Three stops, three R given back, a peak-to-trough drawdown of 5.77 percent, and a longest losing streak of three across two trading sessions. What each stop tau
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.
The week of May 25 closed three losing trades and produced no winners inside the loss-attribution sample. Total damage: three R, or roughly six thousand dollars on a one-hundred-thousand-dollar simulated account sized at two percent risk per trade. The peak-to-trough drawdown on the system equity curve bottomed at 5.77 percent on Thursday afternoon. The longest losing streak inside the window ran to three. Compared against the year-to-date record, the give-back is modest in dollars and shallow on the curve. Through May 31, 2026, the system has banked +20.00R YTD across 121 trades since the Jan 12 inception. A $100,000 simulated account at 2% risk per trade sits at $139,996 (static) or $145,328 (compounded). This week's give-back is approximately $5,776 of that figure on a static-balance basis — a normal drawdown event against a year still net positive in the high teens of R-multiples. We publish red weeks for the same reason we publish green ones: because a journal that only shows winners is marketing, and a system whose edge plays out over hundreds of trades is going to have weeks the edge does not show up. Each of the three entries this week was a setup our rules cleared. None was forced. None was a skipped veto. Every stop was a position the system sized in good faith and the market declined to extend. The week also delivered something specific worth slowing down on: a B-grade EURUSD long that stopped at the same trigger window as the Wednesday US30 long, and a US30 short on Thursday that took the second hit on the same instrument family inside twenty-four hours. They are the two trades we teardown below — not because they are the worst losses on the board (all three printed identical -1R), but because they are the most instructive.
The week's losses cannot be read without the Tuesday print that preceded them. May 26 carried three trades, all between 14:05 and 14:32 UTC, all cleared TP1 or better: a Claude NAS100 long for +0.78R, a GPT GBPUSD short for +0.63R, and a GPT USDJPY long for +0.57R. Cumulative equity climbed from $100,000 at the open to $103,970.45 by 14:32 UTC. The peak the drawdown curve is measured against was set on Tuesday, before any of the loss-side trades had printed. By the end of Tuesday the architecture had built a +1.99R cushion above its starting equity.
Wed May 27 produced two trades at the same 14:42 UTC trigger window, both losses. A Claude US30 long on a pullback continuation stopped for -1R at C+ grade. A GPT EURUSD long on a pullback-to-VWAP setup stopped for -1R at B grade — the cleanest read of any trade on the week. Two simultaneous full stops took cumulative equity from $103,970.45 to $99,970.45, erasing Tuesday's print and crossing flat by 14:42 UTC. The drawdown curve walked from 0 percent at peak to -3.85 percent at the second 14:42 UTC print. Two of the three losses on the week were on the board inside the same minute.
Thu May 28 carried a single trade. At 14:36 UTC a GPT-5.5 US30 short on a pullback-failure pattern stopped for -1R at C+ grade. Cumulative equity moved from $99,970.45 to $97,970.45, and the drawdown curve hit its trough at -5.77 percent — the deepest point of the window. The US30 column closed the week at -2.0R across two losses on opposite directions across consecutive days. Friday May 29 produced no trades; the line held at the trough into the weekly close. The week's full loss arc ran inside two trading sessions on consecutive days, with the architecture's no-trade Friday holding the line at -5.77 percent rather than reaching for a recovery print.
| Date | Time | Instrument | Dir | Model | Setup | Grade | R | $ Sim | Result | Details |
|---|---|---|---|---|---|---|---|---|---|---|
| May 27 | 14:42 UTC | US30 | Long | Claude Opus 4.7 | US30 Long Pullback to Broken Resistance | C+ | -1.0R(SL) | -$2,000(SL) | Stop hit | - |
| May 27 | 14:42 UTC | EURUSD | Long | GPT-5.5 | EURUSD Long Pullback Continuation | B | -1.0R(SL) | -$2,000(SL) | Stop hit | - |
| May 28 | 14:36 UTC | US30 | Short | GPT-5.5 | Short Failed Reclaim of OR High / VWAP | C+ | -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.
The pattern of the week was the runner refusing to extend on confluence cards that scored above the published floor at entry.
