The desk is the hand-traded SPX method with the rules the hand record showed were missing. Before the open it computes the floor pivots from the previous session's high, low and close: P in the middle, R1 to R3 above, S1 to S3 below (Barchart's "Pivot Points, Previous = Day"), and R4 / R5 / S4 / S5 beyond them. The read is the index against P: above P leans calls, below P leans puts.
The rules it enforces, each one taken from a loss in the October hand record: one entry a day in the 09:30 to 11:00 window, no re-entry after a loss, no adding to an open position, a stop at −50% of the fill (switchable off for holds), a time stop at 15:30, a daily loss lock at −$2,500. The record's entries before 09:30 lost $55k on four trades; its entries after a same-day loss won 11% of the time and lost $62k. The morning window won two of three and paid $3,500 a trade.
Autodesk is the desk left to its own devices, watched from one screen. A standing arm sets every session the same way by itself. From its start time (09:30 for the current plan) the trader watches SPX on every tick; the first time the index trades through one of the eleven trigger levels, or comes within half a point of one, the arm fires: it buys the strike nearest the index on the read's side, the size in the Order settings, at market.
The first look is the session's own print, never the prior close, because an opening gap would otherwise read as a touch of every level it jumped (that happened on 5 of 10 sessions in the backtest). The state word on the page says where it stands: ARMED before the start, WATCHING inside the window, IN while a position is held, DONE with the day's result, FLAT when no level was touched.
The hold: at the first pivot touch after 09:30, buy 10 SPX at the money, on the read's side, on the expiration about 14 days out, no stop, out at the 15:30 time stop. On the seven real sessions of Oct 21 to 30 this made +$107,917 at 10 contracts, 5 of 7 up, because the read was right 5 of 7. The same trade at 0DTE made +$83,017 and its two wrong days cost −$18,512 and −$15,112 against −$1,312 to −$6,712 further out: further out, a wrong day costs the move, not the whole premium.
Plus the written leg: one SPX of the opposite side sold at the same touch and bought back at 15:30 (see "Writing instead of buying" below). Both are on the paper account. Every number above is on real option prints, not a model.
Instead of buying the read's side, sell the opposite side's at-the-money option at the touch and buy it back at 15:30. On the seven real sessions, 10 SPY contracts: +$12,186 at 0DTE (7 of 7 up), +$7,066 to +$7,946 further out, against the bought hold's +$7,876 (5 of 7). The short keeps the time decay on the flat days the bought side loses, and earns less on the trend days.
The catch is the tail. On six months of real 0DTE prints the written read made +$149,172 (87 of 121 days up) against the bought's +$126,472 (46 up), but its losses are the big-move days: −$82,512 on 2026-06-09, a −2.4% session, marked −$202,750 at the low. The written side collects the premium and pays the crash, uncapped until 15:30. That is why it runs at one SPX contract next to the ten bought: one SPX short is the collateral of ten SPY, about $115k to $150k of margin.
The lock is the rule set enforced at execution, in the trader, not on the page: the window, one entry a day, no re-entry after a loss, no adding, the daily loss limit, a 20-contract cap. A cascade's re-buys and the armed pair (the hold plus the written leg) count as one entry. "Release the lock until the close" lifts every rule but the contract cap for the day, and every trade taken under it is marked as such.
Not reliably, and the literature is unambiguous. $100 to $300 a day on $25k is 100% to 300% a year. The one complete study of every day trader in a market (Brazil, 2013 to 2015, 19,646 people) found 97% of those who persisted lost money and the single best trader out of 20,000 averaged $310 a day with a daily swing eight times that. Taiwan's full-market data says under 1% are reliably profitable after fees.
The edges the literature still supports after publication are few and structural. Tested on our own data (doc 08): the overnight premium is real in both windows and every symbol (the market's return accrues close-to-open), but exploiting it halves the return; turn-of-month and pre-FOMC drift are gone; the 200-day moving average filter cuts drawdown at a return cost; vol-managed exposure fails out of sample. Nothing earns more than the market with less risk. The money in markets is the equity premium, market-making technology, fees on other people's assets and private information; only the first is open to us on equal terms.
