March 2026 Monthly Recap: 35 Trades, Three Model Phases, Break-Even Net

Restated: Gold (XAUUSD) was part of SkyAnalyst's coverage from inception (Jan 12, 2026) through May 2026. We've since narrowed coverage to six instruments — EURUSD, GBPUSD, USDJPY, US30, NAS100, US500 — and these numbers are restated for the current lineup. The original publish date is preserved; cumulative figures have been recomputed.

March 2026 was the second full calendar month on the published record, and the first to cycle through three different models inside a single window. Mar 1 to 16 ran on GPT-5, the carryover engine from the launch era. Mar 17 to 25 introduced GPT-5.4 (gpt-5.4-2026-03-05) as a nine-day A/B replacement. Mar 26 onward the master automations consolidated onto Claude Opus 4.6. The system took thirty-five trades across the six canonical instruments, closed at minus 0.36R net, and produced a 51.4 percent win rate. Eighteen winners, seventeen losers. The simulated $100,000 account at 2 percent risk per trade ended March at $99,273.27, essentially flat. Through Apr 1, 2026, the system has banked +8.64R YTD across 62 trades from Jan 12 inception. The simulated $100,000 account at 2 percent risk per trade sits at $117,283.56 on a static basis. The break-even number is the surface reading. Underneath it sat both extremes: a Mar 16 to 24 stretch that drew the equity curve to roughly minus 9R on the loss side, and a Mar 25 to 27 closing run that produced six TP3 winners across three sessions and clawed plus 4.62R back. One methodology note up front: every R-multiple here is computed on a TP1 exit for every winner. That is our conservative baseline. A subscriber running the published scale-out plan would have closed March meaningfully positive because twelve of the month's winners ran past TP1 to TP3 in the live broker fills. The minus 0.36R is the floor of the credited projection, not the realized ceiling.

Act 1: Mar 2-15, GPT-5 carryover and the first cold week

The month opened on the launch-era engine. GPT-5 carried the first ten trades of March under the prior posture. Mar 2 ran a US30 short to TP1 for plus 1.2R. Mar 4 produced two winners: a US500 long to TP3 for plus 1.25R credited and a NAS100 long to TP1 for plus 0.93R. The opening week closed at plus 1.38R on 60 percent across five trades.

Then the tape turned. Mar 9 to 13 produced four trades and zero winners on the same GPT-5 engine: a US30 short on Mar 10, a US500 long and a NAS100 long on Mar 11, a NAS100 long on Mar 13. All four stopped. The first all-loss week of the published record closed at minus 4R. Cumulative MTD moved from plus 1.38R to minus 2.62R into the weekend. The GPT-5 carryover panel closed its March run on Mar 16 with the US500 long that hit TP1 for plus 1.5R, the largest credited single winner of the month.

Act 2: Mar 17-25, GPT-5.4 introduced into the drawdown

Mar 17 the master automations rotated to gpt-5.4-2026-03-05. The first nine days under the new model produced the deepest drawdown on the published record. Across Mar 17 to 20 the GPT-5.4 panel fired twelve entries and stacked seven minus 1R losers. The cumulative drawdown widened day by day: minus 4.12R by Mar 17 close, minus 3.99R after the Mar 18 EURUSD short pair clawed back ground, minus 5.03R by Mar 20 close. The Mar 16 to 20 week landed at minus 2.41R on 41.7 percent. Three TP3 winners inside the same week, the Mar 18 EURUSD short pair and the Mar 20 USDJPY long, kept the damage from compounding further.

The bottom came on Tuesday Mar 24 at 14:54 UTC. Three Mar 24 stops in a row (US500 short, US30 short, NAS100 short) pushed cumulative MTD to minus 7.24R, the deepest published reading of the year. From the Mar 4 peak of plus 2.38R the equity curve had traveled roughly 9.6R on the loss side over thirteen sessions, the first nine of them on GPT-5, the last four on GPT-5.4.

Mar 25 began the GPT-5.4 panel's recovery sequence. Inside a 21-minute window the system cleared three TP3 winners: NAS100 short for plus 0.7R credited, US500 short for plus 0.97R credited, USDJPY long for plus 0.9R credited. Plus 2.57R in one session, all attributed to GPT-5.4. That was the GPT-5.4 window's last full day before the model rotation.

The architecture did not change through any of it. The Macro Agent did not flip its gate to flat. The Trend Agent did not raise its threshold. Risk did not cut size. Same library, same evaluation rhythm, two different OpenAI checkpoints reading it. The reading: the drawdown was not a model symptom, it was a regime symptom.

