fix KeyError: directional-mode diagnostics (no more entered/hedged/exit_type columns)

train.py

@@ -229,30 +229,24 @@ def _build_training_dataset(
     if len(df) == 0:
         raise RuntimeError("No valid training rows produced")
     df = df.sort_values("slug_ts").reset_index(drop=True)
-    # Policy-simulation diagnostics (the "real" picture)
+    # Directional-mode diagnostics
+    import numpy as _np
     if "realized_pnl" in df.columns:
-        import numpy as _np
         pnls = df["realized_pnl"].to_numpy()
         pcts = _np.percentile(pnls, [1, 5, 25, 50, 75, 95, 99])
-        log(f"[data] realized_pnl percentiles "
+        log(f"[data] realized_pnl(buy_up_@120) percentiles "
             f"1%={pcts[0]:.2f} 5%={pcts[1]:.2f} 25%={pcts[2]:.2f} "
             f"50%={pcts[3]:.2f} 75%={pcts[4]:.2f} 95%={pcts[5]:.2f} 99%={pcts[6]:.2f}")
-        log(f"[data] realized pnl buckets: "
-            f">$20={float((pnls>20).mean()):.3f} "
-            f">$0={float((pnls>0).mean()):.3f} "
-            f"=0(never_entered)={float((pnls==0).mean()):.3f} "
-            f"<0={float((pnls<0).mean()):.3f}")
-        entered_rate = float(df["entered"].mean())
-        if df["entered"].sum() > 0:
-            hedge_rate_given_entry = float(df.loc[df["entered"] == 1, "hedged"].mean())
-        else:
-            hedge_rate_given_entry = 0.0
-        log(f"[data] policy fire rates: entered={entered_rate:.3f} "
-            f"hedged|entered={hedge_rate_given_entry:.3f}")
-        if "exit_type" in df.columns:
-            vc = df["exit_type"].value_counts(normalize=True).to_dict()
-            log(f"[data] exit_type breakdown: "
-                + " ".join(f"{k}={v:.3f}" for k, v in vc.items()))
+    if "up_ask_120" in df.columns:
+        ua = df["up_ask_120"].to_numpy()
+        da = df["dn_ask_120"].to_numpy()
+        log(f"[data] up_ask_120: median={_np.median(ua):.3f} "
+            f"5%={_np.quantile(ua, 0.05):.3f} 95%={_np.quantile(ua, 0.95):.3f} | "
+            f"dn_ask_120: median={_np.median(da):.3f}")
+        cheap_either = float(((ua <= 0.40) | (da <= 0.40)).mean())
+        cheap_both_lo = float(((ua <= 0.35) | (da <= 0.35)).mean())
+        log(f"[data] entry opportunity: "
+            f"some_side≤$0.40={cheap_either:.3f} some_side≤$0.35={cheap_both_lo:.3f}")
     # re-sort eff_cache to mirror df row order (df was just sorted by slug_ts)
     slug_to_idx = {c["slug"]: i for i, c in enumerate(eff_cache)}
     reordered = [eff_cache[slug_to_idx[s]] for s in df["slug"].tolist() if s in slug_to_idx]