exp 15: fix KellyWeightStrategy - subclass WeightStrategyBase (order_generator), add topk gate; sync venv

workflows/exp15-kelly-size/a_topk_baseline.yaml

@@ -1,15 +1,14 @@
 # -----------------------------------------------------------------------------
-# EXP 14 - Strategy A (baseline): reference model + TopkDropout.
+# EXP 15 - Strategy A (baseline): parallel reference model + TopkDropout.
 #
-# 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.
+# 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).
 #
 # Run:
-#   rd_run_workflow config_path=experiments/workflows/exp14-optstop-v2/a_topk_baseline.yaml \
-#       experiment_name=tac-rd-optstop-v2
+#   rd_run_workflow config_path=experiments/workflows/exp15-kelly-size/a_topk_baseline.yaml \
+#       experiment_name=tac-rd-kelly-size
 # -----------------------------------------------------------------------------
 {%- 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" %}
@@ -43,7 +42,7 @@ qlib_init:
         module_path: qlib.workflow.expm
         kwargs:
             uri: "sqlite:///{{ LAKE }}/mlruns.db"
-            default_exp_name: "tac-rd-optstop-v2"
+            default_exp_name: "tac-rd-kelly-size"
 
 task:
     model:
@@ -64,6 +63,7 @@ task:
             reg_alpha: 0.1
             reg_lambda: 1.0
             seeds: "42,7,2026,99,123"
+            parallel: 5
 
     dataset:
         class: DatasetH

workflows/exp15-kelly-size/b_kelly_weight.yaml

@@ -1,21 +1,18 @@
 # -----------------------------------------------------------------------------
-# EXP 14 - Strategy B (enhanced): reference model + OptimalStopControlV2.
+# EXP 15 - Strategy B: parallel reference model + KellyWeightStrategy.
 #
-# 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.
+# 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.
 #
 # Run:
-#   rd_run_workflow config_path=experiments/workflows/exp14-optstop-v2/b_optstop_v2.yaml \
-#       experiment_name=tac-rd-optstop-v2
+#   rd_run_workflow config_path=experiments/workflows/exp15-kelly-size/b_kelly_weight.yaml \
+#       experiment_name=tac-rd-kelly-size
 # -----------------------------------------------------------------------------
 {%- 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" %}
@@ -49,7 +46,7 @@ qlib_init:
         module_path: qlib.workflow.expm
         kwargs:
             uri: "sqlite:///{{ LAKE }}/mlruns.db"
-            default_exp_name: "tac-rd-optstop-v2"
+            default_exp_name: "tac-rd-kelly-size"
 
 task:
     model:
@@ -70,6 +67,7 @@ task:
             reg_alpha: 0.1
             reg_lambda: 1.0
             seeds: "42,7,2026,99,123"
+            parallel: 5
 
     dataset:
         class: DatasetH
@@ -119,20 +117,17 @@ task:
           kwargs:
               config:
                   strategy:
-                      class: OptimalStopControlV2
-                      module_path: tac_qlib.contrib.strategy.optimal_stop_v2
+                      class: KellyWeightStrategy
+                      module_path: tac_qlib.contrib.strategy.kelly_weight
                       kwargs:
                           signal: "<PRED>"
                           topk: 10
-                          entry_pct: 0.85
-                          exit_pct: 0.70
-                          max_hold_days: 10
-                          min_hold_days: 2
-                          sl: -0.08
+                          lookback: 20
+                          min_obs: 10
+                          kelly_fraction: 0.5
+                          max_weight: 0.15
+                          min_weight: 0.0
                           risk_degree: 0.95
-                          notional: 20000
-                          rebalance_band: 0.05
-                          cooldown_days: 3
                           max_turnover: 0.30
                   backtest:
                       start_time: 2026-01-04