Week 4 cost normalization review design

Week 4 cost normalization review design

Purpose

Separate whether the performance differences in existing EXP-002~EXP-006 come from the strategy signals or from an execution contract that misapplied per-product taxes and CREON commissions. Existing outputs are not overwritten; as EXP-007, the same data and the same signals are rerun.

Fixed contract

• Initial capital: the existing experiment contracts are kept. The Connors portfolio is 100,000,000 KRW for daily bars and 3,000,000 KRW for 5-minute bars.
• Compounding: the balance after costs from each fill is the base for the next trade. 5-minute bars use only 95% of current cash for entry.
• Execution: the existing next-day open / next 5-minute bar open rules are kept.
• Slippage: fixed at 0bp in this 1st normalization, leaving the limitation that no real quote data exist.

Cost matrix

The commission in normalized_actual_proxy is an assumption made before checking actual account fills. Once the actual CREON fee plan is confirmed, this profile is re-fixed as the new baseline.

Execution scope

1. Connors RSI(2) daily: the same 4 A069500 exit variants as EXP-004 and EXP-005
2. BNF C4 daily: the existing single baselines of EXP-002 (2024~) and EXP-003 (2005~)
3. Connors RSI(2) 5-minute: EXP-006's A005930 and A069500, 7 variants across tracks A/B
4. Each run records the trade ledger, costs, MDD, win rate, profit factor and equity curve.

Decision rules

• A strategy is not promoted on a single positive cell under the normalized actual assumption.
• Check that locked OOS and the 2x stress point the same way.
• If the ETF cost fix improves results, the earlier verdict is corrected from "strategy rejected" to "earlier cost contract rejected".
• If, as with BNF, target price, holding period and signal structure conflict even at lower costs, it is put on hold separately from the cost issue.

Run command

python scripts/backtest/run_cost_normalization_review.py
python -m pytest -q
python scripts/report/build_site.py