Weekly Losses: the GBPUSD short we kept paying for
Three losses, two of them on the same instrument, same direction, forty-six hours apart. The desk gave back 2.5R against a +21.60R YTD line, then closed the wee
Before the trade, meet the system.
SkyAnalyst is not one AI trader. It is four specialist agents — each with its own data pipeline, each maintaining state between evaluations, and each required to agree before a position is sized. They don’t chat in prose. They write structured messages to a shared state object that each reads on every evaluation cycle.
Through Jun 8, 2026, the desk has banked +21.60R YTD across 107 trades, YTD through Jun 8, 2026. A $100,000 simulated account at 2 percent risk per trade sits at $143,204.12 on the static path, or $150,174.39 on the compounded path, from Jan 12 inception. This week we gave 2.5R of that figure back. Three losses, two of them the same instrument in the same direction, forty-six hours apart. The simulated curve walked to a 5 percent trough on Wednesday afternoon and recovered to a fresh peak by Friday close. The Friday recovery is in the weekly recap. The two GBPUSD shorts are the only thing worth writing about here, so that is what this article does.
Act 1: The Monday cluster
The first loss landed at 15:06 UTC on Monday, a GBPUSD short stopping for a clean 1R on the post-ISM pullback read. Nineteen minutes later, a USDJPY long stopped at the broker 1R, published in this article at minus 0.5R because the Claude USDJPY trader currently carries a 0.5x marketing risk scalar (see the disclaimer at the foot of the page). Two entries inside twenty minutes, both off the same dollar-tape inflection: the post-ISM read had been a fade of dollar strength, and the dollar caught a bid before either trade could work. The simulated curve walked from $100,000 to $97,000 inside that twenty-minute window.
Act 2: The Wednesday second chance
Tuesday was a stand-down. The desk did not place a trade. Wednesday afternoon at 14:33 UTC, the Setup Agent flagged a second-chance GBPUSD short, a grade B read on the post-data lower high. It stopped at flat 1R. That is the second-worst kind of loss: same instrument, same direction, same outcome as Monday's first trade, forty-six hours apart. Drawdown deepened to 5 percent on the simulated equity walk. That was the trough.
Act 3: The walk-back
The desk closed the GBPUSD book after the Wednesday stop and stood down on the pair for the rest of the week. Wednesday's late afternoon and Thursday produced a string of winners that we cover in this week's recap, and by Friday at 14:36 UTC the simulated equity printed a fresh peak. The week closed green on the recap side of the books. It still cost us 2.5R, a 5 percent drawdown, and the GBPUSD lesson written into the Pattern of the Week section below.
Every trade the system took.
| Date | Time | Instrument | Dir | Model | Setup | Grade | R | $ Sim | Result | Details |
|---|---|---|---|---|---|---|---|---|---|---|
| Jun 1 | 15:06 UTC | GBPUSD | Short | Claude Opus 4.7 | GBPUSD Post-ISM Pullback Short into VWAP/Fibonacci Resistance | C+ | -1.0R(SL) | -$2,000(SL) | Stop hit | - |
| Jun 1 | 15:25 UTC | USDJPY | Long | Claude Opus 4.7 | USDJPY Pullback Long to VWAP/Support Continuation | C+ | -0.50R(SL) | -$1,000(SL) | Stop hit | - |
| Jun 3 | 14:33 UTC | GBPUSD | Short | Claude Opus 4.7 | Post-Data Second-Chance Short | B | -1.0R(SL) | -$2,000(SL) | Stop hit | - |
Dollar figures are simulated on a $100,000 account at 2% risk per trade. Actual subscriber P&L varies with account size. Past performance is not a guarantee of future results.
Pattern of the week
The pattern this week is a recurring read on a single pair, and the system did not weight it.
Two of the three losses were GBPUSD shorts. Read independently, both setups graded well: the Monday entry at 15:06 UTC was a clean post-ISM pullback into VWAP and Fibonacci resistance, a C+ read on confluence. The Wednesday entry at 14:33 UTC was a grade B second-chance short against the post-data lower high. Each setup, on its own evaluation cycle, looked right. The shared structural premise was that the pound had run too far against the dollar and would mean-revert into the macro tape that had set up earlier in the week.
