dsh-quant-factor-pipeline

Factor research failures rarely come from the model itself but from skipped engineering discipline: verifying panel coverage before any downstream use, anchoring signals to the T-1 close to kill look-ahead bias, and relying on Newey-West t-statistics instead of naive t for IC significance. dsh-quant-factor-pipeline wraps the full chain — probe, panel build, IC/5-gate test, decile long-short backtest, and regime detection — into six MCP tools that a model inside DeepSeek Harness can orchestrate step by step. Every tool emits machine-readable, traceable, auditable output, enforcing the correct order of operations at the workflow level.

The toolset covers A-share daily panel fetching with turnover-coverage probing (panel_probe), CGO disposal-effect factor construction with RankIC and R-squared mechanical health checks (ic_gate), coarse-grained regime and correlation-break detection on signal-return series (regime_check), a pure-Node decile long-short backtest approximation (factor_backtest), and persistent project manifest save and load (project_save, project_load). The service carries zero third-party dependencies, uses only Node built-in modules, communicates over stdio with NDJSON framing, restricts data sources to verified Eastmoney and Tencent snapshot endpoints, and resolves every local path through environment variables for drop-in deployment on any machine.

Built for quant researchers and factor-strategy developers already working inside dsh, or for teams that want a model to auto-orchestrate multi-step research while retaining a human audit trail. If you are tired of manually running scripts and visually re-checking coverage and look-ahead issues every cycle, this pipeline bakes the probe-first, verify-then-run discipline into tool-call ordering — a practical addition to any factor research stack.