dsh-jesse — strategy optimization for DeepSeek Harness, with the overfit meter on
Grid-sweep your strategy parameters on real data — then read the number that most optimizers hide: the train→test overfit gap, with a multiple-testing discount attached. The optimizer that tells you when it's lying to you. 在真实数据上扫策略参数——然后读出多数优化器不给你看的数字:train→test 过拟合缺口,附多重检验折减警示。会告诉你"我什么时候在骗你"的优化器。
Implements the core concept of Jesse (8.8k★) — the optimization loop — natively for DSH, with anti-overfit discipline borrowed from the factor-research world. Not affiliated. Family: dsh-TradingAgents · dsh-nautilus_trader (engine) · dsh-hummingbot (MM) · dsh-OpenBB · dsh-CCXT · dsh-Vibe-Trading.
Why (Jesse vs this vs the family's backtester)
| Jesse (framework) | dsh-nautilus_trader (sibling) | dsh-jesse | |
|---|---|---|---|
| Focus | accurate backtests + Optuna tuning | event-driven engine depth | the tuning loop itself |
| Runs where | Python lab | DSH tools | DSH in-session — agent sweeps params, reads the verdict |
| Anti-overfit | — | — | train/test split + overfit-gap column + Bonferroni-style discount + test-only-once rule |
Quick start
git clone https://github.com/vicfei/dsh-jesse && cd dsh-jesse
python3 scripts/optimize.py --data <klines.json> \
--grid '{"fast":[5,10,20],"slow":[30,60,90],"rsi_buy":[25,35]}' --fee_bps 5
Real smoke (binance BTC/USDT 365d, 2026-11-17): best train config −1.21% in a −30.6% train market — but the table shows its +2.96% overfit gap and the 18-config multiple-testing discount. In a year where buy-and-hold lost a third, trend-following cut the loss; the honest read is on the page.
Status & roadmap
v0 grid optimizer smoke-tested on live data. ROADMAP: M1 Optuna Bayesian search + walk-forward; M2 strategy DSL (Jesse-style lifecycle hooks).
License & disclaimer
MIT; concept inspired by MIT upstream — no upstream code used. Not investment advice. Optimization results are in-sample until proven otherwise.
No comments yet. Be the first to write one.