What the system did not weigh is that the Monday read had already paid its rent. When the same instrument, in the same direction, fires a second time within forty-six hours of a stop, the prior result is information. The Setup Agent does not currently look back at recent stops on the same pair before scoring a new confluence read. The Wednesday short was sized as if Monday had not happened, because in the agent's working state, Monday had not happened. The teardowns below break down both entries; the fix lives in the What we tune next section.
Decision highlights
The Setup Agent sized the Wednesday GBPUSD short at 14:33 UTC without weighting that the same instrument, in the same direction, had stopped on Monday. The setup was a clean read on its own merits, grade B on confluence. The portfolio memory of two days earlier was not part of the gate. This is the single most expensive judgment call of the week, and it lands at the top of the tuning list.
The Risk Agent held flat 1R sizing through every entry of the GBPUSD sequence, refusing to escalate after the first stop and refusing to reduce after the second. That discipline is what kept the give-back at 2.5R instead of the 4R or 5R that revenge sizing would have produced. The decision to do nothing is the decision the most experienced traders make and the one untested systems usually fail.
The desk stood down on Tuesday and stood down on GBPUSD entirely after the Wednesday stop. There is no rule that says we have to trade every session, and the absence of a Tuesday entry is itself a decision: the Trend Agent did not see a setup that cleared the threshold, and the desk did not force one. Quiet days are not failed days; they are the cost of a confluence-gated system.
Claude vs GPT: who led the week.
SkyAnalyst runs multiple foundation models in parallel across its four-agent system. When two models trade the same instrument in the same week, the results are directly comparable. This is that comparison.
- GBPUSD-2.0R · 2 trades
- USDJPY-0.5R · 1 trade
Same signals, same risk framework, different foundation model.
Six instruments, six stories.
EURUSD: no losses this window. The EURUSD trader flagged one setup mid-week that scored below threshold and did not size.
All EURUSD this week →GBPUSD: two losses, both shorts. The Monday entry at 15:06 UTC stopped for 1R on the post-ISM pullback read. The Wednesday entry at 14:33 UTC stopped again for 1R on the post-data second-chance read, taking the simulated curve to its 5 percent trough before the recovery began. Both grades held on their own merits; the pattern is in the repeat.
All GBPUSD this week →US30: no losses this window. The US30 trader sat through the Monday risk-on rotation without an in-confluence short, which is what we would expect on a session with that macro tape.
All US30 this week →NAS100: no losses this window. The NAS100 trader flagged two long candidates Thursday morning, both filled on the winning side and reported in the weekly recap.
All NAS100 this week →USDJPY: one loss. The Monday pullback long at 15:25 UTC took the full broker 1R stop as the post-ISM dollar pop reversed the continuation read. The trade publishes at minus 0.5R here because the active marketing risk policy scales Claude USDJPY R-multiples by 0.5x; the broker ledger shows the full minus 1R.
All USDJPY this week →US500: no losses this window. The US500 trader graded one short below threshold mid-week; it later filled on the recap side of the curve as a different read.
All US500 this week →Loss of the week: GBPUSD Short · -1R
Losses worth learning from
what was right
The grade B read on the Wednesday GBPUSD short at 14:33 UTC. Post-data lower high against the prior session's resistance shelf, VWAP rejected on the underside, volume on the rejection candle above the 60-period average. The Setup Agent had every on-its-own-merits confluence factor it needed to size in.
what was wrong
The portfolio memory was not in the score. Monday's GBPUSD short at 15:06 UTC had stopped on the same direction, forty-six hours earlier. The probability of a second-chance read paying off does drop measurably when the first read stopped two days earlier on the same instrument; the Setup Agent did not weight that signal because the gate does not currently check for it.
what we'd do same
The structural read. The grade. The flat 1R sizing.
what we'd do different
A look-back window in the Setup Agent that downgrades confluence on a same-instrument, same-direction re-entry inside 72 hours. Probably by half a grade. The fix is in the next section.
what was right
The grade C+ read on the Monday GBPUSD short at 15:06 UTC. The setup was a textbook post-ISM pullback into a VWAP and Fibonacci confluence zone, with a 5-minute rejection candle on volume above the 60-period average. The Trend Agent had the confluence math; the entry was the right read on the right tape.
