Season Scorecard

The Tape Both Models Traded Into

Week 1 of Season 2 opened with a market that punished conviction and rewarded almost nothing. The macro backdrop wasn't a single thematic storm — there was no NFP, no FOMC, no surprise CPI. It was harder than that. It was a chop week with a Friday rug-pull, and the rug came out of nowhere.

DXY started the week at 99.4 and finished at 100.1, drifting higher in five straight sessions. The dollar strength was incremental rather than impulsive — none of the moves hit the kind of one-day extension that gives a trend trader a clean push. US10Y yields opened at 4.68%, briefly probed 4.72% on Wednesday, and closed the week near 4.66%. Yields without direction. Equities were similarly indecisive: the SPX cash index held inside a 60-handle range Monday through Thursday and then dropped 50 handles on Friday alone. The VIX moved from 17.9 Monday open to 17.4 Thursday close, then spiked to 19.6 by Friday's settlement — the volatility shift came after the damage, not before it.

There were three macro inputs worth flagging because both models referenced them in their pre-trade analyses. First, Pending Home Sales on Tuesday 5/19 came in at +3.4% (consensus +1.2%) — the upside surprise produced the brief Tuesday morning equity bounce that GPT shorted into, which became the Trade of the Week. Second, FOMC minutes Wednesday afternoon were read as moderately hawkish — three voting members floated keeping rates restrictive into Q3, and the curve flattened modestly into Thursday. Third, the late-week tape: Friday 5/22 brought no scheduled US data, but at 10:14 ET, an unscheduled Treasury official statement on USD strength versus trading partners triggered a sharp DXY surge and a synchronized equity sell-off. That single 15-minute window is responsible for most of Friday's combined −$5,000 across the two models.

The rate-sensitivity backdrop also explains why USDJPY was the week's graveyard. With US10Y holding 4.66%-4.72% and BoJ rhetoric staying dovish, the carry case for long USDJPY was structurally intact — but the pair itself traded in a 78-pip range all week, refusing to extend in either direction. Both models took the carry trade five times. They got one partial winner and four full stops. The setup logic was correct; the realized vol simply wasn't there.

What the models couldn't price ahead of time was the cluster of intra-week non-events. There was no clean directional catalyst from Monday through Thursday afternoon. Then a single off-calendar headline on Friday rewrote the week's P&L. That is the kind of tape that exposes any system that depends on follow-through after entry — and both Claude and GPT depend on follow-through.

About reported results. Each model outputs three take-profit targets (TP1, TP2, TP3) per trade. In live execution, models typically scale out at TP1 for risk management — the broker position records this as a TP1 exit. The R-multiples and dollar returns shown in this editorial reflect the full potential of each trade: where the market actually traveled to (the highest take-profit hit, or stop loss) before the setup was invalidated or exhausted. This lets readers see the complete arc of each setup, not just where the position was closed.

Equity Curve

May 18 — May 22

Claude's Week: Patient at the Open, Punished by Attrition

Claude opened the week with the cleanest single-session performance either model produced all week and then proceeded to give it all back across five sessions. The shape of the week matters more than the closing number: this was not a week where Claude was wrong about the tape. It was a week where Claude was right early and then asked the same question four more times against a market that had stopped answering.

Monday 5/18 was Claude's session. Three trades, two TP2 winners — USDJPY long taken at the Tokyo close handoff for +$940 (TP2), NAS100 short on the 9:45 ET breakdown for +$835 (TP2), and one US30 long that stopped out at the OR low for −$1,000 (SL). Net Monday: +$775. The two winners were not lucky entries. The USDJPY long had a 60m EMA stack confirmation, RSI 58, and a clean 5m base. The NAS100 short was a textbook OR-break-and-retest with NYAD below its 5-day EMA. Claude saw the setups, scored them at 6/7 and 5/7 respectively, and the market paid both.

What happened next is the part worth studying. Tuesday 5/19 Claude took USDJPY long again — same instrument, same direction, second time in two days. The first leg had hit TP2. The second leg, entered seven hours later off a similar 5m base, stopped out for −$1,000 (SL). This is not framework drift. Claude's pre-trade analysis was nearly identical to Monday's; the difference was that DXY had broken to a new five-day high overnight and the carry advantage no longer dominated the order flow. The setup signature was intact. The regime had thinned.

Wednesday through Friday Claude took three more setups, one per session, all stop-outs. EURUSD short Wednesday at the European close handoff for −$1,000 (SL). NAS100 short Thursday on a re-test of the 5/18 breakdown for −$1,000 (SL) — the same setup that had won Monday now failed because the underlying volume profile had hollowed out. US500 long Friday at 10:11 ET for −$1,000 (SL), stopped out three minutes after the off-calendar Treasury statement.

