In-Sample Discovery ArchiveNow Live: Dynamic ShadowForward-Validating2026-07-09T12:24:34+00:00 UTC
IN-SAMPLE DISCOVERY RESULTS. This page documents the backtest results on the
canonical feature panel (May 25-Jun 5 2026, 928 markets) that led to this candidate's deployment.
The KISS distance_only logit model used by some strategies was trained on a subset of this same
panel, so these in-sample figures are not forward-looking and cannot be used to infer live
performance. This candidate was deployed as a live dynamic-shadow paper monitor on 2026-06-13
(fresh inception, no backfill) --
see the live deep dive
for current forward fires and realized PnL.
patient_kiss_3ct_fallback_tte10
Waits for a 3-cent NO price improvement from TTE=14 anchor price before entering. Falls back to TTE=10 entry if the improvement threshold is never met. Uses KISS distance_only logit edge >= 0.07. Best in-sample dynamic candidate on canonical panel.
PnL q=10 (IS)
+233.01
Trades
636
Win Rate
47.3%
Max Drawdown
-109.5
Sharpe (ann)
10.16
Sortino (ann)
14.65
Strategy Description
Entry logic: TTE=14 anchor; enter on 3ct improvement; fallback TTE=10 Entry TTEs: TTE 14-10 (includes TTE=10 fallback) Median entry TTE: 13 min Side: NO (buy NO contract) Quantity: q=10 contracts (backtest convention) Model: KISS distance_only logit (2-feature: distance_to_strike_bps, abs_distance_to_strike_bps)
Comparison vs Benchmarks (In-Sample)
All figures below are in-sample on the canonical panel. Not live realized PnL.
mean_daily / std_daily * sqrt(365); in-sample only
Sortino (annualized)
14.65
mean_daily / downside_std * sqrt(365); in-sample only
Best day
84.80
Highest single-day PnL
Worst day
-54.80
Worst single-day PnL
PnL excl. best day
148.21
Robustness check: not one-day-dependent
Avg entry price
0.4366
Mean NO VWAP at entry
Median entry price
0.3900
Median NO VWAP at entry
Promotion verdict
PROMOTE_CANDIDATE
Gate: beat KISS, DD ok, positive excl best day, no overfitting
Deployment Status
The figures below are in-sample, from the same panel used to discover KISS thresholds. The KISS model was trained on this panel, so in-sample improvement over KISS does not by itself imply forward edge. This candidate was deployed as a live dynamic-shadow paper monitor on 2026-06-13 (fresh inception, no backfill) to begin forward validation -- see the live deep dive for current fires and realized PnL. Promotion to champion/benchmark status still requires accumulating sufficient forward (OOS) trading days.
Forward (OOS) Gate for Champion/Benchmark Graduation
Gate: Deployed as a live dynamic-shadow paper monitor since 2026-06-13 (fresh inception, no backfill). Champion/benchmark graduation still requires sustained forward (OOS) outperformance over more trading days -- in-sample results below predate deployment and are not validated alpha. What to check forward: Total PnL vs KISS on forward (OOS) days; drawdown vs KISS; win rate stability;
no evidence of regime-specific overfit (check performance across BTC up/flat/down days separately). Candidate registry:reports/consistent_panel_dynamic_yesno/shadow_strategy_registry_proposed.csv
Panel and Split Used
Panel: feature_panel_v2.parquet (local only, gitignored) Date range: May 25-Jun 5 2026 Markets: 928 (all markets with valid TTE=14 snapshot) Alignment: btc_left_forward_causal, 5s tolerance Columns: 254 (158,601 rows total across all TTE snaps) Split used for evaluation: Full 928-market pool (same as all published benchmark results) ML split: Documented separately (train/val/retro-test); NOT applied to strategy evaluation here Script:scripts/research/run_consistent_dynamic_yesno_v1.py