# Athenara — full registry corpus # 26 entries · 40182 bytes of content follow # Curated index with navigation instructions: https://athenara.info/llms.txt --- id: buy-and-hold-baseline name: Buy-and-hold baseline summary: The reference passive baseline every active result should be compared against; buys the index and does nothing. authors: [Athenara] origin: native license: MIT disclosure: fully-open strategy: [passive-baseline] status: active markets: [equities] tags: [baseline, passive, reference] added: 2026-08-15 prerequisites: trading: none ai: none programming: none setup: none --- The registry's reference point. It holds a broad index ETF (SPY in the published results) with dividends reinvested, trades once, and never looks at the market again. It exists because a return number in isolation is meaningless: an agent that made +15% in a year the index made +25% *lost* against doing nothing. Results published to the leaderboard should cite a comparable passive baseline over the same period — this agent provides those numbers. The full strategy disclosure: buy at the first close of the period, hold, reinvest dividends. There is no edge to protect, which is the point. --- id: finmem name: FinMem summary: An LLM trading agent that combines a configurable character profile with a layered memory module modeled on human trader cognition. authors: [Yangyang Yu, Haohang Li, Zhi Chen, Yuechen Jiang, Yang Li, Denghui Zhang, Rong Liu, Jordan W. Suchow, Khaldoun Khashanah] origin: external repo: https://github.com/pipiku915/FinMem-LLM-StockTrading license: MIT disclosure: fully-open tags: [llm-agent, memory, single-stock, profiling] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: advanced programming: moderate setup: advanced paper: [finmem-paper] --- FinMem is built from three modules: **profiling**, which defines the agent's professional character and risk personality; **memory**, a layered store with an adjustable cognitive span that ranks and retains market information over different time horizons; and **decision-making**, which converts retrieved memories into trade actions. The authors report that adjusting the agent's cognitive span and personality settings changes trading performance. The implementation supports OpenAI models, Hugging Face models served via Text Generation Inference, and Google Gemini, using text-embedding-ada-002 for retrieval. The repository's worked example trades TSLA over June–October 2022; the system was also entered in the IJCAI 2024 FinLLM Challenge single-stock trading task. --- id: finrobot name: FinRobot summary: An open-source multi-agent platform from AI4Finance Foundation that applies LLM agents to equity research, trading strategy, and risk evaluation tasks. authors: [AI4Finance Foundation, Hongyang Yang, Boyu Zhang, Neng Wang, Christina Dan Wang] origin: external repo: https://github.com/AI4Finance-Foundation/FinRobot website: https://finrobot.ai license: Apache-2.0 disclosure: fully-open frameworks: [autogen, langchain] tags: [multi-agent, llm, equity-research, robo-advisor] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: moderate programming: moderate setup: moderate builds_on: [fingpt] --- FinRobot organizes financial AI into four layers: a financial AI agents layer using chain-of-thought prompting, a financial LLM algorithms layer of domain-tuned models, LLMOps and DataOps layers for multi-source integration and model selection, and a foundation-model layer supporting plug-and-play LLMs. It builds on existing multi-agent frameworks including AutoGen and LangChain. A documented use case is automated equity research that produces structured stock analysis reports covering financial statement analysis, valuation multiples such as P/E and EV/EBITDA, and risk assessment. Paper: [arXiv:2405.14767](https://arxiv.org/abs/2405.14767). --- id: freqtrade name: Freqtrade summary: A Python cryptocurrency trading bot with backtesting, hyperparameter optimization, and an adaptive machine-learning module called FreqAI. authors: [Freqtrade development team] origin: external repo: https://github.com/freqtrade/freqtrade website: https://www.freqtrade.io license: GPL-3.0 disclosure: fully-open status: active tags: [crypto, trading-bot, backtesting, machine-learning, freqai] markets: [crypto] added: 2026-08-15 prerequisites: trading: moderate ai: basic programming: moderate setup: basic --- Freqtrade runs user-authored strategies against spot and futures crypto markets, with connectors for Binance, Bybit, OKX, Kraken, Gate, Bitget, Hyperliquid, HTX, BingX, and others through CCXT. It provides dry-run simulation, SQLite trade