Vibe Investing is a research and curation repository on using large language models as a tool for investing in US equities (Nasdaq and S&P 500), cryptocurrencies, and related sectors. It collects strategy write-ups, comparative guides, investment columns, working papers, technical notes, and the source code of several self-built trading and analysis tools. Most content is written in Korean; an English version of the main overview is provided.
The repository defines vibe investing as an agentic pipeline that starts from a natural-language instruction, has the model call tools to collect market data, news, and on-chain signals, and produces a decision that a person can verify. Its stated components are natural-language strategy definition, active tool calling, multi-agent consensus among bull, bear, risk, and portfolio-manager roles, and a reasoning trail a human can audit.
A recurring theme is a multi-model investment committee. In the ARDS (Adaptive Recession-Defensive Strategy) study, the same prompt and market data were given to Claude, Gemini, ChatGPT, and DeepSeek: macro diagnoses converged (composite standard deviation 2.29 percentage points), while position sizing diverged (coefficient of variation 52%), and each model settled into a distinct portfolio-manager persona. Later projects extend the committee idea to disclosure analysis and secret scanning.
Five self-authored guide series form the core: Claude quant scripts that pair eight quantitative strategies with their academic sources, Python skeletons, and prompts; an evaluation of 30+ AI investing tools for US stocks scored on five axes; an LLM crypto-trading guide centred on the Alpha Arena benchmark; a comparison of 9 conventional and 8 AI/LLM open-source trading bots that catalogues ten structural risks of LLM bots; and a review of AI quant prompt repositories on GitHub.
Features
- Toss × AMQS dashboard: applies an adaptive momentum strategy to Korean stocks and ETFs through the Toss Securities Open API, with a mock mode and a three-year domestic backtest
- ARDS-X regime classifier: classifies drawdowns as correction, oversold, structural decline, or recession rebalancing from FRED and yfinance data, with a two-axis regime map, hysteresis, and a 2014–2026 backtest
- Earnings Momentum Agent: a pipeline that ranks a top-30 list from low-point rebounds and earnings surprises
- Nasdaq–BTC coupling bot: tracks 30-day rolling correlation between BTC and QQQ, classifies six regimes, and emits signals
- Harness Quant v2: an in-development LLM trading platform with a six-scenario matrix, backtesting, MCP, and multi-agent debate
- LAON VaultGuard: an npm package and VS Code extension that scans git repositories for cloud secrets using keyword filters followed by multi-model cross-checking, with an offline Ollama mode
- CASSANDRA AI: collects Korean DART disclosures, analyses relationships between companies, people, and entities with multiple models, and visualises them as a graph
- Investment columns: data-driven pieces on macro trends, industry shifts, and market structure, indexed by topic
- Working papers: four SSRN papers on cryptocurrency market microstructure and multilingual LLM evaluation
- Technical notes: guides on serverless and free tiers, token optimisation, RAG, local model deployment, time-series models, and comparisons of LLM wiki tools