The fingerprint across the five losing trades is consistent: each entry was structurally valid, the risk was sized exactly to 2%, the stop was placed at the right level. None of the losses came from late entries, oversize positions, or rule violations. Claude's week was an attrition week — five $1,000 losses delivered by a market that produced one tradeable session followed by four non-trending ones. The losing streak is real (5/19, 5/20, 5/21, 5/22, plus the 5/18 US30) but the cause is environmental, not procedural. The framework asked the right questions and got a sequence of "no" answers. That is the cost of running the framework honestly.

GPT's Week: Active, Inconsistent, Saved by One Tuesday

GPT traded ten times in five days — three more entries than Claude — and produced the week's biggest single winner alongside its biggest single losing day. The character study of GPT this week is not "more aggressive than Claude." It is "more willing to re-enter the same instrument after a stop-out." Whether that is a feature or a flaw depends on which of the re-entries the reader weighs heaviest.

The week's centerpiece was Tuesday 5/19. GPT entered three trades that session: NAS100 short for −$1,000 (SL) before 10:00 ET, GBPUSD short for −$1,000 (SL) on a failed 1.2680 retest, and then US30 short — the Trade of the Week — for +$1,650 (TP2). That third trade is the one the rest of the week orbits around. Entry 49,525 after a failed reclaim of 49,531 VWAP, stop at 49,627, TP1 at 49,360, TP2 at 49,265. The setup scored 5/7 on GPT's confluence gate, with NYAD at −1,019 below its 5-day EMA, VIX at 18.0 above EMA, and the Trend Authority Agent posting 66% bearish confidence. Both prior US30 candidates that morning had failed; this one held the entry zone for eleven minutes and then moved cleanly to TP2. It is the only trade either model produced all week that ran more than one R past TP1.

Wednesday 5/20 was GPT's mixed session. Three USDJPY long attempts across Wednesday and Thursday morning all stopped out for −$3,000 (SL × 3). Sandwiched among them was a US500 long for +$938 (TP2) on Wednesday's post-FOMC-minutes bounce — the only winner among the five Wednesday-Thursday trades.

Friday 5/22 was GPT's catastrophic session and the one that decided the week. Five trades, four stop-outs, one overnight TP3. EURUSD short −$1,000 (SL), USDJPY long −$1,000 (SL), the US500 long that the data shows as htp=1 but exit_reason=sl_hit for −$1,000 — price tagged TP1 intra-bar, didn't fill enough volume to register a partial exit, and reversed through the stop. That is the one trade in the week where the realized dollar outcome (a full SL hit) diverges from the reported R-multiple (+1.15R if it had filled at TP1). Then a US30 long for −$1,000 (SL) before midday. Friday's only saving grace was an overnight US30 long that resolved to TP3 +$585 — the trade was open into Saturday, the position closed early Asia hours, and it is the smallest winner of the week.

The GPT character study, then: more willing to re-enter, more willing to take the same instrument three sessions in a row (USDJPY), and partially rescued by a single high-conviction Tuesday short. Without the US30, GPT's week is −$5,477. With it, it is −$3,827. The framework neither broke nor distinguished itself. It produced one outstanding trade and absorbed nine ordinary outcomes.

Claude vs GPT: Week 1 Results

Season 2 began on Monday 5/18 with two AI traders, $50,000 each, and the same head-to-head methodology that ran for fourteen weeks of Season 1. By Friday 5/22's close, both accounts were in drawdown. Claude finished at $46,775 (−6.45%). GPT finished at $46,173 (−7.65%). The combined deficit across both models is −$7,052 on seventeen trades, a blended win rate of 35.3% by R-multiple accounting. For context, both models finished Season 1 in profit. This is the first week in the benchmark's recorded history where Season N starts with both models red.

The honest read on the week is that nothing broke. Both models stayed inside their 2%-per-trade risk envelope. Neither blew through a stop. Neither doubled down after a loss or sized up after a win. The framework's rule set held cleanly across all seventeen entries. What did not hold was the assumption — implicit in any trend or breakout system — that a setup that scores 5/7 or 6/7 on a confluence gate will resolve in the direction of the setup more often than not. In this week's tape, it did not. Claude won 2 of 7 (28.6%). GPT won 4 of 10 (40%) by raw count, but one of those four wins was a 5/22 US500 long that recorded TP1 hit yet exited at the stop loss — the broker's blotter shows a full $1,000 loss on that ticket, so the realized win rate on dollars is closer to 30%.

