research: markout_flow — exploratory markout-exit curve (VERDICT: no scalp edge; hold wins everywhere)

Forward days, 559 resolved fills: exit +60s = -$3.43/fill (burst-top entry
is briefly underwater), +30m = +$13.63, +2h = +$35.51 vs hold-to-resolution
+$43.56. Monotone accrual to resolution in every slice — the under-reaction
converges over hours, not seconds. v1's >3h 'bleed bucket' is POSITIVE on
the full signal (+$44.66/fill): another confirmation the v1 book's losses
were cash-gate selection, not hold-time. All entry bands positive forward.
No A2-scalp pre-registration warranted; markout re-reads stay on to verify
with real bids.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

research/markout_flow.py

@@ -0,0 +1,112 @@
+#!/usr/bin/env python3
+"""EXPLORATORY (2026-07-22, user ask) — markout-exit curve for the surge
+signal: does the in-play under-reaction edge peak early and decay, i.e.
+should Study A2-scalp EXIT at a fixed horizon instead of holding to
+resolution? Motivated by the v1 book post-mortem: >3h holds bled −21.9%
+while <3h holds were +4.1%.
+
+NOT pre-registered — output shapes a possible A2-scalp pre-registration,
+nothing more. Prints-based v0: exits are marked at the LAST PRINT <= t+H
+(sim.markout), which is optimistic vs hitting the real bid — the live
+harnesses now record best bid/ask at +60/300/1800s per fill precisely to
+haircut this in v1 of the study. Exit fee charged same as entry.
+
+Method: per tape day — informed set as-of 00:00 UTC (frozen method),
+signals() with frozen params (flow>=$300/60s, band 10-90c, cooldown 900s),
+worst-print $100 entries at p50 lag / calibrated hold. For each RESOLVED
+fill: hold-to-resolution PnL (res_tok) vs exit PnL at each horizon.
+Coverage (no print in window => no exit) reported per horizon."""
+import json
+import os
+import time
+
+import tape
+import sim as simmod
+import study_flow as sf
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+HORIZONS = (60, 300, 1800, 7200)
+FEE = simmod.FEE_RATE
+
+
+def day_bounds(d):
+    lo = time.mktime(time.strptime(d, "%Y-%m-%d")) - time.timezone
+    return lo, lo + 86400
+
+
+def main():
+    db = tape.connect()
+    cal = json.load(open(os.path.join(HERE, "params", "sim_calibration.json")))
+    fz = json.load(open(os.path.join(HERE, "params", "study_flow.json")))["frozen"]
+    t_min, t_max = db.execute("SELECT min(ts), max(ts) FROM trades").fetchone()
+    days = []
+    t = t_min
+    while t < t_max:
+        days.append(time.strftime("%Y-%m-%d", time.gmtime(t)))
+        t += 86400
+    tape.build_resolved(db)
+    payout = {a: (p, lts) for a, p, lts in db.execute(
+        "SELECT asset, payout::DOUBLE, last_ts FROM res_tok").fetchall()}
+    sim = simmod.Sim(db, lag_s=simmod.LAG_P50, hold_s=cal["hold_s"],
+                     fill="worst")
+    rows = []
+    for d in days:
+        lo, hi = day_bounds(d)
+        hi = min(hi, t_max)
+        S = sf.informed_set(db, lo, fz["top_n"])
+        trig = sf.signals(db, S, lo, hi, fz["window_s"], fz["flow_usd"])
+        forward = d >= "2026-07-21"          # after the 07-20 freeze
+        for t_ in trig:
+            r = sim.try_buy(t_["asset"], t_["ts"], t_["p_ref"],
+                            stake_usd=sf.STAKE)
+            if not r["filled"]:
+                continue
+            pay = payout.get(t_["asset"])
+            if pay is None:
+                continue                     # resolution leg required
+            row = {"day": d, "fwd": forward, "px": r["price"],
+                   "shares": r["shares"], "efee": r["fee"],
+                   "hold_s": pay[1] - r["fill_ts"],
+                   "hold_pnl": r["shares"] * (pay[0] - r["price"]) - r["fee"],
+                   "win": pay[0] == 1.0, "mo": {}}
+            for H in HORIZONS:
+                m = sim.markout(t_["asset"], r["fill_ts"], H)
+                if m is None:
+                    continue
+                xfee = FEE * r["shares"] * min(m, 1 - m)
+                row["mo"][H] = r["shares"] * (m - r["price"]) - r["fee"] - xfee
+            rows.append(row)
+        print(f"{d}: {len(trig)} triggers scored")
+
+    def report(tag, rs):
+        if not rs:
+            return
+        n = len(rs)
+        hold = sum(r["hold_pnl"] for r in rs)
+        print(f"\n== {tag} — {n} resolved fills · hold-to-resolution "
+              f"EV/fill {hold/n:+.2f} · hit {sum(r['win'] for r in rs)/n:.2f}")
+        for H in HORIZONS:
+            sub = [r for r in rs if H in r["mo"]]
+            if not sub:
+                continue
+            mo = sum(r["mo"][H] for r in sub)
+            hold_sub = sum(r["hold_pnl"] for r in sub)
+            print(f"  exit +{H:>5}s: EV/fill {mo/len(sub):+7.2f} vs hold "
+                  f"{hold_sub/len(sub):+7.2f} on same {len(sub)} "
+                  f"({100*len(sub)/n:.0f}% coverage)")
+
+    fwd = [r for r in rows if r["fwd"]]
+    ins = [r for r in rows if not r["fwd"]]
+    report("IN-SAMPLE days (<= 07-20)", ins)
+    report("FORWARD days (>= 07-21)", fwd)
+    report("FORWARD · hold > 3h (the v1 bleed bucket)",
+           [r for r in fwd if r["hold_s"] > 10800])
+    report("FORWARD · hold < 3h", [r for r in fwd if r["hold_s"] <= 10800])
+    for lo_, hi_, tag in [(0, .3, "entry 0-30c"), (.3, .5, "entry 30-50c"),
+                          (.5, .7, "entry 50-70c"), (.7, .95, "entry 70-95c")]:
+        report(f"FORWARD · {tag}",
+               [r for r in fwd if lo_ <= r["px"] < hi_])
+
+
+if __name__ == "__main__":
+    main()