ML · Data Eng2026
Chaos Arbitrageur
An event-driven research platform that ingests physical-world alternative data instead of price charts — live port congestion and global news — then uses an LLM correlation agent and a vector memory of historical analogues to estimate equity impact on the most-exposed public companies.
- Live alerts on the globe
- 200+
- Historical analogues in memory
- 12
- Event-study validation
- CAR vs SPY
The problem
By the time a supply-chain disruption is visible in a price chart, the information is already priced in. The signal exists earlier — in port congestion data, in news wire volume, in the physical movement of goods — but it arrives as unstructured, geographically scattered noise that no conventional financial data pipeline is built to consume.
The approach
Ingest the physical world directly. Stream port congestion from IMF PortWatch and global news from GDELT, geolocate the disruptions, then hand the correlation problem to an LLM agent with a vector memory of historical analogues — because "what happened last time something like this occurred" is exactly the query a vector store answers well and a regression does not.
Markets react to physical-world shocks like port collapses and conflict long before the price charts catch up. I wanted to build the pipeline that watches the physical world directly and turns a disruption into a ranked list of exposed tickers.
Architecture
- 01
Alt-data ingest
Live IMF PortWatch congestion data and GDELT global news, geolocated onto an interactive globe.
- 02
Correlation agent
A LangChain + Claude agent that reasons from a disruption to the public companies most exposed to it.
- 03
Analogue memory
Pinecone vector store of historical shocks, queried for precedents matching the current event.
- 04
Event study
Simulator measuring cumulative abnormal returns against SPY, reported with t-statistics.
Stack
- Backend
- FastAPI · Python · DuckDB
- Intelligence
- LangChain · Claude · Pinecone
- Data
- IMF PortWatch · GDELT
- FastAPI
- Python
- Next.js
- LangChain
- Claude
- Pinecone
- IMF PortWatch
- GDELT
- DuckDB
- React-Leaflet
Outcome
A working pipeline from physical disruption to a ranked list of exposed tickers, with an event-study simulator to check whether the ranking would have meant anything historically. Over 200 live alerts surface on the globe, backed by a memory of 12 historical shocks.