificationMulti-Agent

LLMs for Quantitative Investment Research: A Practitioner's Guide

A.-H. Mihov, N. Firoozye, P. Treleaven·2025·Quantitative FinanceEN

What this means for traders

Industry consensus after three years of deployment: LLMs add real value as signal extractors and interpreters inside governed, deterministic pipelines, not as standalone forecasting engines.

Abstract EN

A practitioner-oriented review (UCL / DWS) of how LLMs are reshaping quantitative investment research across three fronts: research assistance, text-based signal extraction, and systemising expert judgment. Documents the field's hard empirical limits (temporal leakage, memorisation, behavioural biases, reproducibility) and provides governance guidelines plus an evaluation checklist for deploying LLMs in production research pipelines.

LLMsQuantitative ResearchGovernanceRAGBest Practices

F2Agent: Financial Fusion of Agentic Intelligence for Multimodal Trading

C. Liu, Y. Jin, S. Cai, P. Fang, X. Xiao, B. C. Ooi·2026·Multi-Agent SystemsEN

What this means for traders

How much each data source counts toward a verdict should be an explicit, inspectable number, not something buried inside a block of text.

Abstract EN

Most LLM trading systems merge market data, indicators, news and sentiment by concatenating them into a single prompt, which can let the textual signal dominate and loses cross-modal dependencies. F2Agent instead assigns one specialist encoder per modality and fuses them through learned modality-aware attention, regularised for prior diversity and for stability under single-modality perturbation. In a six-month backtest (April to September 2025) on five US stocks and Bitcoin, it ranks first on annualized return on all six assets against sixteen baselines, and its ablation shows the fusion design carries much of the gain: with the same specialist agents, replacing it with plain concatenation drops annualized return from 50.1% to 19.2% on AAPL.

Multi-AgentMultimodal FusionRobustnessSignal Weighting

Calibration-Induced Degeneracy: When an Expensive LLM Feature Receives Zero Weight

A. Mohanty·2026·Risk ManagementEN

What this means for traders

Before paying for a sophisticated signal, check two things: that it can mechanically change the answer at all, and that it beats a trivially cheap alternative.

Abstract EN

A fully audit-trailed case study in which an LLM scores the market importance of daily news headlines, and that score is added to implied-volatility baselines to forecast next-day equity risk. Calibration assigned the feature a weight of exactly zero, making every augmented forecast mathematically identical to its baseline, after the full-history inference had already been purchased. A near-free control that merely counts headlines did improve variance forecasting. The paper prescribes a viability checkpoint, fit first, perturb second, acquire last, that detects a feature which cannot move the output before any inference is paid for.

Feature ValidationVolatility ForecastingCost DisciplineNegative Results

MemArbiter: Decision-Time Memory Arbitration for Long-Horizon LLM Agents

J. Dong et al.·2026·System ArchitectureEN

What this means for traders

A constraint the system knows about but does not put in front of the model at decision time is, in practice, a constraint that does not exist.

Abstract EN

Identifies the Memory-Action Gap: in long-horizon agents, information can be stored and retrievable yet still fail to guide the action, because of how it is organised, prioritised and presented at the moment a decision is made. MemArbiter organises memory into functional banks (goal, task state, constraint, episodic, reference), scores decision relevance per bank and per item, and applies a temporal gate that deliberately shields goals and constraints from time decay. On 134 unseen ALFWorld tasks it reaches 82.8% success at a fixed 500-token memory budget against 61.9% for flat retrieval, and cuts the rate at which the agent repeats an action that has just failed from 47.1% to 26.7%.

Agent MemoryContext AssemblyConstraintsLong-Horizon Agents

Harness the Memory: A Holistic Benchmark of Agent Memory Substrates

Z. Huang et al.·2026·System ArchitectureEN

What this means for traders

Retrieving more context is not automatically better: past a point it crowds out the information the model actually needs to act on.

Abstract EN

Evaluates eleven memory substrates across seven families, three backbones and four task suites under twenty-six metrics. No substrate dominates: the winner inverts by regime, with structured graphs leading on dialogue question answering and cheap flat retrieval leading on code. Widening retrieval helps question answering but measurably degrades sequential decision-making, with attention probes showing mass draining from the action context into the retrieved block. Production-grade memory systems are also shown to cost roughly 2,700 to 9,000 auxiliary LLM calls and 11 to 23 hours on a single run of a ten-conversation dialogue benchmark, one to two orders of magnitude above lighter retrieval, an overhead usually reported nowhere.

Agent MemoryBenchmarkingRetrievalInference Cost

Stealing Reasoning Traces from Proprietary LLM APIs

A. Panfilov, D. Schmotz, I. Shumailov, L. Beurer-Kellner, J. Schaeffer, A. Prabhu, J. Geiping, M. Andriushchenko·2026·AI SecurityEN

What this means for traders

A model's hidden reasoning routinely contains sensitive material that never appears in its visible answer, so it must be treated as data worth protecting rather than as a harmless artifact.