Seven years, 14 ETFs, real daily bars. Buy-and-hold made more money on 11 of 14 symbols (+14.0% a year against +8.5%). The wheel's smoother path was just what holding 0.45 beta looks like, and holding 0.45 beta directly is smoother and cheaper. 58% of its P&L was share beta, 22% the options business (which lost on 5 of 14 symbols), 20% interest and dividends. Dropping the covered call recovered most of the gap and still lost. After tax the wheel loses to a two-trade index / T-bill blend on 12 of 14. Retired; its history is on the Archive page.
The idea: enter SPY with a cash-secured put, hold five years, sell a five-year call struck above the average return, since long-dated calls "seem badly priced". They are mispriced, the other way. A five-year call at 1.61× the index is priced as a 14% event; SPY has exceeded it in 47% of the 341 five-year windows since 1993. The seller is on the wrong side. The whole plan: −1.35% a year after tax on SPY, −3.35% on QQQ, negative at every strike and vol tested. The only era it wins is a start in 2000 to 2009, when the index went nowhere. Not built.
The lunar effect is a published anomaly (returns around new moons ran about double those around full moons, 25+ markets), so it was tested properly rather than dismissed: exact rotation nulls (a 29.5-day cycle has only about 30 distinct shifts, so the p-value floor is 0.033) and an equal-weight composite, since 11 of the 14 ETFs are one bet. Absent as a tradeable effect: the right sign in the two older windows, reversed in the newest, never significant. Mercury retrograde: a crash-coincidence artifact, 12 of 22 windows negative against 11 by chance. The study can rule out a large effect; it cannot resolve a small one.
The dollar and the 10-year move together every year (+0.04 to +0.41) because the dollar is priced off the rate differential. That is a co-movement, not a lead. Nothing predicts a month ahead in both windows: no 21-day cell clears p 0.05, the tightening sign is right at five days and gone at 21, and every regime overlay that beats buy-and-hold in one window loses in the other. "Skip tight regimes" improved the old put-spread ladder on one decade whose only tightening bear was 2022: shadowed, never switched on.
Its real entry logic replayed over 30 days of 1-minute bars on seven large caps, expressed through 30-day at-the-money options with its own exit ladder: 151 trades, a 59% win rate, −2% of premium a trade, profit factor 0.70. Losses were uniform across six of seven symbols and both directions, which is what zero edge before costs looks like. Its exit ladder guaranteed the average win was smaller than the average loss. Signal killed; the code was deleted on 2026-11-01.
The hand plan written as a machine: the open against P sets the side, buy the SPX 0DTE option priced nearest $3, 20 contracts, sell 18 at $4, let 2 run, a stop at $1.50, out at 15:30. On 250 sessions, model-priced along the real path: the open's position against the pivot does not predict the day (p 0.75 against its own shuffled labels; always-put and a coin both did better), oil's prior day adds nothing, the 10-year above 5% changes nothing. A $3 option sits 100 to 130 points out and needs a 67% hit rate to break even on that take / stop geometry. Re-priced at the real 0DTE vols the losses roughly halve and the verdict stands.
42 same-day positions over eleven sessions, transcribed from Robinhood: −$26,680 in all, but the first trade of the day made +$11,313 (45% win, $1,028 a trade) and the 09:30 to 11:00 entries made +$38,805 at a 64% win rate. Two things lost the month: four entries before the cash open (−$55,450, none won, including a $40,695 put at 09:05) and nineteen entries after a same-day loss (−$61,748, 11% win). Those became the desk's rules.
"The market most nights goes up; is there an opportunity starting 8:15 PM Sunday?" No. 56% of nights are up, but the average night is +1.6 index points with a 39-point standard deviation, Sundays no better. A $3 out-of-the-money option bought at the 20:15 open gives up a third to a half of its value by 09:15 on a flat market (the night carries about 0.3 of a day's variance and is priced for it) and the overnight spread takes another 8% a side. Every variant lost on every data set, profit factors 0.05 to 0.5.