Act 3: Mar 26-31, Claude Opus 4.6 takes over

Mar 26 the master automations switched to Claude Opus 4.6 (claude-opus-4-6). The Claude panel opened with two TP3 hits inside twenty-five minutes: US500 short for plus 0.99R credited at 14:16 UTC, EURUSD short for plus 1.09R credited at 14:40. Mar 27 closed the week with the largest TP3 print of the month under Claude: the US500 short pullback to opening range for plus 1.28R credited.

Across Mar 23 to 27 the system took eleven trades for a 72.7 percent win rate at plus 3.72R net. Six ran to TP3. The closing-week reversal more than offset the Mar 9 to 13 cold streak and the Mar 17 to 18 stop cluster, but the deeper Mar 16 to 24 drawdown left the cumulative March ledger short of break-even by the time the book closed.

The Mar 30 to 31 bridge ran on Claude as well. Mar 30 stopped a USDJPY short for minus 0.25R. Mar 31 produced two TP3 winners (USDJPY short for plus 0.3R credited, EURUSD long for plus 0.75R credited). The rest of the Mar 30 to Apr 5 calendar week lands in April's recap. We close the March book at minus 0.36R net, with the master automations now consolidated on Claude Opus 4.6 for April.

Case studies from the closing run: Mar 25 US500, Mar 25 USDJPY, Mar 26 US500, Mar 26 EURUSD, and Mar 27 US500.

Pattern of the week

The pattern that defined March was the fade into resistance against a confirmed intraday counter-trend, applied across instruments and across all three model phases. Most of the closing run's TP3 winners fit this template: VWAP rejections, pullbacks to broken support, fades of counter-trend squeezes. The Mar 25 to 27 sequence ran the same setup on US500, EURUSD, NAS100, and USDJPY in turn and got paid on each, with GPT-5.4 firing the Mar 25 leg and Claude Opus 4.6 firing the Mar 26 to 27 legs off the same setup library.

Why the same pattern lost in the middle and won at the end

The cleanest signal in March is that the setup library did not change between the minus 9R loss-side stretch and the plus 3.72R closing run, and it did not change when the model rotated. From Mar 17 through 24 the same VWAP-rejection logic kept stopping out because price kept reclaiming the rejected level inside the next session. From Mar 25 through 27 the rejections held and the moves followed through to TP3. The system did not adapt to the regime; the regime stopped fighting the system. That is how a positive-expectancy system survives variance. It does not consult the recent loss tally to size the next entry. Each entry was sized off its own confluence read, independent of every entry that came before it.

Decision highlights

The Mar 17 sequence is the cleanest example in March of the system holding posture against a hostile tape, and it landed on the first day of the GPT-5.4 window. Inside one 26-minute window it fired a US500 long at 14:10, a NAS100 long at 14:31, and a US30 long at 14:36. All three stopped at minus 1R. The Mar 18 morning fired another USDJPY long that stopped, then a EURUSD short pair that ran to TP3. Each entry used the same threshold, the same sizing, and the same evaluation rhythm. The system did not "learn" from the four-trade losing string and tighten on the next setup, and the model rotation a day earlier did not alter that behavior.

The Mar 25 sequence demonstrated what the same logic looks like when the tape pays, again on GPT-5.4. Inside a 21-minute window it fired a NAS100 short at 14:11, a US500 short at 14:14, and a USDJPY long at 14:32. All three ran to TP3 for a combined plus 2.57R credited. Three decisions inside one session, each independent, each sized off the same confluence math that had stopped seven trades in a row earlier in the week. A discretionary trader emerging from the Mar 16 to 24 drawdown would have hesitated on the third entry after the first two had already cleared. The system did not.

The Mar 26 model rotation was the cleanest decision moment of the month. The master automations consolidated onto Claude Opus 4.6 the day after the Mar 25 recovery cleared, and the first Claude session printed two TP3 winners across correlated setups inside twenty-five minutes (US500 short for plus 0.99R credited at 14:16 UTC, EURUSD short for plus 1.09R credited at 14:40). Cross-Asset flagged the correlation and let both through because the structural premises were independent at the bar level. That cluster, plus the Mar 27 US500 follow-through, is the evidence base for keeping the master automations on Claude through April.

Loss worth learning from

The cleanest loss the month produced was not a single trade but the Mar 17 to 18 cluster, the first 24 hours of the GPT-5.4 window. Four entries inside one day: US500 long, NAS100 long, US30 long, USDJPY long. All four stopped at minus 1R. Same setup library, same threshold, same Macro Agent gate as the EURUSD shorts that printed two TP3 winners in the same window.