Abstract EN

Reasoning models return their chain-of-thought to the client as an opaque encrypted block that the client must replay on later calls. These blocks are shown to be interchangeable across sessions, users and sibling models, so a weaker model from the same provider can be coerced into transcribing a stronger model's hidden reasoning verbatim. Decoding 315,320 reasoning blocks scraped from public repositories recovered 704 distinct privacy artifacts from genuine user sessions, including 62 API keys and 33 passwords; 64 of these 704 appeared nowhere in the visible transcript, and because the blocks work across users, sanitizing only the visible text does not protect them. The displayed reasoning summary is also measured to hide roughly five times more content than it shows.

LLM SecurityData LeakageChain-of-ThoughtResponsible Disclosure

Model Discovery Agent: LLM-Assisted Bayesian Experiment Design

K. Murphy·2026·Machine LearningEN

What this means for traders

Let the model propose and let explicit, auditable machinery score and decide: when every number behind a decision is computed outside the model, the decision can be checked rather than trusted.

Abstract EN

Couples an LLM used strictly as a proposer of candidate mechanisms with standard Bayesian machinery: sequential Monte Carlo for posteriors and evidence, and value-of-information to choose the next experiment. On existing physics, chemistry and regression benchmarks it cuts held-out prediction error and recovers the true mechanism in far fewer experiments than the published LLM-driven agents, and it also performs well on a new, harder neuroscience benchmark. The most transferable idea is the separation of roles: the LLM only proposes, while every number behind the decision (posterior, evidence, experiment score, forecast) is computed by Bayesian inference, never by the model. All experiments use a single model, so stability across models is not measured.

Bayesian InferenceExperiment DesignHypothesis GenerationModel Robustness

Why Better Models Can Create Riskier Systems: Evidence from LLM Agents in Financial Markets

J. Ross, E. So, Z. De Simone, C. Pozniak, A. W. Lo·2026·Multi-Agent SystemsEN

What this means for traders

Several analysts running on the same model and reading the same news are closer to a single opinion than to many: real diversity comes from independent information, not just from different personas.

Abstract EN

Argues that as language models grow more capable, their errors become more correlated with one another, so a market populated by many LLM agents can carry a risk floor that adding more agents never diversifies away. Across the models studied, the correlation of residual decisions between pairs of models rises with capability, while sharing a provider shows no significant effect. In a simulated single-asset market, more LLM traders improve price discovery under normal conditions, but when every agent reads the same misleading commentary, tracking error rises well above a noise-trader baseline for two of the three model families tested. The authors also show that mixing model families removes only the family-specific part of the correlation; the part that comes from a shared information environment remains.

Correlated ErrorsAgent DiversitySystemic RiskMisinformation

MemRiskBench: Trace-Aware Risk-Preserving Evaluation for Long-Horizon LLM Agents

J. Jiang, D. Yuan, W. Li·2026·System ArchitectureEN

What this means for traders

An assistant that remembers you should also know when a fact is out of date and forget it when you ask; a good average score says nothing about either.

Abstract EN

Argues that average scores hide the rare but serious failures of agents with long-term memory. Defines five risk types (stale facts after an update, unresolved conflicting facts, leakage across users, reuse of revoked memories, and gradual decay of standing constraints) and checks each one deterministically against the agent's execution trace instead of relying on an LLM judge. Across 120 scripted episodes and five small open models, overall pass rates conceal large per-risk gaps, with cross-user leakage the weakest category for every model tested. A subset that keeps only the episodes where models disagree reproduces the full ranking at a fifth of the cost.

Agent MemoryEvaluationData LeakageConstraints

Recursive Self-Improvement of AI Research Agents

D. Srikanth, B. Zhao, D. Xu, Y. Wu, Z. Jiang·2026·Machine LearningEN

What this means for traders

A system tuned and graded on the same examples ends up learning the exam; the only honest score comes from cases held back from every adjustment.

Abstract EN

Uses one AI agent to rewrite the code of another research agent in a two-level loop: the inner agent optimizes each task against a public score, while the outer loop keeps a rewrite only if it improves a private, held-out score the inner agent never sees. Over an eight-day autonomous run the system accepted seven improvements, including a bandit over drafting strategies and bounded summaries of long logs, and the resulting agent matched or beat a human-designed baseline on four held-out benchmarks. Its measured rate of reward hacking on a kernel-optimization test fell from 55% to 32%, on 38 kernel cases reported without error bars.

Self-Improving AgentsHeld-Out EvaluationReward HackingAgent Design