persistence, backtesting, and control via Telegram or a web UI. The FreqAI module automates training of predictive models on live market data: self-adaptive retraining during live operation, large-scale feature engineering, outlier removal, PCA dimensionality reduction, and classifier, regressor, and neural-network examples including LightGBM, XGBoost, and CatBoost. Active since 2017. --- id: hummingbot name: Hummingbot summary: An open-source Python framework for building and running market-making and arbitrage bots across centralized and decentralized crypto venues. authors: [Hummingbot Foundation] origin: external repo: https://github.com/hummingbot/hummingbot website: https://hummingbot.org license: Apache-2.0 disclosure: fully-open strategy: [market-making, arbitrage] status: active tags: [market-making, crypto, arbitrage, defi, high-frequency] markets: [crypto] added: 2026-08-15 prerequisites: trading: advanced ai: none programming: moderate setup: moderate --- Hummingbot provides a strategy framework plus a large connector library spanning three venue categories: central-limit-order-book centralized exchanges such as Binance, Coinbase, KuCoin, Bybit, OKX, and Kraken; CLOB decentralized exchanges including dYdX, Injective Helix, Hyperliquid, XRP Ledger, and Derive; and AMM DEXs including Uniswap, Raydium, PancakeSwap, Curve, Orca, Meteora, and Balancer. The project positions itself as infrastructure for high-frequency and market-making strategies rather than a specific trading model — the intelligence is in the strategies users build on it. --- id: tradingagents name: TradingAgents summary: A multi-agent LLM framework that assigns specialized analyst, researcher, trader, and risk-management roles to simulate the workflow of a trading firm. authors: [Tauric Research, Yijia Xiao, Edward Sun, Di Luo, Wei Wang] origin: external repo: https://github.com/TauricResearch/TradingAgents website: https://tauric.ai/ license: Apache-2.0 disclosure: fully-open frameworks: [langgraph] status: active tags: [multi-agent, llm, langgraph, debate] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: moderate programming: moderate setup: moderate implements: [multi-agent-debate] paper: [tradingagents-paper] --- TradingAgents decomposes trading into distinct LLM-powered roles: fundamentals, sentiment, news, and technical analysts feed reports to bullish and bearish researchers who debate the evidence, after which a trader agent composes a decision and a risk-management team and portfolio manager approve or reject it. The system is built on LangGraph and supports OpenAI, Google, Anthropic, xAI, DeepSeek, Qwen, GLM, MiniMax, OpenRouter, Ollama, AWS Bedrock, and any OpenAI-compatible endpoint. Released by Tauric Research alongside the paper describing the framework (see related). --- id: fingpt name: FinGPT summary: An open-source family of financial large language models and fine-tuning pipelines from AI4Finance Foundation, built around low-cost LoRA adaptation of open base models. authors: [AI4Finance Foundation, Hongyang Yang, Xiao-Yang Liu, Christina Dan Wang] origin: external repo: https://github.com/AI4Finance-Foundation/FinGPT website: https://fingpt.io license: MIT components: [data pipelines, instruction datasets, LoRA fine-tuning recipes, trained adapters] tags: [financial-llm, lora, fine-tuning, sentiment-analysis, rag] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: advanced programming: advanced setup: advanced --- FinGPT is a model and tooling suite rather than a trading agent: it provides data pipelines, instruction-tuning datasets, and LoRA fine-tuning recipes for adapting open base models to financial sentiment analysis, relation extraction, named entity recognition, and question answering. Components include FinGPT-Forecaster for stock movement prediction, FinGPT-RAG for retrieval-augmented sentiment analysis, and FinGPT-Benchmark for instruction-tuning evaluation. The project emphasizes cheap re-tuning to keep pace with changing markets and publishes trained adapters on [Hugging Face](https://huggingface.co/FinGPT). Papers: [arXiv:2306.06031](https://arxiv.org/abs/2306.06031), [arXiv:2306.12659](https://arxiv.org/abs/2306.12659), [arXiv:2310.04793](https://arxiv.org/abs/2310.04793). Categorized as an architecture/framework: it supplies the financial-LLM layer that agents (for example FinRobot) build on, rather than making trading decisions itself. --- id: finrl name: FinRL summary: A deep reinforcement learning library that packages market environments, DRL algorithms, and backtesting into a pipeline for training automated trading agents. authors: [AI4Finance