Three setups, on two instruments, across two sessions, all cleared the published threshold and none produced the second leg the runner window required. Across most weeks, a B-grade entry like Wednesday's EURUSD long gets the confirmation print and adds to the curve; a C+ pullback continuation like Wednesday's US30 long gets the same shape in roughly half the cases. This week neither did. The confluence math at entry was correct on each. The tape extension after entry was absent on each. The drawdown curve is the visible footprint of that absence.
The lesson is not that the reads were wrong. The B-grade EURUSD entry below was the cleanest read of the week and it stopped at a clean -1R, which is the position behaving exactly as designed under an invalidation. The lesson is that a thin window can sit inside a regime where the structural confluence at entry and the runner extension after entry disagree, and the rules will keep sizing the structural confluence because the structural confluence is what the rules can see at trigger. That is a tape problem, not a rules problem, and we do not adjust the engine to chase it on a three-trade sample.
The Risk Agent allowed the Claude US30 long and the GPT EURUSD long to size at full risk simultaneously at 14:42 UTC Wednesday, on two different model families and two different instruments. Each evaluation cleared its own confluence card honestly and the architecture's per-trade isolation logic correctly treated them as independent draws rather than a correlated position. They stopped together at the same trigger window, and the joint loss accounted for -2R of the week's -3R loss column. That is the design — independence at entry — not a discipline failure.
The Macro Agent did not veto any of the three entries, and we are not flagging that as a Macro error. The regime read at the time of each entry did not warrant a veto; the confluence card was there and the rules cleared the threshold. A non-veto that precedes a loss is not a Macro mistake when the information available at decision time supported the entry. We grade decisions on the read, not on the outcome.
The system declined nothing it should have taken and forced nothing it should have skipped. The honest decision review for this week is that the process ran exactly as designed and the runner declined to extend on three consecutive entries inside two trading sessions. Friday's no-trade outcome is the cleanest single discipline read of the window: the architecture did not fire on any instrument despite the three-trade loss streak sitting at -3R, holding the setup-grade floor at threshold rather than reaching for a recovery print.
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.
Same signals, same risk framework, different foundation model.
EURUSD: one loss, -1.0R. The Wednesday GPT long on a pullback-to-VWAP entry at B grade — the highest grade on any trade of the week — stopped at a clean -1R at the same 14:42 UTC trigger as the Wednesday Claude US30 long.
All EURUSD this week →GBPUSD: no losses this window. The pair produced one win (Tuesday GPT short, +0.63R) and no triggered losses across either model family.
All GBPUSD this week →US30: two losses, -2.0R combined. The week's instrument concentration: a Claude pullback-continuation long Wednesday and a GPT pullback-failure short Thursday, both at C+ grade, both cleared the confluence floor at trigger, neither produced the runner inside the entry window. The 24-hour interval between the two stops on opposite directions is the cleanest single-instrument signal of the week.
All US30 this week →NAS100: no losses this window. The index banked one winner (Tuesday Claude long, +0.78R) and did not produce a loss-side entry.
All NAS100 this week →USDJPY: no losses this window. The pair banked one winner (Tuesday GPT long, +0.57R) and did not produce a loss-side entry.
All USDJPY this week →US500: no losses this window. The index did not trigger on either model family across the available windows; the architecture sat the index out rather than reaching for a setup that did not clear threshold.
All US500 this week →Loss of the week: US30 Short · -1R
What was right: The cleanest read of the week and the highest grade on any of the six entries. A NY-afternoon GPT-5.5 long on a pullback-to-VWAP setup, graded B by the Trend Agent on a genuine confluence card rather than a reach. Macro did not veto regime. Sizing was the standard 2 percent risk on equity at trigger. The structural premise was honest: the session pullback into a VWAP and Fibonacci confluence had retraced cleanly and the entry shape carried the markers of an absorption-style continuation.
What was wrong: The runner the trade needed did not extend inside the entry window. Confluence said the level was good; the tape's response after entry said the second leg was not there. The stop printed at a clean -1R at the documented invalidation, which is the position behaving exactly as designed under a confluence card that fails to extend.
