Strategy Backtesting Pipeline

LLM-generated strategies → VectorBT screen → Lumibot adjudication → DeepSeek judge → Alpaca paper data: tiingo · window 2015-01-01→2024-12-31 · universe 25 · survivorship-verified from 2013-01-01

Pool reconstructed by the causal liquidity rule as of 2015-01-01 from the survivorship-free universe; FIXED as of t0 (free-tier quota) — names newly liquid after 2015-01-01 are not added mid-window.

GENERATED

PASSED SCREEN (VectorBT)

ADJUDICATED (Lumibot)

PROMOTABLE TO LIVE

Screen vs. adjudication

The VectorBT screen is infinite-capacity (ranking only). Lumibot is the real-budget source of truth (official metrics). Unfunded entries are the capacity teeth — signals the screen took for free but the finite budget / max_positions cap rejected. Budget: $100,000.

Strategy	Screen Sharpe	Passed	Adj Sharpe	Adj Return	Max DD	Fills	Unfunded
mom_above_ema	+0.39	✓	+0.40	+64.3%	-30.5%	2699	1166
trend_sma	+0.29	—	screened out (did not reach adjudication)
trend_capped2	+0.29	—	screened out (did not reach adjudication)
meanrev_rsi	+0.14	—	screened out (did not reach adjudication)