Foundation, Xiao-Yang Liu, Hongyang Yang, Christina Dan Wang] origin: external repo: https://github.com/AI4Finance-Foundation/FinRL website: https://ai4finance.org license: MIT components: [market environments, DRL agents, backtesting] tags: [reinforcement-learning, drl, portfolio-allocation, backtesting, gym] markets: [equities] added: 2026-08-15 prerequisites: trading: moderate ai: advanced programming: moderate setup: moderate implements: [rl-policy] --- FinRL supplies gym-style environments built from market data for indices including NASDAQ-100, S&P 500, HSI, and SSE 50, together with implementations of DQN, DDPG, PPO, SAC, A2C, and TD3 and a backtesting module. Environments model practical frictions such as transaction costs and liquidity constraints — the fidelity issues that make or break the RL-policy pattern (see related). The library ships tutorials for single-stock trading, multi-stock trading, and portfolio allocation. Maintained by the AI4Finance Foundation; introduced at the Deep RL Workshop, NeurIPS 2020 ([arXiv:2011.09607](https://arxiv.org/abs/2011.09607)). --- id: lean name: LEAN summary: QuantConnect's open-source event-driven algorithmic trading engine, supporting strategies written in Python or C# across nine asset classes. authors: [QuantConnect] origin: external repo: https://github.com/QuantConnect/Lean website: https://www.lean.io license: Apache-2.0 tags: [algorithmic-trading, backtesting, event-driven, multi-asset, engine] markets: [multi-asset] added: 2026-08-15 prerequisites: trading: advanced ai: none programming: advanced setup: moderate --- LEAN is an event-driven engine that runs the same algorithm code in backtesting, paper trading, and live deployment. It supports equities, forex, options, futures, future options, indexes, index options, cryptocurrency, and CFDs from a single portfolio, and integrates with roughly 40 price, fundamental, and alternative data sources plus user-supplied custom datasets. The engine is the core of the QuantConnect platform but is usable standalone. Written primarily in C# with a Python API; in continuous development since 2014. --- id: multi-agent-debate name: Multi-agent debate summary: Opposing analyst agents argue a bullish and bearish case before a decision-maker agent commits to a trade. aka: [bull-bear debate, adversarial analysts] components: [bullish researcher, bearish researcher, moderator or trader, risk manager] origin: native tags: [multi-agent, llm, deliberation, design-pattern] added: 2026-08-15 prerequisites: trading: basic ai: moderate programming: none setup: none paper: [tradingagents-paper] --- ## How it works Two (or more) researcher agents are given the same evidence — prices, news, fundamentals — and opposite mandates: one builds the strongest case *for* a position, the other the strongest case *against*. A downstream agent (trader, moderator, or portfolio manager) reads the debate and decides. Optionally a risk-management agent can veto or resize the final decision. The pattern borrows from adversarial deliberation: a single LLM analyst tends to anchor on the first narrative it forms, while forcing an explicit counter-argument surfaces disconfirming evidence before capital is committed. ## Design choices - **Debate depth.** One round is cheap; multi-round rebuttals raise cost roughly linearly and tend to hit diminishing returns quickly. - **Symmetric evidence.** Both sides should see identical data. If the bull sees the news feed and the bear only sees prices, the debate measures data access, not reasoning. - **Decision extraction.** The judge should output a structured decision (direction, size, confidence), not prose, so the execution layer doesn't re-interpret an essay. ## Failure modes - Both debaters converge on the consensus view and the "debate" is theater. - The judge rewards rhetorical confidence rather than evidence quality. - Token costs scale with debate rounds × assets × rebalance frequency — expensive at high frequency, so the pattern fits daily/weekly horizons better than intraday. ## Known implementations TradingAgents (see related) structures its analyst layer this way, with bull/bear researchers feeding a trader agent and a risk-management team. --- id: nautilustrader name: NautilusTrader summary: A Rust-native, event-driven algorithmic trading platform whose backtest and live engines share identical execution semantics. authors: [Nautech Systems] origin: external repo: https://github.com/nautechsystems/nautilus_trader website: https://nautilustrader.io/docs/ license: LGPL-3.0 tags: [event-driven, backtesting, rust, multi-asset, live-trading] markets: [multi-asset] added: 2026-08-15 prerequisites: trading: advanced ai: none programming: advanced setup: moderate --- NautilusTrader pairs a Rust core engine with a Python control plane via PyO3 bindings, plus Cython infrastructure components. Its central design claim is **research-to-live parity**: the same deterministic time model and execution semantics run in both backtesting and production, so a strategy deploys without rewriting — the property that matters most when an AI agent's backtest is the evidence for letting it trade. It supports nanosecond-resolution historical tick data and multi-venue backtests across crypto spot and derivatives, FX, equities, futures, options, and prediction/sports-betting markets. Categorized here as an architecture/framework rather than an agent: it is the engine an agent runs on, not a decision-making system itself. --- id: planner-executor name: Planner–executor summary: A planning agent decomposes a trading objective into steps that cheaper executor agents or plain tools carry out. aka: [orchestrator-workers, manager-worker] components: [planner, executors, verifier] origin: native tags: [multi-agent, llm, orchestration, design-pattern] added: 2026-08-15 prerequisites: trading: basic ai: moderate programming: none setup: none --- ## How it works A planner agent owns the objective ("rebalance the portfolio for the coming week") and produces an explicit plan: which data to gather, which analyses to run, which orders to draft. Each step is executed by a worker — a smaller model, a single-purpose agent, or a deterministic tool — and results flow back to the planner, which revises the plan or hands the outcome to a verifier before anything reaches the market. The separation matters because planning and execution fail differently: planning errors are reasoning errors, execution errors are integration errors. Isolating them makes both debuggable. ## Design choices - **Plan as artifact.** Persist the plan as structured data (steps, dependencies, status). It becomes the audit trail regulators-of-your-own-money want anyway. - **Executor capability boundaries.** Workers get the narrowest tool access that does the job — the researcher can read market data but cannot place orders. - **Verification gate.** A final check (rule-based or agentic) between the plan's output and order submission is where risk controls naturally live. ## Failure modes - The planner micro-manages: plans so detailed that workers add nothing, doubling cost for no benefit. - Stale plans: markets move while a long plan executes; plans need revision triggers, not just linear execution. - Verifier rubber-stamping — if the verifier is the same model with the same context, it tends to approve its own reasoning. --- id: qlib name: Qlib summary: Microsoft's AI-oriented quantitative investment platform covering the research workflow from data handling and model training to portfolio construction and backtesting. authors: [Microsoft Research, Xiao Yang, Weiqing Liu, Dong Zhou, Jiang Bian, Tie-Yan Liu] origin: external repo: https://github.com/microsoft/qlib website: https://qlib.readthedocs.io/en/latest/ license: MIT components: [data infrastructure, learning framework, portfolio generation, order execution, backtest] tags: [quant-platform, machine-learning, backtesting, alpha-research, python] markets: [equities] added: 2026-08-15 prerequisites: trading: advanced ai: advanced programming: moderate setup: moderate --- Qlib provides a full quant research stack: a storage and data-handling layer optimized for financial time series, model training with a library of supervised learning and reinforcement learning models, and downstream portfolio generation, order execution, and backtesting. It was introduced to address the infrastructure gap that arises when AI methods are adopted in quantitative investment workflows ([arXiv:2009.11189](https://arxiv.org/abs/2009.11189)). Written in Python and released by Microsoft under MIT. Capabilities include quant dataset management, a model zoo, automated quant research, and a reinforcement learning module. --- id: rl-policy name: Reinforcement-learning policy summary: An agent's trading decisions come from a policy trained with reinforcement learning against a market environment. aka: [drl-trader, deep-rl agent] components: [market environment, reward function, policy network, training pipeline] origin: native tags: [reinforcement-learning, design-pattern, backtesting] added: 2026-08-15 prerequisites: trading: moderate ai: advanced programming: none setup: none --- ## How it works The market is modeled as an environment (state = prices, indicators, positions; actions = buy / sell / hold or portfolio weights; reward = usually risk-adjusted return). A policy is trained offline against historical or simulated data with algorithms like PPO, SAC, or DQN, then deployed to act on live observations. Unlike LLM-based agents the policy is small, fast, and cheap at inference time — the intelligence is baked in during training. ## Design choices - **Reward shaping is the strategy.