What we'd do same: Take the setup again. A B-grade pullback-to-VWAP entry that invalidates at -1R for one tape session is the system working, not the system failing. We do not adjust an honest B-grade read because a single session refused to extend.
What was right: A NY-afternoon GPT-5.5 short on a pullback-failure pattern, graded C+ on a genuine continuation read on the index. The Trend Agent saw the prior session's pullback hold and the failure shape clear the published floor; the Macro Agent did not veto regime. Sizing was the standard 2 percent risk. The pairing with the prior day's opposite-direction US30 long was not by itself a problem; each entry stands on its own confluence card and the architecture's per-trade isolation logic correctly treated them as independent draws.
What was wrong: The same runner-refusal that killed Wednesday's US30 long killed Thursday's US30 short on the opposite direction. The continuation refused to extend inside the entry window. Stop was a clean -1R at the documented invalidation.
What we'd do same: Take the read. The instructive lesson is in the pairing, not in either entry alone: two C+ pullback patterns on the same instrument in twenty-four hours on opposite directions, both stopped by the same runner refusal. That belongs in the tuning paragraph below, not in a verdict on the reads themselves.
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 | -3R | −$6,000 |
We did not enjoy writing this one and we are publishing it anyway, because the weeks we most want to skip are the weeks the gate exists to catch. Three losses, three R given back, a 5.77 percent peak-to-trough drawdown on the equity curve, no winners in the loss-attribution sample. The publish gate fired on its own arithmetic — loss count above the threshold, with no override from us. That is the design working. We did not force a drawdown story onto a quiet week, and we did not bury a real one when it arrived.
Put the dollars on the table. A $100,000 simulated account at 2% risk per trade gave back approximately $5,776 across the window's three stops on a static-balance basis. The peak-to-trough on the equity walk sat at -5.77 percent at Thursday's low and the line held at the trough into the Friday close. Through May 31, 2026, the system has banked +20.00R YTD across 121 trades since the Jan 12 inception, with the simulated account sitting at $139,996 static or $145,328 compounded. The $5,800-or-so compounded figure between the two — driven by 121 trades reinvesting risk against a slowly growing base — is the visible footprint of disciplined fixed-fractional sizing through both green and red weeks. The discipline is what holds the give-back bounded specifically in weeks like this one.
The honest close is the unromantic one. The system did this week what it does every week: it read the tape, it sized to its rules, and three times the runner declined to extend. We found one real thing to scope — the same-instrument repeat-exposure gap the two US30 entries expose — and we are scoping it narrowly rather than rewriting an engine that ran exactly as designed across a three-trade window the tape would not extend. The next report will look different. They always do.
From the desk, the SkyAnalyst Team.
The one concrete tuning candidate this week is the same-instrument repetition the two US30 entries expose. The Risk Agent evaluates each entry on its own confluence card and does not currently apply a discount when a second C+ read appears on the same instrument inside a twenty-four-hour window, even on opposite directions. Wednesday's US30 long and Thursday's US30 short were not independent draws on the underlying tape; they were two confluence cards on the same index across two consecutive sessions, both stopped at fixed-R. The change we are scoping is a narrow same-instrument repeat-exposure check that flags this specific pattern before the second entry sizes at full risk, rather than after both have stopped.
We are explicitly not tuning the reads themselves and we are not tuning the three entries collectively. Three stops in a window is not a signal that the entry rules are broken; it is a signal that the runner did not extend on confluence cards the rules read correctly at trigger. Overfitting the entry logic to one bad regime on a three-trade sample is exactly the mistake the statistics section warns against. The fix belongs narrowly inside the same-instrument repeat-exposure layer, scoped tightly, validated against the longer 91-trade record before it ships, and shipped only if the longer-record validation says it would not have cost more winners than it would have saved losers.
Here is the framework for reading this week's three stops without flinching or spinning them. The loss-attribution window carries a 34.1 percent win rate over a 91-trade trailing sample, an average winning trade targeting roughly 0.66R, a current peak-to-trough drawdown of 5.77 percent inside this window, and a longest losing streak of three. A three-trade window that goes zero-for-three on the loss-attribution sample looks heavy in isolation and is statistically routine in context. The math is the same math every legitimate fund publishes when it reports a drawdown.