what was wrong
The stop sat where the rule says it sits, just above the prior swing high. The dollar caught a bid faster than the post-ISM read implied, and the level we were fading held for nineteen minutes longer than the trigger needed. That is a structural loss, not a mis-read. The setup was the C+ that the grader said it was. The market disagreed.
what we'd do same
Everything about the read, the grade, the entry, the stop placement, and the size. The Monday GBPUSD short is the median losing trade we ask the system to take.
what we'd do different
Nothing on this specific entry. The fix from the Wednesday teardown applies in retrospect: had we shipped the cooldown gate before this week, the Wednesday short would not have re-entered. The Monday loss would still have happened.
On a $100k account at 2.0% risk per trade.
Each trade risks +$2,000 (1R). The system's actual scale-out behavior may differ, see disclaimer.
| Scenario | R-multiple | Profit on $100k |
|---|---|---|
| Window drawdownActual | -2.5R | −$5,000 |
All six agents combined.
From the desk
A $100,000 simulated account at 2 percent risk per trade enters this week at $145,704.12 on the static path, or $152,632.86 on the compounded path, against an inception equity from Jan 12. It exits the week at $143,204.12 on the static path, or $150,174.39 on the compounded path. This week's give-back is approximately $2,500 on the static figure, and approximately $2,458 on the compounded figure. The static path and the compounded path tell a slightly different story week over week, and the gap between the two ($6,970.27 in the desk's favor on the compounded side, as of Jun 8, 2026) is the receipt for disciplined sizing across 107 trades since Jan 12.
We almost did not publish this article. The week closed green on the recap side of the books. By the time the Friday session ended, the curve was at a fresh peak. The natural editorial instinct is to write only that story and not this one. We are publishing this one because the GBPUSD pattern is research data we owe the people who read this site, and because a desk that only writes about its winning weeks is not a trading desk, it is a marketing channel.
The Claude instrument-traders had a structurally bad GBPUSD week. They did the parts of the job we ask of them, well: the Trend Agent flagged a clean confluence read on Monday and a clean second-chance read on Wednesday, the Risk Agent held the line on sizing across both, the Macro Agent had the dollar regime tagged correctly on both sides of the post-ISM pop. The part of the job we have not asked the GBPUSD trader to do, yet, is the look-back at its own recent record on the same pair. We are asking it to do that now.
What we're tuning
One specific change ships in this week's iteration cycle. The Setup Agent gets a per-instrument cooldown gate: any new entry on a pair that has already stopped in the same direction inside the prior 72 hours has its confluence score downgraded by half a grade. The Wednesday GBPUSD short would have re-evaluated as C+ on this gate instead of B and almost certainly declined entry, on a model that requires grade B or better on the second-chance read.
The change is mechanical, testable on the historical record, and small enough to ship without a model swap or a retrain. The back-test on the YTD trade ledger runs next week. If the gate produces fewer false positives across the full sample than the count of avoided same-instrument re-entry losses, it ships permanently. If the back-test surfaces meaningful winners the gate would have vetoed, we publish that analysis instead and the gate does not ship.
What the numbers actually mean
A YTD win rate of roughly 59 percent across 107 trades sounds high to anyone who has not read Van Tharp's `Trade Your Way to Financial Freedom`. Tharp's central point in that book is that a trading system is not graded on win rate; it is graded on expectancy, which is win rate times average win minus loss rate times average loss. A system that wins 59 percent at a positive average R and loses 41 percent at flat 1R has positive expectancy if the win-side R distribution is wide enough. The system's YTD net of +21.60R across 107 trades is the receipt for that math. Expectancy, not win rate, is the number the desk runs on.
This week we drew a 2.5R give-back with a longest losing streak of three. Jack Schwager catalogs in `Market Wizards` the streak distributions that come out of trend-following and pattern-based systems with win rates in the 35 to 60 percent range; a three-loss streak is well inside the expected range for a system in our window, and longer streaks show up on the historical math several times a year. The current 5 percent peak-to-trough drawdown is uncomfortable to look at because the curve is in the public, but it is not anomalous. We publish weekly losses because every legitimate trading fund publishes its drawdown reports; the alternative is showing readers only the wins and hoping they do not check the equity curve.