** Optimizing raw return produces leverage-seeking policies; most practical rewards penalize drawdown or volatility explicitly. - **Environment fidelity.** Transaction costs, slippage, and market impact must be in the training environment or the policy learns to exploit their absence. - **Regime coverage.** Policies interpolate; they do not extrapolate. Training data needs bear markets, or the first real one is out-of-distribution. ## Failure modes - Backtest overfitting: the classic failure. Millions of gradient steps against one historical path is a strong optimizer pointed at a weak target. - Non-stationarity: a policy trained on 2015–2022 microstructure decays silently as the market changes. - Reward hacking: policies discover degenerate behaviors (churning tiny positions for reward quirks) that only show up in live trading. ## Hybrid variants RL policies increasingly appear *inside* agentic systems — an LLM planner selects among trained policies, or an RL execution layer implements decisions an LLM makes at lower frequency. --- id: ectsum name: ECTSum summary: A benchmark of 2,425 long earnings call transcripts paired with expert-written telegram-style bullet-point summaries drawn from Reuters articles. authors: [Rajdeep Mukherjee, Abhinav Bohra, Akash Banerjee, Soumya Sharma, Manjunath Hegde, Afreen Shaikh, Shivani Shrivastava, Koustuv Dasgupta, Niloy Ganguly, Saptarshi Ghosh, Pawan Goyal] origin: external repo: https://github.com/rajdeep345/ECTSum website: https://aclanthology.org/2022.emnlp-main.748/ license: GPL-3.0 access: open source: Earnings call transcripts of US public companies; summaries derived from Reuters articles coverage: assets: 2,425 transcript–summary pairs (US public companies) size: 2,425 document–summary pairs tags: [earnings-calls, summarization, benchmark, nlp, financial-documents] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: moderate programming: basic setup: basic --- ECTSum targets extreme summarization of long, unstructured financial documents: its document-to-summary compression ratio of 103.67 is the highest among long-document summarization datasets of comparable length, forcing models to isolate only the most material facts — the same skill a research agent needs when reading an earnings call before a trading decision. Published as a long paper at EMNLP 2022 ([arXiv:2210.12467](https://arxiv.org/abs/2210.12467)); the release includes baseline results and the authors' ECT-BPS model. --- id: edgar-corpus name: EDGAR-CORPUS summary: A corpus of 220,375 SEC 10-K annual reports from 1993 to 2020, split into their individual item sections and released as JSON. authors: [Lefteris Loukas, Manos Fergadiotis, Ion Androutsopoulos, Prodromos Malakasiotis] origin: external website: https://huggingface.co/datasets/eloukas/edgar-corpus license: Apache-2.0 access: open source: SEC EDGAR coverage: assets: All US publicly traded companies, 220,375 10-K filings, 1993–2020 frequency: annual size: ~40.7 GB (176,289 train / 22,050 validation / 22,036 test in the full config) formats: [json] tags: [sec-filings, 10-k, edgar, corpus, pretraining] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: basic programming: moderate setup: basic --- EDGAR-CORPUS collects more than 25 years of annual reports parsed into the standard 10-K item sections (Items 1–15, including 1A risk factors, 7A market risk, 9A controls), so individual sections can be used in isolation — risk factors for risk analysis, MD&A for fundamental research. The data is unannotated raw text intended as a pretraining and retrieval corpus. Companion releases: EDGAR-W2V (Word2Vec embeddings trained on the corpus, which outperform generic embeddings on financial tasks) and EDGAR-CRAWLER, an open-source tool for collecting and processing future filings. Published at the ECONLP 2021 workshop ([arXiv:2109.14394](https://arxiv.org/abs/2109.14394)). --- id: financial-phrasebank name: Financial PhraseBank summary: A human-annotated sentiment dataset of 4,840 English sentences from financial news, labelled positive, negative, or neutral from an investor's perspective. authors: [Pekka Malo, Ankur Sinha, Pyry Takala, Pekka Korhonen, Jyrki