Start with streaks. Van Tharp's work on R-multiples in <em>Trade Your Way to Financial Freedom</em> makes the point plainly: a system with a sub-fifty-percent loss-attribution win rate will, over a long enough record, produce losing streaks that feel uncomfortable in the moment and are statistically expected. A strategy with a 34.1 percent loss-attribution hit rate expects clusters of consecutive stops as a structural feature, not a malfunction. A three-trade streak inside a single window is not the deep tail of that distribution; it is the kind of streak the distribution requires the system to produce regularly. Jack Schwager's <em>Market Wizards</em> interviews return to the same theme from the practitioner side: the traders who survive size so an ordinary streak cannot end them, and they do not change the system because a normal short run feels abnormal on a small sample.
Now the arithmetic of this week. Three stops for a combined -3R is, by design, survivable: the risk layer holds each trade near a fixed fractional risk, so even a full window of losers caps the give-back at a known, bounded figure rather than an open-ended one. The drawdown curve confirms it: the trough sat at -5.77 percent, well inside the historical drawdown band for this system. That is uncomfortable. It is also inside the operating envelope a 34-percent loss-attribution win rate and a fixed-fractional sizing rule predict. We do not publish a computed Kelly fraction the engine does not use, because a number the system never trades on is theater. The honest statement is bounded: at this hit rate, three-trade losing streaks happen, sizing keeps them bounded, and three trades is far too small to update a 91-trade base rate.
No. The loss-attribution sample for this report is the three stops; the broader trailing window carries a 34.1 percent loss-attribution win rate over 91 trades. A zero-for-three stretch on a system whose loss-attribution hit rate sits near a third is uncomfortable and well inside the normal distribution of streaks. Year-to-date the system sits at +20.00R across 121 trades, win rate 57.9 percent. The arithmetic does not change because one window ran three for three against.
Because the publish gate was lowered to lossCount ≥ 1 in May 2026. Transparency beats thinness: a journal that hides drawdowns until they cross some arbitrary depth threshold is marketing, and a system whose edge plays out over hundreds of trades should publish its red weeks at the same cadence it publishes its green ones. This window cleared the gate cleanly, and we report what happened rather than what we would prefer to report.
One thing: a same-instrument repeat-exposure check inside the Risk Agent, so a second pullback-pattern entry on the same instrument inside a twenty-four-hour window does not size at full risk without a clustered-exposure discount, even on opposite directions. We are not changing the entry reads or the threshold logic; three trades is far too small a sample to retune the engine.
At a $100,000 simulated account with 2% risk per trade, the three -1R stops gave back approximately $6,000 in nominal terms, and the equity curve trough sat at -5.77 percent below the Tuesday peak before the Friday no-trade outcome held the line. That figure sits well inside the system's historical drawdown band; the YTD account balance still sits at +$39,996 static or +$45,328 compounded for the year.
Because three losses constitute a documentable streak under our gate, and the educational reader value of statistical context — Tharp on R-multiples, Schwager on streak distributions, fixed-fractional sizing inside expected ranges — does not require a deeper drawdown to land. The honest portfolio practice is to report what the gate fires on, not to wait for a deeper print to write about.
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: +20.00R YTD across 121 trades, see stats strip.
121 trades since Jan 12 inception, 70 winners, 51 losers, +20.00R net at TP1 baseline. A $100,000 simulated account at 2% risk sits at $139,996 static, $145,328 compounded. A system that launched slow, ramped staged, consolidated by May.
Forty-seven trades, twenty-seven winners, twenty losers, +8.34R net at the TP1 baseline. Claude ran +6.89R on 26 trades at 61.5 percent; GPT printed +1.45R on 21 trades at 52.4 percent. The two-model lineup stabilized this month.
Six trades, three winners, three losers, -1.01R net at the TP1 baseline. Tuesday opened with a 3-for-3 morning printing +1.99R inside thirty minutes; Wednesday and Thursday gave it back through two US30 stops and an EURUSD stop.