Loses about $500 a day once the options are priced the way the market prices them at the open (the study found real 0DTE prints at the open run at 0.4 to 0.7 of the prior VIX close, two to three times cheaper than the model had assumed; that recalibrated every later study). A −30% to −50% stop brings it to break-even and no better. The play is a bet that 0DTE options at the open are cheap relative to the day's realised move; over a year, on average, they were not. "Sell 1 at some %" changes nothing: a runner that has doubled rarely triples.
Six months of Robinhood's own bars for the expired SPXW series (121 sessions, 1,936 contracts). Neither pays. Twenty of the ~$2.50 contract at 10:00, everything out at +40% or +60% with a −50% stop, loses about $800 a day with every direction rule: the 10:00 trend, the pivot read, both, the first-candle break, their opposites, always call, always put, random. They are indistinguishable from each other and from the shuffled null (p 0.56). The geometry decides: a 0DTE option crosses a +40% / −50% band inside the first hour nearly every day, the take fills on 45% of days and that band needs 62%. At the money sold with momentum loses less in dollars under every exit. If the hand record's edge is real it lives in the timing inside the window, which a fixed 10:00 entry cannot see.
0DTE spreads have no premium to sell. On the real prints, an iron condor sold at 10:00 and held to the settlement breaks even (+$1 a spread, 90% win) and every intraday stop or early take makes it a loser (−$200 to −$400 a day at the desk's risk). No sideways filter known by 10:00 beats random days. Three independent tests agree: the day's implied move under-priced the realized one by 7% over 2022 to 2026, the real prints broke even, and at 2026's real 0DTE vols every 0DTE cell lost over 2008 to 2026.
The one that paid 2006 to 2026 is the monthly put spread: about 30 days out, the short strike near 10 delta (8% below the index), a wing about 2% of the index, held to expiry: +1.98% of the risk a trade, 97% of trades keep the credit, 3 losing years of 21, the worst 2008 at −4.9% a trade. Calls, condors and butterflies do not pay at any tenor (Cboe's real condor index: −0.3% a year over T-bills). Taking 50% early or closing at 21 days cost money in every version; the wing is the stop. The desk has a scanner for it on the Manual page; trading it needs combo orders and a month-long lock, not built.
The 0DTE plan carried to 5, 10 and 20 days: a smaller losing trade with a different texture, fewer takes, fewer stops, a steady bleed instead of a daily swing. The $3 option at two weeks sits about 300 points out, a tail ticket. The read stays inside its null at every expiry. The later term studies (doc 21 §25 to §27, on real prints) found something different: the at-the-money hold, not the $3 option, is where a longer expiry helps, by shrinking the wrong day.
The 10-year, oil, ES, NQ and VIX on their last hour vote call or put; buy the at-the-money option at a pivot touch in that direction, +25% take, −40% stop. No indicator called the session better than a coin, alone or together, at 09:30 or 10:30; the plan loses $823 a day, below its own shuffled labels. This is the test the gradients topic below refers to.
Friday 2026-10-30's real SPY 0DTE minute prints, then ten sessions. The cascade buys at a pivot touch, sells at +5%, re-buys at the money at once, with a stop that widens a notch each profitable sale. On Friday's continuous drop: 43 sales, +$1,454 a contract. Friday was not typical. Over the ten sessions the cascades lost (−$221 options, −$414 written, −$174 shares) while their plain holds made +$651 / +$1,094 / +$969. How it loses: its 117 sales made +$885, the four stops on choppy days −$511, the four last positions held into the close on trend days −$420, fees −$175.
Every variant was tried on real bars: a +2% or +3% take (worse), the turning-point exit from Barchart's cheat sheet (helps the cascade, cuts the hold's trend days short), entries from 09:50 (helped), the grid, the scalper, the dip ladder, Robinhood's nickels and costs, 15-second bars. The common finding: the cascade earns the distance the option travels, not its wiggles; waiting for a pullback sits out the runs; seeing faster bars does not help; and on one strike, selling and buying straight back is the hold cut into pieces, less the fees. Nothing beat the plain hold robustly.