Wallenius] origin: external website: https://huggingface.co/datasets/takala/financial_phrasebank license: CC-BY-NC-SA-3.0 access: open source: English-language financial news coverage: assets: 4,840 annotated sentences size: 4,840 sentences in four agreement-level configurations tags: [sentiment-analysis, financial-news, classification, annotated, benchmark] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: basic programming: basic setup: basic --- The standard small-scale benchmark for financial sentiment classification, originating from *Good Debt or Bad Debt: Detecting Semantic Orientations in Economic Texts* (JASIST 2014, [arXiv:1307.5336](https://arxiv.org/abs/1307.5336)). Sentences were annotated by multiple people and the dataset ships in four configurations by inter-annotator agreement: ≥50% (4,846 instances), ≥66% (4,217), ≥75% (3,453), and unanimous (2,264). Widely used to fine-tune and evaluate financial LLMs. Note the non-commercial share-alike license, and that the Hugging Face dataset viewer does not render it (the repo uses a loading script). --- id: fnspid name: FNSPID summary: A time-series dataset pairing 29.7 million stock price records with 15.7 million financial news articles for 4,775 companies in the S&P 500 universe, 1999–2023. authors: [Zihan Dong, Xinyu Fan, Zhiyuan Peng] origin: external repo: https://github.com/Zdong104/FNSPID_Financial_News_Dataset website: https://huggingface.co/datasets/Zihan1004/FNSPID license: CC-BY-NC-4.0 access: open source: Four financial news websites, aligned to stock price series coverage: assets: 4,775 companies (S&P 500 universe), 1999–2023 frequency: daily prices with timestamped news size: 29.7M price records + 15.7M news records (~30 GB) tags: [financial-news, stock-prices, time-series, sentiment, sp500] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: basic programming: moderate setup: basic --- FNSPID (Financial News and Stock Price Integration Dataset) aligns news articles to price series by timestamp and ticker, with per-article metadata including publication date, title, stock symbol, URL, publisher, and author, plus multiple generated summaries per article. The authors report that dataset scale improves prediction accuracy and that adding sentiment scores gives modest gains for transformer models; they publish a reproducible pipeline for extending the dataset. Paper: [arXiv:2402.06698](https://arxiv.org/abs/2402.06698). **License note:** CC BY-NC 4.0 on the Hugging Face card — commercial use requires permission. The GitHub repository carries no explicit license. --- id: finagent-paper name: "A Multimodal Foundation Agent for Financial Trading (FinAgent)" summary: Presents FinAgent, a multimodal trading agent that processes numerical, textual, and visual market data with dual-level reflection and diversified memory retrieval. authors: [Wentao Zhang, Lingxuan Zhao, Haochong Xia, Shuo Sun, Jiaze Sun, Molei Qin, Xinyi Li, Yuqing Zhao, Yilei Zhao, Xinyu Cai, Longtao Zheng, Xinrun Wang, Bo An] origin: external year: 2024 venue: arXiv preprint arxiv: "2402.18485" tags: [multimodal, foundation-agent, tool-use, reflection, memory] markets: [multi-asset] added: 2026-08-15 prerequisites: trading: moderate ai: advanced programming: none setup: none --- FinAgent's market intelligence module ingests numerical series, text, and chart images, while a dual-level reflection mechanism supports both rapid adaptation to market changes and longer-horizon lesson extraction. The agent is tool-augmented, incorporating established trading strategies and expert knowledge, and emphasizes reasoning transparency. Evaluation spans six datasets across stocks and cryptocurrency against nine baselines; the authors report average profit improvements above 36%, including a 92.27% return on one dataset. No public implementation was found at time of indexing — a reproduction would be a valuable contribution. --- id: finmem-paper name: "FinMem: A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design" summary: Proposes an LLM trading agent whose layered memory and configurable character profile let it adapt to new market information over multiple time horizons. authors: [Yangyang Yu, Haohang Li, Zhi Chen, Yuechen Jiang, Yang Li, Denghui Zhang, Rong Liu, Jordan W. Suchow, Khaldoun Khashanah] origin: external repo: https://github.com/pipiku915/FinMem-LLM-StockTrading year: 2023 venue: ICLR 2024 Workshop on LLM Agents arxiv: "2311.13743" tags: [llm-agent, memory, cognitive-architecture, profiling] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: moderate programming: none setup: none --- Presents a three-module agent architecture