Same-rule semiconductor extension and fee correction V3

2026-09-08 JST. Frozen before inspecting new strategy outcomes. Research only; no exchange writes, account changes or schedules.

Authority and scope

The user's quoted successful reference is CASH_15m_BREAK_NEXT_DWM in the latest55-scenario roleflip-v2.1 study:77 trades,63.636% win,mean+0.063318R at an assumed6bps total cost. This is a cash-price reference, not an XYZ fill strategy or a profit guarantee. Use exact existing engine pivot216_roleflip_engine_v2.ts; preserve prior55 results. The earlier75-scenario JS engine is a different study and must not be mixed in. Inputs NVDA archive research:pivot216:NVDA:20260908. New data/results prefix research:pivot216:semiconductors:20260908:fee_v3. Target sampleEnd2026-09-08T00:30:00Z; exclude in-progress cash sessions/current daily data. Raw payloads, requests and hashes retained.

Rules unchanged

15m, ATR14 based on preceding closed bar,50-bar warm-up. Same D/W/M visible anchors and full-grid next target,0.25ATR strong-close buffer,0.50ATR body,upper/lower25% close; short mirrored. Same0.50ATR beyond-anchor SL,1ATR proximity for BREAK_NEAR,1h pending limit lifetime,6clockhour maximum hold,session-close liquidation for cash reference,stop-first within-bar ambiguity,no future or same-bar entry. Modes BREAK_NEXT_DWM, BREAK_NEAR, RETEST_LIMIT, REJECT_2, HYBRID, BREAK_NEXT_DP_LONG. No new volatility/RSI/regime filters, stop optimization or winner-selected changes. Keep the old entry reward>6bps selection gate even when rescoring fees, to preserve exact trade selection. It remains a buffer, NOT a claim that the actual fee is6bps. Separately label any future gate revision rather than silently changing the experiment.

Fees vs execution friction

User supplies taker0.0081% (0.81bps) and maker0.0027% (0.27bps) per fill. Baseline rescoring charges taker on entry AND exit using each actual modeled fill notional, not leverage/margin. Approx roundtrip1.62bps=0.0162%, but exact fee changes with exit/entry ratio. Alternative passive assumption: limit entry maker, other entry taker, TP maker, protective SL/time exits taker. This is a fee scenario, not verified queue/fill evidence. Show fees-only and additional aggregate spread/slippage2/4/8bps separately. Historical funding excluded unless separately acquired/validated. Applying HL rates to cash price paths is a transfer-reference cost model, not the broker's real equity commission/tax/stock-borrow fee.

Instruments and sessions

NVDA/MU/SNDK cash regular NY09:30-16:00; exclude Yahoo zero-volume terminal quote exactly as old NVDA. KRX000660.KS is SK hynix common stock (KRW), NOT SKHY ADR. Korea09:00-15:30 KST; preserve valid OHLC bars with zero reported volume (missing volume metadata is not a missing price; volume not used). KRX price-dependent tick per official rule; if all test prices>=500000KRW, tick1000KRW. USD perps MU/SNDK/SKHX price tick needs exact observed/API constraints and is configured separately (not blindly copied0.01). SKHX official oracle is000660KRW price divided by USD/KRW. Never overlay rawKRW levels on USD XYZ candles. CashKRW reference needs noFX conversion for dimensionless R/return. XYZ_UTC self-source is a separate fully local-price test. A cash-source-to-XYZ SKHX test may use explicitly labeled lagged-FX approximation only: each completed dailyOHLC divided by last observed FX before its market open, then previous completed D/W/M aggregation; never label that exact oracle/Pine parity. Include FX alignment/freshness caveat and verify no forward-looking FX input. Do not use today's FX to translate all past prices.

Comparisons

For each symbol run six modes on cash reference and on XYZ with cash-derived levels and with XYZUTC-derived levels where data/mapping support it. Explicitly label unavailable/approximate series. Reprice archived NVDA trades, verify reproduction of cited77 count/gross mean before any conclusions. Preserve native full histories and also compare a fixed common calendar window shared by all four symbols. No optimization on new symbols. Report counts,gross/fee-only/friction-adjusted meanR,bps,PF,net-win%,TP/SL/time,early70/late30,weekly-cluster1000-bootstrap intervals as exploratory, best-trade concentration, D/W/M/long-short/session diagnostics where useful. Do not claim full portfolio independence across semiconductor names or untouchedOOS for NVDA already observed histories.

Deliverables

Archive data/results/config/source separately from live trade ledger. Provide user-readable comparison/report and reusable scripts; export factual summary tables and fee formula audit. Preserve all failures and do not transform research into live orders.

Data-quality amendment before new strategy outcomes

Yahoo000660 regular-only metadata incorrectly truncates at15:00KST, whereas official KRX regular ends15:30. includePrePost=true successfully retrieves15:00-15:15 and15:30 auction prints. Restrict the extended feed to09:00-15:30KST; aggregate15m, assigning the15:30 closing auction print to the15:15-15:30 bar, not a new executable post-close bar. No synthetic15:20/15:25 prices. Exclude sessions with missing interior15m bins rather than fill prices; retain official prior cash dailyOHLC for pivots. Audit completion counts and all excluded dates. Current09-09:30 bars ofSep8 excluded from cash reference, which endsSep7KST. For common-window comparison use UTC2026-07-20T00:00 to2026-09-05T00:00 (aligned date range, not identical cross-time-zone hours). Price grid for new XYZMU/SNDK/SKHX is conservatively fixed0.10USD, legal but sometimes coarser than actual5-significant-digit grid under1000USD. CashMU/SNDK0.01USD, cashKRW1000, archivedNVDA0.01USD. These instrument-specific grids are disclosed; no invalid 0.01USD target prices on XYZ instruments priced above1000.