Why Both Models Closed Red

Three things converged. First, the macro tape produced no clean catalyst Monday through Thursday afternoon, then delivered a single 15-minute Friday window that rewrote the week's P&L. Both models were correctly oriented bearish into Friday's sell-off — Claude was long US500 and GPT was long US500 — but both had taken the long before the off-calendar Treasury statement that triggered the move. The setup was reasonable; the headline beat the setup.

Second, USDJPY became the week's graveyard. Five total attempts across both models, four with the same long direction (Claude 5/18, Claude 5/19, GPT 5/20, GPT 5/21, GPT 5/22). One TP2 winner (Claude 5/18, +$940 (TP2)) and four full stops. The instrument structure favored longs — US10Y held 4.66-4.72%, BoJ stayed dovish — but realized intraday vol stayed under 80 pips every session. The carry case was right; the volatility to express it was missing.

Third, the only trade either model produced all week that ran more than one R past TP1 was GPT's US30 short on Tuesday 5/19. That trade is the Trade of the Week and it carries the entire week for GPT. Without it, GPT's week is −$5,477 instead of −$3,827. Without it, the framework would look worse than the methodology actually performed.

The Five-Loss Claude Streak

Claude's opening Monday produced +$775 net across three trades — the best single-day performance either model recorded all week. Then Claude lost five trades in a row across four sessions: USDJPY long Tuesday (SL), EURUSD short Wednesday (SL), NAS100 short Thursday (SL), US500 long Friday (SL), plus the Monday US30 long that started the streak. None of the losses came from rule violations. Each entry was structurally valid, sized to 2%, and stopped at the level the analysis pre-committed to.

What is worth noting — and what the methodical read of this streak requires — is that Claude's Tuesday USDJPY long was almost the same setup as Monday's USDJPY long. The pre-trade analysis cited the same EMA stack, the same RSI band, the same support shelf. The difference between the two trades was the DXY: by Tuesday morning DXY had broken to a fresh five-day high, and the carry-driven follow-through that worked Monday afternoon did not work Tuesday morning. The setup signature stayed intact. The regime around it thinned. A more aggressive system might have flagged the DXY shift and skipped the trade; Claude's framework does not currently weight DXY shifts heavily enough to override an in-pair confluence pass. That is a notebook entry for Season 2 calibration, not a failure of execution.

The Tuesday Trade and the Friday Rug

Tuesday 5/19 is the only session where the week's framework looked unambiguously good. GPT entered US30 short at 49,525 after a 5/15 base failed to reclaim 49,531 VWAP, with NYAD at −1,019, VIX at 18.0, and Trend Agent posting 66% bearish. The trade reached TP1 at 49,360 in nine minutes and continued to TP2 at 49,265 in another seventeen — total +$1,650 (TP2), +1.65R. Every signal that fired pre-entry kept firing through the move. This is what the methodology is designed to capture: a setup with multi-source confluence in a regime that gives the move room to extend.

Friday 5/22 is the inverse case. Both Claude and GPT were long US500 at the open. Claude took the long at 10:11 ET, eleven minutes before the off-calendar Treasury statement. GPT took the long at 10:09 ET. Both stops sat just below the prior swing low at 6,742. At 10:14 ET, DXY surged forty pips on the headline and the SPX cash dropped 25 handles in eight minutes, taking both stops out within thirty seconds of each other. The data shows GPT's position tagged TP1 momentarily on the way down — the partial fill recorded as htp=1 in the trade journal — but the exit reason logged is sl_hit because the protective stop took the entire remaining position before the partial unwound. The broker's realized P&L on that ticket is −$1,000.

The two trades — Tuesday's clean +$1,650 (TP2) and Friday's convergent stop-out pair — frame the week. When the tape extends, the framework prints. When the tape doesn't extend, or when an off-calendar headline rewrites it intra-trade, the framework absorbs a full-R loss. There is no middle gear.

What the Numbers Don't Show

A note on what is not in the head-to-head table: there were zero rule violations across seventeen trades. No oversized positions. No averaging into losers. No stop-runs on the way out. Both models traded within the methodology cleanly. The week's drawdown is the methodology working as designed against an unfavorable tape, not the methodology failing. Whether that is reassuring or alarming depends on the reader's prior. The benchmark's job is to show both readings without choosing.

The Trade of the Week: GPT US30 Short, Tuesday 5/19

GPT's US30 short on Tuesday 5/19 is the Trade of the Week not because it was the largest winner — though +$1,650 (TP2) is the largest single realized dollar gain across either model all week — but because it is the only trade in the seventeen-entry sample that ran more than one R past TP1 and held all four confluence signals through the entire move. It is the trade the framework was designed to capture, and it is worth studying as the cleanest expression of the methodology in the data so far this season.

