May 18-24, 2026: A Seven-Eleven Week That Closed -2.82R Across Claude and GPT

Eighteen trades, seven winners, eleven losers, -2.82R net on the TP1 baseline. That is the scorecard for May 18-24, 2026 across the canonical-instrument set, and it sits in the lower band of the published expectancy. Cumulative equity opened at $100,000, climbed to $102,910.66 on Tuesday morning after GPT's US30 short ran TP1, gave back through Wednesday and Thursday into a -4.11R intraweek trough on Friday's second print, then recovered most of the week's giveback through a midday GPT cluster to close at $94,364.13. The five published winners are linked in parallel: the May 18 USDJPY long ran TP1 for +0.47R on Claude, the May 18 NAS100 short ran TP1 for +0.84R on Claude, the May 19 US30 short ran TP1 for +1.65R on GPT as the win of the window, the May 20 US500 long ran TP1 for +0.94R on GPT, and the May 22 US500 long ran TP1 for +1.15R on GPT. The remaining two winners are aggregated in the trade index alongside the eleven losses below.

Act 1: Monday belongs to Claude, with a stop in the middle

Mon May 18 carried three trades, all Claude Opus 4.7. At 14:52 UTC a USDJPY long on a NY AM continuation triggered at C+ grade and ran TP1 for +0.47R. Nine minutes later, at 15:01 UTC, a US30 long on a pullback continuation stopped for -1R at the same grade. At 15:42 UTC a NAS100 short on a NY AM bearish pullback triggered at C+ and ran TP1 for +0.84R. Monday closed at $100,610.66 cumulative, +0.31R net on a two-winner, one-loser session, with both wins running through Claude.

Act 2: Tuesday and Wednesday hand the green side to GPT

Tue May 19 produced four trades, three GPT and one Claude. At 14:42 UTC a Claude USDJPY long stopped at -0.5R on a partial structural exit. At 14:45 UTC a GPT US30 short on a continuation pattern ran TP1 for +1.65R, the largest single contribution of the window. At 15:09 UTC a GPT GBPUSD short on a post-data second-chance entry stopped for -1R, and at 15:37 UTC a GPT NAS100 short on a pullback failure at the prior-day low stopped for -1R. Wednesday's three trades extended the chop. A GPT USDJPY long stopped partial at -0.5R, a Claude EURUSD short stopped for -1R, and a GPT US500 long on a pullback retest ran TP1 for +0.94R. Wednesday closed at $97,785.66 cumulative, -1.11R net at the end of Act 2, with the index side carrying both sessions' green print.

Act 3: Thursday-Friday dig the trough, then GPT pulls back toward flat

Thu May 21 produced two losses, a GPT USDJPY long for -0.5R partial and a Claude NAS100 short for -1R. Fri May 22 ran a six-trade cluster between 14:05 and 15:38 UTC that swung the week's full equity arc. A Claude US500 long stopped for -1R, a GPT USDJPY long stopped partial at -0.5R, then a GPT US30 long ran TP1 for +0.59R, a GPT US500 long ran TP1 for +1.15R, a GPT EURUSD short stopped for -1R, and a Claude GBPUSD long ran TP1 for +0.56R. From the intraweek trough at $91,785.66, the architecture clawed back $2,578.47 across the final four prints to close the window at $94,364.13, -2.82R net.

Pattern of the week

The week's recurring pattern was pullback continuation entries that triggered cleanly at C+ grade and then split on what the tape did inside the trade lifecycle, seven cleared TP1 or better, eleven did not. Twelve of the eighteen trades were longs and six were shorts; both directions printed the same shape, with index longs and pair longs leaning on retest continuations and the few shorts leaning on pullback failures into resistance. The C+ grade described the entry card on most trades, the outcome varied with whether the second leg formed inside the runner window.

Inside the eighteen canonical trades, US30 carried the cleanest green at +1.24R net across three trades for a 66.7 percent win rate, US500 banked +1.08R across three trades for a 66.7 percent win rate, GBPUSD split one-for-one for -0.44R, NAS100 went one-for-three for -1.16R, USDJPY went one-for-five for -1.53R, and EURUSD took two losses for -2.0R. The index side carried the week and the FX side gave it back. Seven winners contributed +6.10R combined at the TP1 baseline; eleven losers contributed -8.92R combined, where four of those losses printed at -0.5R partial exits rather than the full -1R stop.

How a seven-eleven week settles at -2.82R

Seven winners summed to +6.10R: +0.47, +0.84, +1.65, +0.94, +0.59, +1.15, and +0.56. Eleven losers summed to -8.92R: seven full stops at -1R apiece for -7.0R and four partial structural exits at -0.5R for -2.0R, with one additional -0.5R partial that paired against the USDJPY entries on Tuesday and Wednesday. The net at -2.82R is the arithmetic of a 38.9 percent win rate with an average winner of +0.87R and an average loser of -0.81R. The expectancy math runs slightly negative when the win rate sits in the high thirties and the average winner does not clear 1R, and that is exactly what landed. The four partial -0.5R exits matter as a line item, they trimmed -2.0R off what would have been a clean -10.92R loss column.

