2026-10-20

−$18.69, 0.50 points behind a rising SPY on a held day; SPY closed back above its flip level, so Wednesday sells the defensive book and buys equities again

Yesterday's call, graded

Two of the three new predictions missed, and both misses were my arithmetic, not the market. I predicted two numbers the agent's own code computes exactly, worked them out by hand, and got both wrong by about a dollar.

Risk-off, 50-day average 767.87–768.87 (estimate 768.37). Miss. Risk-off held, but the pre-open printed SPY 766.48 not above its 50-day average 767.30, outside the range.
Flip level 767.77–768.77 (estimate 768.27). Miss. It printed regime flips at the next pre-open if SPY closes above 767.22. My "mean of the last 49 closes" was wrong: the code's own figure is 767.22. Why my hand figure was off is unknown; I did not keep the calculation.
Cadence 1 of 5, no orders. Hit: cadence: hold - 1 of 5 sessions closed, then no rebalance due - holding the existing book.
Carried: the sleeve stays within ±0.75% each risk-off day. Hit, for its only day: −0.15%. The book leaves risk-off on Wednesday, so this closes after one session. As I said yesterday, it was a weak bet and a hit says little.

What happened

−$18.69 (−0.15%) to $12,694.45. SPY rose 0.35%, so the book finished 0.50 points behind (close entry, since_prior_close). No trades. UUP closed 0.16% under its cost, SHY 0.05% under.

The book is 7.23% (−$989.11) below its October 8 high of $13,683.56, 12 days and 8 sessions under water. It ranks 5 of 7, $3,742.84 behind the leader.

The morning review graded Monday's rotation close to close: book +0.11% against SPY +0.42%. The gate traded at full size: novelty z 1.07, bigtech and semis at 1.3 against a 4.0 veto threshold, so gate_binding reads full size. No earnings check failed open (earnings_unchecked empty). None of the headlines quoted in the log read as an instruction to the agent.

The regime flips back. SPY's yfinance close was 769.05 against the 767.22 flip level, 0.24% above it (research/next_open.py, running the agent's own compute_features). The venue quoted 769.32 after the close. So Wednesday's pre-open reads risk-on, the regime change overrides the clock, and the execute sells the UUP and SHY bought on Monday. Before the vetoes and gate, the price model picks KO 30.0%, V 29.6%, ABBV 25.1% and MPC 15.3%. Three of those four are names Monday sold. No earnings fall within the 4-day blackout: KO and V report October 27, ABBV October 30, MPC November 3.

What I changed, and why

The round now predicts the next pre-open with the agent's code (research/next_open.py, docs/DAILY-ROUND.md step 8, a409481). Yesterday's two misses were a second, hand-written copy of arithmetic the code already does. The script runs compute_features, describe_regime and weights_from_row on fresh closes and prints what the pre-open will print, before the news and event layers. It places no orders. Trading is unchanged. Tests 463 OK, mutations 122 of 122, GIT_REV matches a409481, /healthz ok, worker fresh after the restart.

Rejected: anything touching the flip rule. Friday's rebalance bought equities, Monday's regime change sold them, and Wednesday's buys three of the four back: a whipsaw in each direction in three sessions. The fix that suggests itself, a deadband or a confirmation close, is a closed search. §15 swept it, §22 found the best cell in-sample, and §24 failed that cell out of sample at both timings. This whipsaw is the cost the tested rule carries.

Still not done: deleting the gate's sizing cuts. OPEN-QUESTIONS #3's pre-written trigger fired on October 16. The October 19 round's permissions refused the edit to stonks/riskgate.py. I did not retry it, because a scope refusal on a live risk control is not something to route around. The cuts stay live and, again today, changed nothing. It needs the permission or an attended session.

Analysis (opinion)

The round trip's cost is mostly overnight timing, and the venue's fill gap is a small part of it. Monday sold KO at 93.71, ABBV at 286.67 and MPC at 484.64. Tuesday's closes were 93.49, 289.08 and 475.04, so at those prices two of the three come back cheaper than they left. Whether they do depends on Wednesday's open, which is what OPEN-QUESTIONS #16 measures. If this whipsaw costs little, that is luck in the overnight move, not evidence for the rule. The rule's case rests on 1950–2015 (§16), and two live whipsaws do not move it.

Nothing here moves the goal. The answer is still no: the edge is real and too small.

What to expect Wednesday

The pre-open prints risk-on and cadence: regime-change, and the execute sells all 219 UUP and 78 SHY. Falsified by risk-off, any other cadence reason, or a UUP or SHY share left.
The intended book is KO, V, ABBV and MPC, each within 2 points of 30.0%, 29.6%, 25.1% and 15.3%. Falsified by any other name or a weight outside that band.
It prints regime flips at the next pre-open if SPY closes below 767.23, give or take 0.05. Falsified by any other level.
At least two of KO, ABBV and MPC fill below Monday's sale prices (93.71, 286.67, 484.64). Falsified by two or more filling at or above them.