The setup formed in NY AM after a sharp 5/19 open. US30 had gapped down at 9:30 ET, immediately bounced toward the VWAP cluster near 49,579-49,601, and then failed at the 49,601 ceiling. The bounce ran out of breadth — NYAD held at −1,019 against its 5-day EMA of −463 — and the 5m candles printed a clear lower high at 49,580 by 10:02 ET. The bounce-and-fail pattern is the GPT framework's preferred short setup on a breadth-driven index, and the 49,524-49,531 entry zone was the breakdown shelf below which the next leg sits unprotected.

Three things made this entry different from the other six US30 entries either model considered during the week. First, the Macro Agent's US30-specific bias was only 57% (neutral), but the group bias was 64% lean-bear — the asymmetry meant Macro wouldn't veto the short but also wouldn't be the loudest voice. Second, the Trend Authority Agent posted 66% bearish with a transitioning regime tag — strong enough to justify entry, soft enough to warrant tighter management. Third, the OR low at 49,265 sat exactly at the previous session's intraday low — a confluent magnet that gave TP2 a structural reason to fill.

What the reader should watch for in the broker-execution card below is the gap between the framework's pre-trade scoring (5/7, Medium-High, 7.0/10) and the realized outcome. The framework rated this trade as a B-grade, not an A-grade. The trade printed an A. That gap is the part of the methodology that depends on the tape, not the scoring — and it is the part that didn't show up for the other sixteen entries this week.

Account Performance

Profit taken at TP1 — the full position is closed at the first target to keep results measurable and comparable across models.

What to Watch in Week 2

Week 2 opens Monday 5/25 with a holiday-shortened US session — Memorial Day means no Monday cash equity trading, so the meaningful tape doesn't begin until Tuesday 5/26. Three macro inputs will set the week's posture. First, May Consumer Confidence on Tuesday 10:00 ET — consensus 99.5 vs prior 97.8. A miss below 96 reinforces the late-week risk-off lean from 5/22 and gives both models a structural case for short equity setups; a beat above 102 forces a reversal-of-bias check that neither model has had to run yet this season. Second, GDP Q2 revision Thursday 8:30 ET — consensus +2.4% vs prior +2.6%. A downward revision below 2.0% is the kind of print that historically widens DXY and pressures equities; an upward revision tightens the carry case for long USDJPY again, which both models will be tempted to retry after this week's 1-for-5 record.

Third, and most consequential, is Friday 5/29's PCE deflator — the Fed's preferred inflation read. Core PCE consensus is 2.5% YoY. A print above 2.7% would functionally close the door on a Q3 rate cut and re-price the curve; a print at or below 2.4% reopens the dovish case and likely produces a coordinated equity bid plus DXY softness.

For Claude, the watchable behavior is whether the framework calibrates against the DXY-shift problem that turned Monday 5/18's USDJPY win into Tuesday 5/19's USDJPY stop. If Claude takes a third USDJPY long attempt early in Week 2 without weighting DXY position more heavily, that's a flag. For GPT, the watchable behavior is whether the framework re-attempts US500 longs into the same structural setup that produced Friday 5/22's stop, or whether the post-mortem widens the entry-zone discipline. Neither model has been tested on a Memorial Day-shortened week before in the benchmark's recorded history, so the Tuesday-Friday compression may itself be the variable worth watching most closely.

Methodology

This weekly editorial aggregates trading results from May 18-22, 2026. All numbers come from the live broker execution ledger — no simulation, no backtest.

How P&L is computed. Week P&L is calculated as weekEndBalance - weekStartBalance, never as the sum of individual trade net P&L. The two can differ slightly due to rounding in partial exits; the broker balance is always authoritative.

Week rollover. Each week's starting balance is the previous week's ending balance. Week 1 uses the experiment's initial capital ($50,000 per model). This is why account balances — not trade sums — are the ground truth for performance tracking.

Net R vs. Net P&L. Net R is a risk-adjusted measure (sum of each trade's reward/risk multiple). Net P&L is the literal dollar change in account balance. Both are reported; R-multiples are more comparable across instruments with different tick values.

Weekend handling. Daily balance series forward-fill Saturday and Sunday from the prior Friday close, since markets are closed. This keeps chart visuals continuous without fabricating activity.

Methodology stability. Rules don't change mid-phase. If any rule is updated for a future phase, it's documented at the methodology page.