The Kelly criterion gives a conceptual answer to the question of how much to risk per trade. We do not publish a computed Kelly fraction because the inputs (true win rate, true average win, true average loss) drift on a system this young and we do not want a reader to take a number off this page and use it as a sizing rule. Fixed 2 percent risk per trade is well below any reasonable half-Kelly we could compute from current data, which means the system is structurally underbet, which means drawdowns of this size are absorbed and not amplified. The 107-trade sample is the YTD sample. The next quarter of execution data will tighten the confidence interval; the structural conclusion (positive expectancy under flat 2 percent risk) will probably not change.
At a Glance
Drawdown questions
Is a 5 percent drawdown normal for a system with this kind of win rate?
It is well inside the expected range. A system that wins around 59 percent of its trades and runs at flat 2 percent risk per trade should expect drawdowns of 5 to 10 percent at multiple points across any given quarter. The distribution of streak lengths in Schwager's research suggests three-loss streaks are normal and longer runs show up several times a year. The number to watch is not any single drawdown, it is the slope of the equity curve across hundreds of trades.
Why did the desk re-enter GBPUSD short on Wednesday after Monday's stop?
Because the Setup Agent grades each entry on its own confluence math, and the Wednesday read was a grade B independent setup against a fresh post-data lower high. The portfolio memory of Monday's stop on the same pair was not part of the scoring gate. After this week, that will change for same-instrument, same-direction re-entries inside a 72-hour window. The change is in the What we tune next section.
How does the compounded path differ from the static path week to week?
The static path assumes the account size never changes; every trade risks 2 percent of the original $100,000. The compounded path grows or shrinks the per-trade dollar risk with the account, so a winning streak compounds faster and a losing streak loses faster. Across this YTD, compounded sits $6,970.27 above static, which is the gap that disciplined sizing buys you on the upside of a winning trajectory.
Are losing weeks like this one a sign the model is degrading?
Not on this sample size. 107 trades is enough to compute the headline statistics but not enough to detect a slow drift in win rate or average R. The desk runs continuous monitoring on both; the alert thresholds are set at multiples of historical standard deviation, and nothing in this week's data tripped them. We will publish a degradation analysis the quarter we see one.
Why publish the losses at all?
Because every legitimate trading fund publishes its drawdown reports. The alternative is showing readers only the wins and treating the losses as marketing inconvenience. We are not running that playbook. The losses are research data. They are also a more accurate picture of what the desk feels like to trade with than the wins.
Trade with the system that publishes its drawdowns.
Subscribers receive every signal — winners and losers — three minutes before entry, with full reasoning.
Dollar figures are simulated on a $100,000 account at 2% risk per trade. Drawdown trajectories shown reflect a small window sample size and are not projections of forward performance. Past performance — including losses — is not a guarantee of future results. Actual subscriber P&L varies with account size and execution. YTD context: +21.60R YTD across 107 trades, see stats strip. Risk-policy note: Claude USDJPY R-multiples publish at 0.5x the broker figure. SkyAnalyst applies a per-Trader marketing risk policy to published R-multiples; the live broker accounts run uniform 1R risk per trade. The dashboard at skyanalyst.ai shows raw broker outcomes, which is why a per-Trader article net can differ from the dashboard net by exactly the de-risked amount. See src/config/instrument-risk-policy.ts for the dated rationale.
From the SkyAnalyst Journal
The second euro short of the day, on a five-pip stop
The system had already shorted EURUSD to TP3 that morning. Hours later it took a second short into the same weakness, and a compressed tape let it risk just five pips for a full-potential 2.4R.
The Dow short the market breadth called before price did
A Dow short where the tell was not the chart but the market internals. Negative breadth kept the bias short, price failed at the open, and the fade ran to TP2 for a full-potential 1.5R.
The week's first trade, selling the failed reclaim of a low
The week opened with a Cable short that did not sell a bounce. It sold the failed retest of a broken low, on a single evaluation, seven of seven confluences, TP1 for a full-potential 1.06R.