Project Blueprint

V1 objective

Create a daily regime-aware exposure overlay for a tradable ChiNext 50 ETF proxy.

State set

trend_up
crowded_up
fragile_range
stress_down
repair_rebound

Why this should be more reliable than a pure classifier

continuous scores + controlled exposure
no optimization on regime separation alone
frozen-hypothesis validation
costs and T+1 constraints built in

Success gates

out-of-sample Sharpe improvement > 0.15
max drawdown ratio vs passive < 0.75
upside capture >= 0.75
positive utility in >= 60% of test windows