Skip to content

Sharpen v1.0.0

Choose a tag to compare

View all tags
@bigcan bigcan released this 18 Sep 09:57
v1.0.0
486d725

First public release of Sharpen: an agentic-AI, full-featured quant strategy builder — from a concept to a validated, trained and deployment-ready system.

Highlights

  • Agent-driven workflow. Built to be driven by an AI coding agent such as Claude Code or Codex, in any language. The rules it follows ship in CLAUDE.md: pre-registration, gates, staged training, and the invariants that must never break.
  • Seven-tier validation funnel (sharpen/signals/). Covers causality tripwires, IC and decay, capturability after costs, stability, purged path resampling, deflated Sharpe with FDR and the Harvey-Liu-Zhu hurdle, and orthogonality to a factor book.
  • Backtesting and falsification toolkit. Walk-forward and recent out-of-sample stages, realistic cost models, circular-shift timing nulls, edge-sign-flip and matched-exposure tests, planted-signal power, matched-null search, gate-reachability checks, multiplicity accounting, deflated Sharpe with PBO, and stress and noise robustness.
  • Crucible alpha mining (sharpen/crucible/, crucible-v14.0). Continuous automated alpha mining over free-data connectors (FRED, CFTC COT, SEC EDGAR, GDELT, Stooq, TWSE, TAIFEX). An LLM hypothesis proposer never sees any verdict. Gates are hash-frozen and every candidate is forward-incubated in a lockbox.
  • Frontier deep RL. SAC with CrossQ-style critics, distributional SAC, IQN, BDQ and PPO. Environments cover multi-asset allocation, execution scheduling and market making. Includes distributed Optuna HPO across GPU fleets and the six-stage SharpOps pipeline with a config validator that rejects leaky or fused runs.
  • Execution. Six broker adapters, a paper-portfolio executor with a parity harness, and a Docker stack with Prometheus, Grafana, Telegram alerts, drift detection and a kill file.

Get started

pip install -e ".[dev]"
python scripts/research/eval_signals.py --batch demo --panel "synthetic:1400,60" --out results/demo

No data and no API keys needed. The headline research numbers reproduce from a fresh clone in about a minute. See docs/REPRODUCE.md.

Also in this release

  • NEGATIVE_RESULTS.md: 76 strategies and probes across eight families, each with the number that decided it.
  • docs/LEAKS_FOUND.md: three bugs that had manufactured results, how each was caught, and the test that now guards it.
  • Test suite: 3261 passed, 25 skipped, 28 deselected, 1 xfailed on the release tree.

Apache-2.0. Not investment advice; see DISCLAIMER.md.

Assets 2
Loading