NG base-spread и BR regime filter: walk-forward
Предварительный разбор публикации · публикация 2026-07-05
Паспорт проверки и источников
BR walk-forward much weaker than fixed-fit result; NG only 13 trades. Verify actual contracts and costs.
Время изменения файла — не время последней рыночной котировки. Статус публикации не подтверждает работу бота; research, paper и live требуют отдельных результатов.
- gas_research/output/br_regime_arena_paper_state.json — 2026-07-20T04:37:06+00:00
- gas_research/output/br_regime_arena_paper_log.jsonl — 2026-07-20T04:32:05+00:00
- gas_research/output/spud_calendar_spread_scan.json — 2026-07-05T07:21:15+00:00
- gas_research/output/spud_calendar_spread_scan_5m.json — 2026-07-05T07:22:59+00:00
- gas_research/output/ng_base_spread_filter_search.json — 2026-07-05T07:31:21+00:00
- gas_research/output/ng_base_spread_filtered_walkforward.json — 2026-07-05T07:47:58+00:00
- gas_research/output/ng_base_spread_filtered_walkforward_trades.csv — 2026-07-05T07:47:58+00:00
- gas_research/output/br_regime_walkforward.json — 2026-07-05T07:43:39+00:00
- gas_research/output/br_regime_candidate_analysis.json — 2026-07-05T07:43:05+00:00
- gas_research/output/br_regime_candidate_fee2_trades.csv — 2026-07-05T07:43:05+00:00
Свежий прогон отделил рабочие направления от слабых: календарные Spud-спреды NG/BR пока не проходят, NG NYMEX-vs-MOEX base-spread дает редкие сигналы, а BR regime filter через BR*USDRUBF выглядит сильнее для forward paper.
- BR fixed regime candidate57 test trades, +41.21%, PF 4.42, max DD 2.93% after 2 bps fee
- BR monthly walk-forward38 trades, +13.60%, PF 1.79, max DD 10.11%; less smooth than fixed candidate
- NG filtered walk-forward13 trades, +3.69%, PF 7.77, max DD 0.38%; too rare for live, keep watchlist
- Calendar spread SpudNG calendar spreads did not pass; BRN6-BRQ6 5m has only a weak 43-trade candidate
The research now treats the spread as the base asset instead of forcing outright futures logic everywhere. A new scanner builds NG and BR calendar-spread OHLC series and runs Spud stochastic threads on the spread itself.
Result: pure calendar-spread Spud is not the main edge. NGM6-NGN6 and NGN6-NGQ6 did not survive train/test filters. BRN6-BRQ6 has a small 5-minute momentum candidate, but the forward sample is too small.
The stronger NG idea remains NYMEX NG versus MOEX NG base-spread mean reversion. After monthly walk-forward selection with hour/Spud filters it is profitable but very rare: 13 trades over the test folds.
The strongest oil candidate is BR directional trading gated by a negative BR*USDRUBF regime. The fixed candidate br_usd_z_24 <= -2, sig_lb=1, sig_th=0.005, horizon=3 has the best current research profile, but it should be moved to forward paper before any live use.
Decision: promote BR regime filter to forward paper monitoring; keep NG base-spread as a rare alert/watchlist; reject calendar Spud spreads as standalone live candidates for now.
Charts
Файлы и источники в проекте
gas_research/br_regime_arena_paper_bot.pygas_research/start_br_regime_arena_paper.ps1gas_research/output/br_regime_arena_paper_state.jsongas_research/output/br_regime_arena_paper_log.jsonlgas_research/spud_calendar_spread_scan.pygas_research/ng_base_spread_filter_search.pygas_research/ng_base_spread_filtered_walkforward.pygas_research/output/spud_calendar_spread_scan.jsongas_research/output/spud_calendar_spread_scan_5m.jsongas_research/output/ng_base_spread_filter_search.jsongas_research/output/ng_base_spread_filtered_walkforward.jsongas_research/output/ng_base_spread_filtered_walkforward_trades.csvgas_research/output/br_regime_walkforward.jsongas_research/output/br_regime_candidate_analysis.jsongas_research/output/br_regime_candidate_fee2_trades.csv