One at-the-money SPX put bought at the touch, sold +$0.20 and bought straight back, no stop: Friday +$2,762 as described, +$739 paying the spread and a tick, against the hold's +$6,339; on 10-29, the wrong day, −$1,239 against the hold's −$1,011. Never beat the hold at any expiry on either day. The same put at 0DTE, 2, 4, 7 and 11 weeks, 10 contracts, on every session with real bars: the hold was best at 2 to 4 weeks (+$108k to +$112k over seven sessions against +$83k at 0DTE) because the wrong days cost a tenth as much. That is the plan now running. On six months of real 0DTE prints the hold made +$126,472 but 46 of 121 days up: five sessions made +$324k and the rest lost −$198k. It is a trade that needs the trend days.
Doc 15 found the model priced 0DTE options two to three times too rich at the open. Every study since runs on real prints (Robinhood's bars for the expired series, SPXW's own minute bars, the desk's recorders) and a model is used only where no print exists, said so each time. Doc 21 §19 measured how the backtests themselves fill: every leg of a bar fills what it crosses, so a rule can trade up to eight times a minute on paper; the live desk sees a fill within its 5-second cycle. A 2008 to 2026 model of the term hold was built and then dropped on the operator's word: only real prints count. The recorders now keep every option tick and the tape so each new session adds real evidence.
Mario's model calls the day right about 62% of the time and he suggested ranking by gradients before we open a position. Here is what that means, with our own numbers.
Stand anywhere on a hill. The gradient at that spot tells you two things: which way is steepest, and how steep. Flat field, zero. Cliff edge, huge. That is the whole idea.
A model is a set of knobs. For any setting of the knobs you can measure how wrong its calls were. The gradient says, for each knob, "turn this one a little this way and the error drops fastest." Training is walking downhill one small step at a time until it stops improving. That walk is how a model arrives at something like 62%.
Once a model is trained, the same slope idea gets used two ways. Mario should say which he means.
Today Fenrir fires the arm on every first pivot touch, on the read's side, with no judgement about how strong the read is. A gradient ranking would turn that into "fire only when the read is strong", or size up when it is.
We already measured the cheap version of this. In doc 19 we let five indicators (the 10-year, oil, ES, NQ, VIX, each on its last hour) vote on the side, then bought the at-the-money option at a pivot touch:
| Test | Sessions | Result |
|---|---|---|
| Five-indicator vote picks the side, bought at the touch, +25% / −40% | six months of real prints | −$823 a day, below its own shuffled labels |
| The plain hold, the read's side, 14 DTE, out at 15:30 (Monday's trade) | 7 real sessions, Oct 21–30 | +$107,917 at 10 contracts, 5 of 7 up |
| The same hold written (sold) instead of bought | same 7 sessions | +$12,186 at 10 SPY, 7 of 7 up |
No single indicator called the session better than a coin on our data. That does not say Mario's model is wrong; his 62% is a different model on different inputs. It says the bar is: does his strong-conviction subset beat the plain hold on our real prints? That is a backtest we can run the way we ran the others, before any setting changes.
What we need from Mario to test it:
A day-by-day file of his model's call and its confidence (not just up or down) for the sessions we have real option prints for, Oct 21 to Oct 30 at minimum, the six months of 30-minute prints ideally. With that we can rank by his gradient, trade only the top third, and put the result next to the hold on the same days.
Nothing here changes what the desk does: the standing arm (the read's side, held to 15:30, 14 days out, plus one SPX written) stands as set.
Pivots: P = (high + low + close) / 3 of the previous session; R1 to R3 above it, S1 to S3 below; R4 / R5 / S4 / S5 extend the ladder. The read: the index against P, calls above, puts below. A touch: SPX trading through a level or within half a point of it. 0DTE / 14 DTE: the option expiring today / about 14 calendar days out. At the money (ATM): the strike nearest the index; out of the money (OTM): a strike the index has not reached. The hold: buy once, hold to the time stop. The take: the resting sell above the fill. The stop: the sell when the option falls to a fraction of the fill. The time stop: 15:30 ET, everything out. Written: sold to open, bought back to close. The cascade: sell a small gain, re-buy at once, repeat. PF: profit factor, gross wins over gross losses; under 1 loses. The shuffled null: the same rule with its labels shuffled; a rule that cannot beat its own shuffle has no signal. Real prints: option prices the market actually quoted, as opposed to a model.