What the system saw that was right

A continuation pullback regime that had paid on Mar 16 (plus 1.5R US500 long to TP1 on GPT-5) and would pay again on Mar 18 (EURUSD shorts to TP3 on GPT-5.4). The Trend Agent's confluence cleared threshold on each of the four losers. The Macro Agent's gate had not flipped. Cross-Asset confirmed correlations were inside tolerance. The model rotation the day before did not alter any of these reads.

What the system got wrong

The Mar 17 tape inverted intraday in a way the system's evaluation rhythm could not detect ahead of time. The 14:10 US500 long, 14:31 NAS100 long, and 14:36 US30 long were three correlated equity-index continuation entries fired into a session that had set up like a continuation but resolved as a fade. Cross-Asset did not veto because the structural premises were independent at the bar level. Three correlated stops followed inside half an hour. The same outcome would have landed under GPT-5; the model swap did not cause the stop cluster.

What we would do the same

The entry rule is the entry rule. The system does not consult the recent trade record to size or decline a new entry. If the inputs cleared threshold, the entries were correct under the same logic that took the Mar 25 to 27 TP3 winners (GPT-5.4 on Mar 25, Claude Opus 4.6 on Mar 26 and Mar 27). Streak-aware filters would lower realized expectancy across calendar windows, not raise it. The Mar 16 to 24 drawdown was the cost of holding posture; the plus 3.72R closing run was the payoff. Both came from the same playbook, and the playbook survived the model rotation intact.

From the desk

Through Apr 1, 2026, the cumulative ledger reads +8.64R YTD across 62 trades from Jan 12 inception. The same $100,000 account at 2 percent risk per trade sits at $117,283.56 on the static line and $117,333.79 on the compounded line. The spread between the two is roughly $50 on a year-to-date basis, the natural consequence of compounding a positive-expectancy edge through 2 percent risk when winners cluster around losses. Disciplined sizing is what keeps that spread small and the path stable; the compounded line traces the static line within a fraction of a percent because no single month dominates the ledger.

The honest reading of March is that the system delivered exactly what a positive-expectancy playbook delivers across a 35-trade window: variance both directions, posture held through both swings, and a net result close to zero with twelve TP3 winners the conservative baseline systematically under-credits. The Mar 16 to 24 drawdown reached roughly minus 9R peak to trough on the loss side. The Mar 25 to 27 closing run paid plus 4.62R on five sessions. Same architecture across both. Same playbook. Same threshold. Three different models. The architecture survived all three.

The simulated $100,000 account closed March at $99,273.27, roughly seven dollars off the starting balance per thousand. Twelve of March's eighteen winners ran to TP3 in the live broker fills, six inside the Mar 25 to 27 stretch alone. A subscriber on the published scale-out plan would have closed March meaningfully positive on the same trades. The credited recap underweights those runners by design.

What carries into April is the same playbook, now running on Claude Opus 4.6 across the full month. The system does not tune based on the calendar month; it tunes based on the rolling 100-trade window. February gave us a clean look at the architecture working through a regime rotation. March gave us a clean look at the architecture surviving variance both directions inside one window, and at the master automations consolidating on Claude after a three-stage model transition. Both readings are data. Neither is a verdict.

Weekly views: Mar 2-8, Mar 9-15, Mar 16-22, Mar 23-29, Mar 30 to Apr 5. Prior monthly: February 2026 monthly recap.

What we're tuning

March did not produce a tuning signal at the setup level. A 51.4 percent win rate at minus 0.36R net across thirty-five trades is inside the rolling-100 distribution for similar setup mixes, and the dispersion across instruments (US500 plus 2.72R, NAS100 minus 3.13R) is the kind of variance the playbook implies. Tuning on a 35-trade sample would be over-fitting noise. The right horizon for any tuning decision is the rolling 100-trade window.

What we did tune was the model layer. Three engines through the month (GPT-5 Mar 1 to 16, GPT-5.4 Mar 17 to 25, Claude Opus 4.6 from Mar 26 forward), and the consolidation decision was settled by Mar 26. April runs on Claude across the full window. The forward-looking variables are unchanged: whether NAS100 stabilizes or the drag continues, whether the Mar 25 to 27 cross-instrument TP3 cadence repeats once Claude has more entries on its own, and whether the TP1-versus-TP3 spread on the runners stays as wide as twelve TP3 winners across thirty-five entries suggests. None are immediate tuning items. They are the variables that, if they shift, will eventually surface as a signal worth acting on.