GPT Stops Out on NAS100 — Three Evaluations, One Quick Loss

Strategy Analysis

The pullback long that GPT rated tradeable

The setup template is the standard intraday continuation. After a directional opening move, price retraces toward a confluence zone with at least three independent levels. The entry trigger is a completed 5-minute bullish candle inside the zone. The structural stop sits below the zone's lower edge. Targets ladder up to extension levels.

GPT's read on the NAS100 chart was direct. The trend authority signal was unambiguous at 73% bullish in a TRENDING regime. The pullback geometry was clean. The four-way confluence at 28740–28770 was the kind of zone the framework is calibrated to take. The entry trigger candle at 14:58 UTC had pushed back above 28780, closing inside what GPT considered the zone's upper resolution. The 15-minute RSI was rolling from overbought toward neutral — which the framework reads as healthy reset, not bearish reversal, when trend authority is strong.

Three evaluations between 14:42 and 15:05 walked through the read. The first two flagged the trigger as marginal. The third — at 82% confidence, GPT's highest read of the day — committed. From the model's own reasoning: "Pullback has completed into 28740–28770 with the entry candle closing back above 28780. Trend authority at 73% BULLISH TRENDING confirms direction. Macro stable. RSI reset on the 15m is supportive, not bearish. Position sized at 13.10 lots with structural stop at 28705 immediately below overnight liquidity."

Three evaluations is not many. GPT's typical evaluation count on a marginal-trigger pullback is five to eight. Today's three reflects high conviction on the trend authority signal and a willingness to act fast on what looked like a clean confluence stack. The framework rewarded that profile twice in the prior week — the May 5 US30 long was a one-evaluation entry at 74% confidence and went to TP3. Today the framework took the same shape and the market did not cooperate.

What happened after the entry

The trade was over in forty-six minutes.

Entry was 15:11:58 UTC at 28781.5 — the exact top of the entry zone on a 13.10-lot fill, $1,067.54 of risk-side dollars. The peak price after entry was 28781.5, which means the position never went into profit at all. Price drifted lower from the entry candle. By 15:25 the tape was at 28760, working under the EMA cluster. By 15:40 NAS100 was inside the 28720s. The structural stop at 28705 — placed just below the overnight low sweep at 28701 — was tagged at 15:58:00 UTC. Exit price was 28704, a single tick below the stop. The broker recorded the position closing at -$1,032.28 (SL) on the 13.10-lot fill, -1.0R on the framework side.

No oscillation. No second-touch test. No bid materializing at the confluence zone. The pullback geometry that drew the entry simply did not produce buyers. The Fib level at 28755 broke. The EMA cluster broke. The overnight liquidity sweep at 28701 was tested, the stop was triggered, and the trade was done.

The number that matters is the peak price after entry: 28781.5. That is the entry price itself. The trade never produced a single tick of favorable excursion. In benchmark trading, this is a specific and recognizable failure mode — the entry was either too early, too late, or against the prevailing micro-structure. Forty-six minutes from entry to stop with zero positive excursion is a clean signal that the trigger was the wrong candle.

Why the trigger fired anyway

This is the diagnostic part of the post-mortem.

GPT's evaluation framework weights trend authority and macro confirmation heavily — by design, the framework is calibrated to act on strong directional signals even when the trigger candle is imperfect. The trade-off is explicit: a higher base-rate of entries in trending regimes, at the cost of occasional losses on entries a stricter trigger-rule would have filtered. Across enough trades, the calibration produces a season-long return curve that depends on how often the trending regime tag accurately predicts continuation.

In the May 7 reading, the regime tag was TRENDING at 73% bullish. The framework treated that as load-bearing. The trigger candle at 14:58 — which closed back above 28780 inside the zone — was rated good enough given the strength of the trend authority signal. A strict-trigger version of the rule would have required a textbook bullish reversal candle with a clean wick into the zone's lower edge. That candle did not print. Price hovered at the zone's upper edge, never tested the zone's lower edge cleanly, and the entry triggered on a tepid mid-zone candle.