Decision highlights

The Monday decision to keep firing Claude entries after the 15:01 UTC US30 stop is the discipline read of the week's first session. Between 14:52 and 15:42 UTC the system sized into three back-to-back triggers on the same automation, and the Risk Agent did not pull exposure on the third entry after the second printed a full -1R. The NAS100 short that came nine minutes after the stop ran TP1 for +0.84R, the largest Claude contribution of the window. Each entry was sized at the standard 2 percent risk on equity at trigger.

The Tuesday decision to take the GPT US30 short at 14:45 UTC against a backdrop of two Tuesday losses already in the column is the cleanest standalone read of the week. The 14:45 UTC entry triggered three minutes after a Claude USDJPY partial stop on the same instrument family and ran TP1 for +1.65R, the window's deepest single contribution. The architecture's per-trade isolation logic did exactly what it is written to do. It read each entry on the confluence math at trigger and did not down-weight the GPT signal because of an adjacent loss on a different automation.

The Friday decision to keep cycling the GPT automation across four consecutive entries between 14:38 and 15:38 UTC is the recovery read of the window's bottom. From the trough at $91,785.66 after the second 14:05 UTC stop, the system fired US30 long, US500 long, EURUSD short, and GBPUSD long in sequence. Three of the four cleared TP1 for a combined +2.30R, the EURUSD short stopped for -1R between them, and the cluster brought the week's net from -4.11R back to -2.82R at close. The recovery was structural, not narrative, each entry triggered on its own confluence card and the Risk Agent ran the fixed-R sizing unchanged through the sequence.

Loss worth learning from

What the system saw that was right

Each of the eleven losers cleared the published confluence threshold at trigger. The May 20 Claude EURUSD short faded a session-high rally into a confluent resistance zone the architecture has banked positive R on in other windows; the May 22 GPT EURUSD short triggered on a pullback into VWAP and Fib resistance; the May 21 Claude NAS100 short ran a VWAP and EMA9 rejection; the May 22 Claude US500 long triggered on a VWAP and breakout retest. Every grade described an entry card that scored above threshold and the Macro Agent did not veto regime on any of the eleven. The four partial -0.5R USDJPY exits show the structural-exit logic reading invalidation correctly before the hard stop printed.

What the system got wrong

The USDJPY column carried the heaviest weight, five trades for one win and -1.53R net, and the pattern inside that column matters more than any single stop. The pair's longs all triggered on pullback continuations or breakout retests above session highs that did not hold inside the runner window. The architecture read the entry card cleanly on each, the tape simply did not produce the second leg. The four partial structural exits on USDJPY trimmed the loss column by roughly 2R against clean full stops, the structural-exit logic was doing its job through the entire FX side of the week. The two EURUSD full stops are the simpler losses, both sold rallies into resistance and the tape continued through the level on both. The recap surfaces the pattern, not the trade-by-trade autopsy.

From the desk

The honest reading is that this week did exactly what the published expectancy says a thirty-nine percent win rate week should do at our risk-per-trade. Eighteen trades, seven winners, eleven losers, -2.82R net. A 38.9 percent win rate with an average winner of +0.87R and an average loser of -0.81R produces a slightly negative R-multiple, and that is what landed. The week also carried a clean head-to-head between Claude Opus 4.7 and GPT-5.5 across both master automations, and the architecture ran the same confluence floor and the same fixed-R policy across both sides through the full eighteen-trade sequence.

The architecture point is that the sizing did not change to reflect the run on either side. Claude's three sequential Monday entries were sized at the same 2 percent risk on equity at trigger that GPT ran through Tuesday's index winners and Friday's four-trade recovery cluster. A discretionary trader closing the morning red on either side would have stepped aside after the third or fourth loss. The system did not. The fixed-R policy and the published confluence threshold ran unchanged through the eighteen-trade sequence across both model families and across the FX-to-index instrument boundary. The recovery from the -4.11R intraweek trough back to -2.82R at close was structural, not narrative, four GPT entries on confluence cards that scored above threshold and a Risk Agent that did not down-weight sizing after the EURUSD stop printed between the green entries.

The TP1-baseline reading of -2.82R is the simplest comparison across periods, broker fills on the seven winners that ran past TP1 will read different on a scale-out or trail strategy. The four partial -0.5R structural exits on USDJPY printed identically across baselines, those are exit-logic outcomes, not baseline differences. From the SkyAnalyst Team.

What we're tuning

The USDJPY column is the one signal worth flagging out of this window. One winner across five trades at 20 percent win rate is the kind of single-instrument print that warrants a closer review of the confluence floor on the pair, not a threshold change on the basis of one week. We are reviewing whether the architecture's NY AM continuation and pullback entries on USDJPY are over-weighting confluence factors that decayed inside the runner window across this specific week's tape. The fixed-R policy and the 2 percent sizing rule are unchanged across both Claude and GPT through the review.

The four partial -0.5R structural exits on USDJPY are the second item on the bench, in a positive sense. The structural-exit logic did exactly what it is written to do across each of the four, reading invalidation before the hard stop printed and trimming the loss column by roughly 2R against clean full stops. We are not opening a new instrumentation track from this week, the eleven-loss column is the published expectancy running on the left side of the distribution rather than a regime drift signal.