Skip to content

Pennant research findings

Append-only record of substantive findings produced by Pennant tests. A finding here is a claim about pennant behavior or strategy performance that future work should treat as established (until and unless superseded). Each finding cites its origin test by ID and its evidence by report path.

Findings are not opinions, recommendations, or hypotheses — they are empirical claims with a specific population, configuration, and window behind them. The "Caveats" column captures the scope of applicability so the finding is not over-generalized.

ID Date Finding Origin Evidence Caveats
F-001 2026-05-11 Tightening pennant duration to 10–20 bars + flagpole duration to 1–5 bars (V1) removes 71% of baseline event volume without improving the hit-rate at +15% MFE. Edge per event is roughly flat; the variant trades raw count for noise reduction without delivering a per-event quality lift. PEN-TEST-001 reports/Pennant/pennant_criteria_ab_test_2026-05-11.md Baseline + V1 only. Cohort-level finding measured at the +15% MFE threshold; does not address strategy-level P&L.
F-002 2026-05-11 V2 detection parameters (pennant 7–17, flagpole 1–5) keep 39% of baseline volume with the +15% MFE hit-rate +0.4pp over baseline. V2 is the sweet spot in the duration × flagpole-tightness sweep — neither the loosest (Baseline) nor the strictest (V4) variant matches it on per-event quality. PEN-TEST-002 reports/Pennant/pennant_criteria_ab_test_v2_2026-05-11.md Measured against the Phase 11a A/B baseline at the +15% MFE hit-rate. Other thresholds may rank differently.
F-003 2026-05-11 V4 (flagpole.max = 2 bars) trims event volume too aggressively. Despite the cleanest entries on flagpole geometry, the 30-day endpoint mean is below baseline — the cohort that's left after the cut underperforms unconditionally. Flagpole.max = 3 (V3) is the practical floor. PEN-TEST-003 reports/Pennant/pennant_criteria_ab_test_v3_v4_2026-05-11.md Per the Phase 11a-3 cohort; results may vary with different combined parameter sets.
F-004 2026-05-11 Under PEN-STOCK-001 (stock-only scaled exit, $500 fixed sizing, $10K start), Baseline maximizes raw return ($40,398, +304%, 7.50% CAGR) by virtue of trade count alone. V2 maximizes risk-adjusted return (Sharpe 0.617 vs Baseline's 0.543) with about half the drawdown (–21.4% vs –38.1%). Across the five detection variants, the trade count × average-profit lever moves the equity curve; per-trade win rate does not (43.7–45.4% range, 1.7pp spread). PEN-TEST-004 reports/Pennant/pennant_strategy_backtest_2026-05-11.md Stock-only at fixed $500 per trade. Options or %-equity sizing would change the trade-count-vs-quality balance.
F-005 2026-05-12 The unconditional baseline pennant population (15,528 events, 2007–2026) is near-symmetric in outcomes: 41% reach +10% MFE within 30 days, 36% reach –10% MAE within 30 days. Median MFE +7.5% vs median MAE –6.6%. The detector produces a balanced distribution at the population level; any edge comes from selection or trade mechanics, not from the unconditional mean. PEN-TEST-005 reports/Pennant/pennant_population_behavior_2026-05-12.md Baseline detection only; tighter detection variants (V1–V4) skew the distribution. Pre-friction outcomes.
F-006 2026-05-12 The median pennant winner gives back 68% of its peak by day 30; the median pennant loser recovers 88% of its trough by day 30. Pennants are bursts that mean-revert, not sustained-trend continuation patterns. The give-back / recovery symmetry is one of the strongest signals in the population data for trailing-stop design. PEN-TEST-005 reports/Pennant/pennant_population_behavior_2026-05-12.md Endpoint-vs-peak basis; a full intraday-trajectory analysis could refine the day-by-day give-back curve.

Format

  • ID = F-NNN, sequential and append-only.
  • Date = date the finding was first recorded (usually = origin test date).
  • Finding = the claim itself, written as a falsifiable empirical statement.
  • Origin = PEN-TEST-NNN (one or more).
  • Evidence = relative path(s) to the canonical report(s).
  • Caveats = scope constraints. A finding without caveats is almost always wrong about something.

Superseding a finding

A finding can be superseded by a later one. Mark the original's status with a footnote (e.g., F-003 ¹ superseded by F-NNN on YYYY-MM-DD). Do not delete the row.