The quiet tell was the MACD histogram on the 5-minute. It had flipped negative ten minutes before the entry candle closed. A strict-momentum filter would have held the position out of the trade on that basis. GPT's framework does not currently weight 5-minute MACD as a primary input — the model treats it as a secondary confirmation that can be overridden by strong trend authority. Today the trend authority signal was strong and the momentum signal was weak, and the framework's calibration weighted the trend authority. The market resolved in favor of the momentum read.

This is not a framework failure. It is a calibration trade-off resolving against the framework on a single trade. Across the season, the same calibration produced GPT's +8.90% finish. The closing-week loss on this entry is the cost of that calibration on Day 18.

The parallel chart — Claude on the same read

The coordinated trap is the part of today worth dwelling on.

Claude took NAS100 long the same morning, nine minutes after GPT, on essentially the same confluence stack. Different evaluation framework, different trigger rule, six evaluations versus GPT's three, and the same conclusion: the pullback is tradeable, the structural read is correct, the entry should fire. Claude entered at 28770 at 15:21 UTC. GPT had entered at 28781.5 at 15:11 UTC. Both stops were placed in the same structural zone — Claude's at 28690, GPT's at 28705. Both stops triggered within minutes of each other. Claude's loss was -$1,014.00 (SL) and -1.0R. GPT's loss was -$1,032.28 (SL) and -1.0R. The aggregate loss across both models was -$2,046.28.

When two independent models with different evaluation frameworks reach the same conclusion on the same chart and the market rejects both, the read on the underlying instrument deserves the credit, not either model's discipline. The NAS100 chart on May 7 was a trap setup — strong structural confluence, weak follow-through, no buyers when buyers were required. Both models read the structure correctly and both models lost the same trade for the same reason.

This is the part of the AI Trading Benchmark that is hardest to control for. Two models can both be right about what the chart shows and both be wrong about what the chart does next. The benchmark measures cumulative outcomes across many trades. Single coordinated losses like today's are not evidence of framework failure — they are evidence that two models can reach the same incorrect conclusion when the structural read invites it.

How this fits GPT's Season 1 profile

GPT closes Season 1 at +8.90%. The May 7 loss is part of that closing balance, not an outlier from it.

GPT's Season 1 profile is a model that takes more trades than Claude, acts faster on high-confluence setups, and pays for that aggression with occasional fast losses on marginal triggers. The cumulative ledger going into Day 18 sits at +$3,377.24 net across ten recorded broker trades — a +6.75% return before today's stop is booked. The wins have come on setups where the trend authority signal aligned with momentum and the trigger candle was unambiguous. The losses have come on setups where the trend authority signal was strong but the trigger was imperfect — exactly the pattern today's NAS100 produced.

Looking at the prior month, the same pattern is visible. The April 22 US30 long was a strict-trigger entry that went to TP3 cleanly. The April 27 US30 long was another textbook entry that produced. The April 24 NAS100 long was a marginal-trigger entry that stopped fast — same failure mode as today. The framework's win/loss distribution is consistent with its calibration. The seasonality of marginal-trigger losses is also consistent: they cluster around regime transitions where trend authority remains strong but momentum starts to break.

Today's stop puts GPT at $52,344.96 — a step back from the May 6 close at $53,377.24. The season balance is intact, but the lead over Claude has compressed to $78 after both models stopped on the same NAS100 setup. The closing week will be decided by the next two trading days, May 8 and May 11. A small swing in either direction reshuffles the season standings entirely.

What the closing-week setup looks like

Friday is May 8. The macro calendar carries the BoC speech in the afternoon and a light US data window. The dollar will be the deciding input — DXY has been ranging between 98.20 and 98.50 for three sessions, and the May 7 close at 98.31 is mid-range with no clear directional bias.

For GPT, the strategic question is whether to take the May 8 setup if it presents. The model's bias has been to act on high-confluence reads even after a fresh loss. The framework treats prior-trade outcomes as independent of the current setup's edge, which is the correct statistical posture but occasionally produces back-to-back losses when the regime continues to disappoint. The Season 1 closing-week record will reflect whether May 8 and May 11 produce setups GPT's calibration is built for, or setups it is calibrated against.

Today's data point is clean. The trade was marginal. The execution was disciplined. The loss was fast and contained. The framework took a high-confluence read, the market rejected it, the position resolved at the structural stop. That is the framework working exactly as designed — every part of it except the part that matters today, which is the market's vote on whether the structural read was right.