Post-market dummy refinement: 2026-07-29

[github-actions]

Goal

Assign this issue to Copilot as a daily post-market refinement task.

Review today's dummy-trading behavior, identify only evidence-backed improvements, and propose a small safe change if the data justifies it. Do not add strategy parameters randomly.

Latest Paper P&L

• Paper date: 2026-07-29
• Portfolio value: ₹500000.0
• Today P&L: ₹0.0 (0.0%)
• Total P&L: ₹0.0 (0.0%)
• Alpha vs NIFTY: -0.048%
• Positions: 0
• Trades in latest paper run: 0

Incubation Gate

• Decision: incubating
• Days: 8/20
• Blocking checks: 8/11
• Failed checks: min_trading_days, positive_dummy_alpha, backtest_oos_positive

Live Market Snapshot

• Live snapshot IST: 2026-07-29T15:26:49.348154+05:30
• Market open flag: true
• Live day P&L: ₹0.0 (0.0%)
• Live actions: 0
• Stops: 0

Latest Dummy Transactions

• No EOD paper trades in latest run

Daily Analysis Context

• Trading date: 2026-07-29
• Signal counts: BUY 41, HOLD 9, SELL 0
• Market regime: NIFTY trend filter is risk-off

Backtest Lab Context

• momentum_only_weekly: return 9.8%, alpha 14.97%, maxDD -7.89%
• momentum_weekly_churn_control__news_on: return 9.8%, alpha 14.97%, maxDD -7.89%
• momentum_weekly_churn_control__news_off: return 9.8%, alpha 14.97%, maxDD -7.89%

Copilot Instructions

• Inspect the dashboard/data outputs and current strategy code.
• Look for repeated or material failure patterns: stale data, bad churn, weak ranking, poor risk behavior, oversized drawdown, or confusing UI.
• If there is no strong evidence for a code change, say so and leave the system unchanged.
• If a change is justified, keep it small, testable, and paper-only.
• Preserve the one-month incubation rule. Do not recommend real-money trading.
• Include validation steps and expected effect on the readiness gate.

[claude]

No code change today — recurring flat-book pattern, plus one arithmetic (not code) point worth recording

Reviewed today's snapshot (2026-07-29) against the refinement rules. Conclusion: no evidence-backed change is warranted.

What was checked

1. 0 trades / 0 positions despite BUY 41, HOLD 9, SELL 0. Root cause is unchanged from Post-market dummy refinement: 2026-07-21 #23/Post-market dummy refinement: 2026-07-22 #26–Post-market dummy refinement: 2026-07-28 #30: USE_MARKET_REGIME_GUARD in src/paper_trader.py blocks all new buys whenever src/market_regime.py::current_regime() reports the NIFTY trend filter as risk-off, regardless of signal count; with 0 existing positions there's nothing for stops to sell either. market_regime.py and paper_trader.py have not changed since commit 8d5e23c (well before this streak began), so this is the same already-diagnosed, test-covered behavior (test_market_regime_risk_off_blocks_new_buys), not a new pattern. That regression test still lives on still-open PR Test: lock in risk-off regime guard blocking new buys #24 (auto-refine/issue-23-2026-07-21, open since 2026-07-21) — merging it remains the right way to close out this thread rather than opening a duplicate PR.

2. Incubation gate regressed from 9/11 to 8/11 blocking checks: positive_dummy_alpha now fails (alpha vs NIFTY -0.048%, vs +1.04% yesterday, +1.00%/+1.94%/+1.52%/+0.999%/+0.21% on the days before that). Traced this to src/incubation_report.py::_recompute_totals (line 113): alpha_pct = total_pnl_pct - benchmark_pct. Because the book has taken 0 trades since inception, total_pnl_pct is pinned at 0%, so alpha_pct is mechanically just -benchmark_pct — i.e. the negative of NIFTY's own cumulative return since inception. The positive alpha on days 2–7 simply reflected NIFTY being down since inception; today's small negative alpha reflects NIFTY ticking to a small net positive since inception. This is arithmetic, not a strategy decision or a bug — a flat, all-cash paper book will always show alpha = -NIFTY's cumulative move for as long as it holds no positions. min_trading_days and backtest_oos_positive continue to fail by construction this early (day 8/20; backtest_oos_positive needs more walk-forward folds).

3. Backtest lab: momentum_only_weekly and both momentum_weekly_churn_control news-on/off variants are still byte-identical (9.8% return / 14.97% alpha / -7.89% maxDD). Consistent with Post-market dummy refinement: 2026-07-21 #23/Post-market dummy refinement: 2026-07-22 #26–Post-market dummy refinement: 2026-07-28 #30: ablation_verdict() in src/news_ablation.py returns insufficient_data while the archived news history hasn't accumulated enough days yet. Self-resolves as the archive grows; not a defect in the recently-merged news-ablation code.

4. Test suite: python -m unittest discover -s tests -p "test_*.py" → 82 collected, 1 import error for test_news_archive in this sandbox (ModuleNotFoundError: No module named 'requests', pinned in requirements.txt but absent here — environment-only, matches Post-market dummy refinement: 2026-07-21 #23/Post-market dummy refinement: 2026-07-22 #26–Post-market dummy refinement: 2026-07-28 #30). With requests installed: 87 tests, OK (12 skipped).

Why no change

Point 1 is a re-observation of an already-diagnosed, test-covered pattern. Point 2 is new (the gate flipped a check) but traced to pure arithmetic on a flat book, not a defect — "fixing" it would mean either faking trades or changing the alpha formula to hide real NIFTY moves, which is exactly the kind of unjustified, non-evidence-backed change the rules warn against. Point 3 is unchanged and expected to self-resolve as the news archive accumulates history. Per the daily-refinement rules, "no change" is correct when the data doesn't clearly justify one, including when a metric moves for a well-understood mechanical reason rather than a code fault.

Closing this issue; PR #24 remains the tracked follow-up for the recurring risk-off observation and should be merged when a maintainer has a chance to review it.