Leading Indicators Information Arbitrage

Identifying divergences between qualitative sentiment and price data and evaluating the strength of qualitative data as leading indicators.

MCP server

Run the MCP server over stdio:

python mcp_server.py

The server name is info-arb.

Tools

Tool	Purpose
refresh_market_data	Fetches ticker-scoped news sentiment from Alpha Vantage and writes ticker sentiment rows plus related OHLCV into ClickHouse.
get_ticker_sentiment_price_analytics	Reads sentiment and OHLCV for a single symbol from ClickHouse, deduplicates similar items via embedding cosine similarity, and returns combined analytics (rolling sentiment vs typical, VWAP, SMAs, RSI, ATR, Bollinger, etc.).

refresh_market_data

• ticker (optional, default AAPL): symbol used for the Alpha Vantage NEWS_SENTIMENT request; co-mentioned tickers can still be ingested for OHLCV follow-up.

get_ticker_sentiment_price_analytics

• ticker (required): symbol to analyze.
• sentiment_lookback_days (default 90, max 365): window for sentiment rows.
• price_lookback_days (default 90, max 365): window for daily OHLCV.
• semantic_similarity_threshold (default 0.8): cosine similarity threshold for deduplicating near-duplicate news.
• sentiment_roll_short / sentiment_roll_long (defaults 5 / 20): rolling mean windows on deduplicated ticker sentiment scores.

Call refresh_market_data (for the symbols you care about) before expecting analytics to have data in ClickHouse.

---

Claude API client

Runs an Anthropic Messages request with the same logical tools as the MCP server, implemented in-process (no MCP transport). Requires ANTHROPIC_API_KEY in the environment or a .env file.

python claude_client.py

Optional flags:

python claude_client.py --model claude-sonnet-4-5 --prompt "Your custom prompt"

Tools exposed to Claude

• refresh_market_data: optional ticker (defaults to AAPL if omitted).
• get_ticker_sentiment_price_analytics: required ticker; optional sentiment_lookback_days, price_lookback_days, semantic_similarity_threshold, sentiment_roll_short, sentiment_roll_long (same defaults and bounds as the MCP tool).

Tool results are JSON-serialized back to the model for the next turn.