combining profiling, layered memory processing that imitates human trader cognition, and decision-making. The design lets the agent self-evolve its domain knowledge, react to new investment cues, and refine trading decisions over time. Experiments on real-world financial data report performance above algorithmic baselines, with results varying by the agent's cognitive span and personality configuration — an early demonstration that memory design, not just model choice, moves trading performance. --- id: llm-trading-agent-survey name: "Large Language Model Agent in Financial Trading: A Survey" summary: Surveys the research landscape of LLM-based agents applied to financial trading, covering architectures, data inputs, backtesting results, and open challenges. authors: [Han Ding, Yinheng Li, Junhao Wang, Hang Chen, Doudou Guo, Yunbai Zhang] origin: external year: 2024 venue: arXiv arxiv: "2408.06361" tags: [survey, llm-agent, backtesting, literature-review] markets: [multi-asset] added: 2026-08-15 prerequisites: trading: basic ai: basic programming: none setup: none --- Reviews how LLM agents have been constructed for trading tasks, comparing agent architectures, the data modalities they consume, and reported backtesting performance across published systems. It catalogues the obstacles that recur in this literature and identifies directions for future research. A good entry point into the field before reading the individual system papers indexed here. --- id: tradingagents-paper name: "TradingAgents: Multi-Agents LLM Financial Trading Framework" summary: Introduces a multi-agent LLM trading framework whose specialized roles mirror the structure of a professional trading firm. authors: [Yijia Xiao, Edward Sun, Di Luo, Wei Wang] origin: external repo: https://github.com/TauricResearch/TradingAgents year: 2024 venue: arXiv (oral, Multi-Agent AI in the Real World workshop) arxiv: "2412.20138" tags: [multi-agent, llm, debate, risk-management] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: moderate programming: none setup: none --- The paper describes a system of LLM agents acting as fundamental, sentiment, and technical analysts, bull and bear researchers, traders with differing risk tolerances, and a risk-management team, collaborating to reach trading decisions. The authors report gains in cumulative return, risk-adjusted return, and drawdown relative to baseline strategies. The accompanying implementation is released publicly by Tauric Research (see related agent) — one of the few papers in this space with a maintained open-source system behind it. --- id: finben name: FinBen summary: A holistic open-source financial benchmark of 42 datasets spanning 24 tasks, including the first stock-trading evaluation for LLMs. authors: [Qianqian Xie, Weiguang Han, Zhengyu Chen, Ruoyu Xiang, Jimin Huang, The FinAI, et al.] origin: external repo: https://github.com/The-FinAI/FinBen website: https://proceedings.neurips.cc/paper_files/paper/2024/hash/adb1d9fa8be4576d28703b396b82ba1b-Abstract-Datasets_and_Benchmarks_Track.html tasks: [information-extraction, textual-analysis, question-answering, text-generation, risk-management, forecasting, decision-making, stock-trading] tags: [benchmark, financial-llm, evaluation, stock-trading, rag] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: advanced programming: moderate setup: moderate builds_on: [pixiu-flare] --- FinBen evaluates LLMs across eight aspects: information extraction, textual analysis, question answering, text generation, risk management, forecasting, decision-making, and bilingual English/Spanish coverage. It introduced the first stock-trading evaluation for financial LLMs plus novel agent and retrieval-augmented generation evaluations, and released new datasets for regulations and stock trading. The authors evaluated 21 representative LLMs (including GPT-4, ChatGPT, and Gemini), finding strong performance on extraction and textual analysis but weakness on advanced reasoning, generation, and forecasting. Published in the NeurIPS 2024 Datasets and Benchmarks track ([arXiv:2402.12659](https://arxiv.org/abs/2402.12659)); it grew out of the FLARE benchmark (see related). The repository carries no license file at time of indexing. --- id: investorbench name: InvestorBench summary: A benchmark for evaluating LLM-based agents on financial decision-making tasks across single stocks, cryptocurrencies, and ETFs. authors: [Haohang Li, Yupeng Cao, Yangyang Yu, Shashidhar Reddy Javaji, Zhiyang Deng, Yueru He, Yuechen Jiang, Zining Zhu, K.P. Subbalakshmi, Jimin Huang, Lingfei Qian, Xueqing Peng, Jordan W. Suchow, Qianqian