exp 14: enhanced stochastic-control allocation - workflow YAMLs (A: TopkDropout baseline, B: OptimalStopControlV2)

workflows/exp14-optstop-v2/a_topk_baseline.yaml

@@ -1,14 +1,15 @@
 # -----------------------------------------------------------------------------
-# EXP 15 - Strategy A (baseline): parallel reference model + TopkDropout.
+# EXP 14 - Strategy A (baseline): reference model + TopkDropout.
 #
-# Model = RankICEnsembleLGBModel with PARALLEL=5 (thread-pool seed training,
-# rank_ensemble.py) - the speedup means the 5-seed 3000-round ensemble trains in
-# ~1/5th the wall time of the serial reference. Strategy = TopkDropout topk=10
-# n_drop=2 risk_degree=0.95 (the reference's recorded strategy).
+# Model = RankICEnsembleLGBModel (5-seed RankIC-early-stopped LGB), the class
+# wired by the tac-rd-rank-ensemble-isolated reference (run 0cea66d9...).
+# Strategy = TopkDropout topk=10 n_drop=2 risk_degree=0.95 (the reference's own
+# recorded backtest strategy), so this run reproduces the reference baseline on
+# the same 50-ETF SP-5d panel.
 #
 # Run:
-#   rd_run_workflow config_path=experiments/workflows/exp15-kelly-size/a_topk_baseline.yaml \
-#       experiment_name=tac-rd-kelly-size
+#   rd_run_workflow config_path=experiments/workflows/exp14-optstop-v2/a_topk_baseline.yaml \
+#       experiment_name=tac-rd-optstop-v2
 # -----------------------------------------------------------------------------
 {%- set LAKE = TAC_LAKE_DIR %}
 {%- set UNIVERSE = "SPY,QQQ,DIA,IWM,MDY,VTI,VOO,VEA,VWO,VT,EFA,EEM,TLT,IEF,SHY,AGG,BND,LQD,HYG,JNK,EMB,GLD,SLV,USO,UNG,DBA,DBC,XLK,XLF,XLE,XLV,XLI,XLY,XLP,XLU,XLB,XLRE,ARKK,SMH,SOXX,IBB,XBI,ITA,XAR,ICLN,TAN,FDN,IGV,ESPO,REM" %}
@@ -42,7 +43,7 @@ qlib_init:
         module_path: qlib.workflow.expm
         kwargs:
             uri: "sqlite:///{{ LAKE }}/mlruns.db"
-            default_exp_name: "tac-rd-kelly-size"
+            default_exp_name: "tac-rd-optstop-v2"
 
 task:
     model:
@@ -63,7 +64,6 @@ task:
             reg_alpha: 0.1
             reg_lambda: 1.0
             seeds: "42,7,2026,99,123"
-            parallel: 5
 
     dataset:
         class: DatasetH

workflows/exp14-optstop-v2/b_optstop_v2.yaml

@@ -1,18 +1,21 @@
 # -----------------------------------------------------------------------------
-# EXP 15 - Strategy B: parallel reference model + KellyWeightStrategy.
+# EXP 14 - Strategy B (enhanced): reference model + OptimalStopControlV2.
 #
-# Model = RankICEnsembleLGBModel PARALLEL=5 (same as A). Strategy =
-# KellyWeightStrategy (tac_qlib.contrib.strategy.kelly_weight): applies the
-# Kelly criterion to position SIZING -
-#     f* = kelly_fraction * mu_i / var_i   (Gaussian Kelly, mu_i > 0)
-# with per-name mu/var estimated from the rolling signal history (no
-# lookahead), floored/capped and normalized to the risk_degree leverage budget.
-# This is the piece the reference's equal-weight TopkDropout never tunes: it
-# weights names by edge/risk instead of equal-weight top-k.
+# Model = RankICEnsembleLGBModel (5-seed RankIC-early-stopped LGB), identical to
+# Strategy A. Strategy = OptimalStopControlV2 (tac_qlib.contrib.strategy.
+# optimal_stop_v2) with the controls that address OptimalStopControl's documented
+# weaknesses:
+#   - turnover / cost control: rebalance_band=0.05 (skip small rebalances),
+#     cooldown_days=3 (no whipsaw re-entries), max_turnover=0.30 (cap daily
+#     traded notional, priority exits > opens > rebalances)
+#   - robust thresholds (no valid-window overfit): entry 0.85 / exit 0.70 /
+#     max_hold 10 / min_hold 2 / sl -0.08
+# Sizing = equal-weight control (risk_degree fraction of total value split
+# across targets) - the "proper allocation" that replaces cash-heuristic sizing.
 #
 # Run:
-#   rd_run_workflow config_path=experiments/workflows/exp15-kelly-size/b_kelly_weight.yaml \
-#       experiment_name=tac-rd-kelly-size
+#   rd_run_workflow config_path=experiments/workflows/exp14-optstop-v2/b_optstop_v2.yaml \
+#       experiment_name=tac-rd-optstop-v2
 # -----------------------------------------------------------------------------
 {%- set LAKE = TAC_LAKE_DIR %}
 {%- set UNIVERSE = "SPY,QQQ,DIA,IWM,MDY,VTI,VOO,VEA,VWO,VT,EFA,EEM,TLT,IEF,SHY,AGG,BND,LQD,HYG,JNK,EMB,GLD,SLV,USO,UNG,DBA,DBC,XLK,XLF,XLE,XLV,XLI,XLY,XLP,XLU,XLB,XLRE,ARKK,SMH,SOXX,IBB,XBI,ITA,XAR,ICLN,TAN,FDN,IGV,ESPO,REM" %}
@@ -46,7 +49,7 @@ qlib_init:
         module_path: qlib.workflow.expm
         kwargs:
             uri: "sqlite:///{{ LAKE }}/mlruns.db"
-            default_exp_name: "tac-rd-kelly-size"
+            default_exp_name: "tac-rd-optstop-v2"
 
 task:
     model:
@@ -67,7 +70,6 @@ task:
             reg_alpha: 0.1
             reg_lambda: 1.0
             seeds: "42,7,2026,99,123"
-            parallel: 5
 
     dataset:
         class: DatasetH
@@ -117,17 +119,20 @@ task:
           kwargs:
               config:
                   strategy:
-                      class: KellyWeightStrategy
-                      module_path: tac_qlib.contrib.strategy.kelly_weight
+                      class: OptimalStopControlV2
+                      module_path: tac_qlib.contrib.strategy.optimal_stop_v2
                       kwargs:
                           signal: "<PRED>"
                           topk: 10
-                          lookback: 20
-                          min_obs: 10
-                          kelly_fraction: 0.5
-                          max_weight: 0.15
-                          min_weight: 0.0
+                          entry_pct: 0.85
+                          exit_pct: 0.70
+                          max_hold_days: 10
+                          min_hold_days: 2
+                          sl: -0.08
                           risk_degree: 0.95
+                          notional: 20000
+                          rebalance_band: 0.05
+                          cooldown_days: 3
                           max_turnover: 0.30
                   backtest:
                       start_time: 2026-01-04