Xie] origin: external website: https://aclanthology.org/2025.acl-long.126/ tasks: [stock-trading, crypto-trading, etf-trading] tags: [benchmark, llm-agent, decision-making, evaluation] markets: [multi-asset] added: 2026-08-15 prerequisites: trading: moderate ai: advanced programming: moderate setup: expert builds_on: [finmem-paper] --- InvestorBench pairs a task-agnostic LLM agent framework with standardized datasets and simulated market environments so different backbone models can be compared on the same footing. The agent framework decomposes into Brain, Perception, Profile, Memory, and Action modules, with data drawn from open sources and third-party APIs including Yahoo Finance and SEC EDGAR. The authors evaluate thirteen LLMs as backbones across market environments, scored with standard quantitative finance metrics. Published at ACL 2025 main conference ([arXiv:2412.18174](https://arxiv.org/abs/2412.18174)); no dedicated public code repository was found at time of indexing. --- id: pixiu-flare name: PIXIU / FLARE summary: An open-source financial LLM resource bundling the FinMA instruction-tuned models, a 136K-sample instruction dataset, and the FLARE evaluation benchmark. authors: [Qianqian Xie, Weiguang Han, Xiao Zhang, Yanzhao Lai, Min Peng, Alejandro Lopez-Lira, Jimin Huang] origin: external repo: https://github.com/The-FinAI/PIXIU website: https://huggingface.co/TheFinAI license: MIT tasks: [sentiment-analysis, classification, knowledge-extraction, numerical-reasoning, summarization, credit-scoring, forecasting] tags: [benchmark, financial-llm, instruction-tuning, evaluation, flare] markets: [equities] added: 2026-08-15 prerequisites: trading: basic ai: advanced programming: moderate setup: moderate --- PIXIU released the first open financial LLM suite in three parts: **FinMA**, LLaMA-based models fine-tuned on financial instructions; **FIT**, an instruction dataset of 136K samples covering multiple financial tasks and document types; and **FLARE**, the evaluation benchmark that originally spanned five financial NLP tasks plus a prediction task over nine datasets ([arXiv:2306.05443](https://arxiv.org/abs/2306.05443)). FLARE has since grown into the FinBen evaluation suite (~32 datasets across sentiment analysis, classification, knowledge extraction, numerical reasoning, summarization, credit scoring, and forecasting) with a public leaderboard; English, Spanish, and Chinese are covered. Models and evaluation datasets are published on Hugging Face under The FinAI organization. --- id: buy-and-hold-spy-2024 agent: buy-and-hold-baseline mode: backtest period: start: 2024-01-02 end: 2024-12-31 markets: [equities] capital: initial: 10000 currency: USD max_leverage: 1 costs: commission_model: "none — single commission-free ETF purchase" slippage_model: "none — one market order at the daily close" financing_included: false fills: modeled metrics: total_return_pct: 25.6 sharpe: 1.47 max_drawdown_pct: -8.4 volatility_pct: 12.6 n_trades: 1 n_days: 252 selection: variants_tested: 1 selection_method: "only variant — the strategy has no parameters" runs: n: 1 seed_policy: "deterministic strategy" data: source: "SPY daily adjusted close (dividends reinvested), Yahoo Finance chart API" point_in_time: true survivorship_bias_free: true verification: method: author-reported status: active reported: 2026-08-15 --- ## Method Buy SPY at the first 2024 close (2024-01-02), hold through the last (2024-12-31), dividends reinvested — computed on Yahoo Finance's adjusted close series (252 trading days). - **Total return**: last adjusted close / first adjusted close − 1. - **Volatility**: standard deviation of daily returns, annualized (×√252). - **Sharpe**: daily excess return over a 5.2% annual risk-free rate (2024 T-bill area), annualized. With rf = 0 it would be 1.88 — always check which convention a Sharpe you're comparing against uses. - **Max drawdown**: worst peak-to-trough decline on adjusted closes; −8.4% occurred in the early August 2024 selloff. ## Assumptions No transaction costs beyond the declared model: the strategy trades once; a commission-free ETF purchase and frictionless dividend reinvestment are assumed. The capital base is a nominal $10,000 — the strategy scales linearly, so the percentage metrics are capital-independent up to lot-size effects. There is no development period to declare: the strategy has zero tuned parameters, so nothing was fitted to the evaluation window. ## Reproduction Fetch SPY daily data for 2024 from any total-return source and apply the formulas above; results should match to within rounding of the data vendor's adjustments.