2026 Year-to-Date: 121 Trades, +19.99R, the Year SkyAnalyst Found Its Cadence

One hundred and twenty-one trades since system inception on January 12, 2026. Seventy winners, fifty-one losers, +19.99R net on the TP1 baseline. That is the year-to-date scorecard through May 31, 2026 across the canonical-instrument set, a 57.9 percent win rate with an average winner of +0.93R against an average loser of -0.91R. A $100,000 simulated account at 2% risk per trade sits at $139,996 (static) or $145,328 (compounded) — a +39.99 percent year-to-date return on the static figure, or +45.33 percent on the compounded figure that reinvests risk against the running balance. The story this year tells in 121 trades is not a single dramatic print. It is the story of a system that launched soft, ramped through staged model rollouts, consolidated by end-March on a primary execution model, and settled into a stable cadence through April and May. The architecture did not change. The Risk Agent's fixed-R sizing rule did not change. The Trend Agent's confluence floor did not change. The Macro Agent's veto logic did not change. What changed was the cadence — which models earned execution time, which instruments became the desk's reliable edge, and how the team learned to read the same confluence cards across different model snapshots. This is the article we want a reader evaluating SkyAnalyst to read first. Every trade in the trade index below is one we published in real time. The +19.99R figure is not a backtest or a curated subset; it is the arithmetic of every master-automation trade the system has logged since launch. The simulated balance is the same arithmetic applied to a standard 2 percent risk policy on a $100,000 base. The compounded figure shows what the same trade ledger produces when risk-per-trade scales with the running balance instead of staying fixed against the original $100,000 — a meaningful gap at 121 trades, and a wider gap if the year continues at the same hit rate.

Act 1: January — soft launch on three trades

Since launch on January 12, 2026, the system spent the back half of January in internal validation rather than published volume. Only three trades cleared the published gate in January, all on the Claude side under the pre-rebrand Opus 4.6 snapshot. The January contribution was +3.02R on a 100 percent win rate across the three entries — directionally encouraging but explicitly too small a sample to read. January closed at +3.02R YTD across 3 trades, the simulated static balance sitting at $106,040. The framing for January is not "incomplete month" but "since launch on Jan 12" — the inception date is the start of the record, not a gap.

Act 2: February — the first real ramp

February doubled down on volume. Twenty-four trades cleared the published gate that month, the first real sample-size month of the year. Claude carried more of the entries early; the GPT-side lineage (running pre-rebrand snapshots: c1/openai/gpt-5/v-20250930 attribution on Claude-side automations during the legacy era) began to appear on the trade index as the architecture's evaluation logic started routing entries across multiple model families on the same instruments. February contributed +6.64R for a 27-trade YTD position at +9.66R. The simulated static balance climbed from $106,040 at January-end to $119,320 at February-end, a meaningful gain on the first real ramp month. Win-of-the-month-class trades included a US30 long on Feb 11 for +2.64R — still the year's largest single GPT-lineage print and the YTD result reveal.

Act 3: March — full force and the Claude consolidation

March was the month every model family ran live in production, and it was the month the architecture's design choice to allow multiple model snapshots to compete on the same setups produced its clearest stress test. Thirty-five trades cleared the published gate, the highest single-month volume of the year through May. March's net was -1.02R — a slightly red month against a roughly flat distribution — but the more important read on March is structural: by the end of the month, the team had consolidated Claude Opus 4.7 as the desk's primary execution model. The analysis held the most consistent discipline across the regime shifts March's tape produced. From that point forward, the published Claude-vs-GPT comparison runs on cleaner attribution; the early-March D-grade GPT-lineage entries running pre-rebrand snapshots gradually retired in favor of the formal AI Trader lineup that activated May 18. March closed at +8.64R YTD across 62 trades; the simulated static balance held at $117,280 — slightly below February-end despite the higher trade count, the visible signature of a month the architecture worked through chop without giving back the February ramp.

Act 4: April and May — settled cadence on +11.32R combined

The closing two months ran on stable cadence. April produced 18 trades for +2.99R, a quiet positive month on the published expectancy band. April closed at +11.63R YTD across 80 trades; the simulated static balance climbed to $123,260. May produced 47 trades for +8.33R, the heaviest single-month contribution of the year. May closed at +19.99R YTD across 121 trades; the simulated static balance hit $139,996. The architecture did not change between April and May; the cadence did. April's flatter month and May's stronger month are two adjacent prints on the same distribution, the same Risk Agent sizing the same fixed-R policy across the same confluence floor. Over the four-act arc, the year tells the story of a system that launched on three honest trades, ramped through a 24-trade February, took its lumps and consolidated on Claude in a 35-trade March, and settled into a 65-trade April-May cadence that banked the majority of the YTD print.

Pattern of the week

The pattern of the year is the same architecture compounding through staged volume — three trades in January, twenty-four in February, thirty-five in March, eighteen in April, forty-seven in May. The confluence floor did not change. The Risk Agent's fixed 2 percent sizing did not change. The Trend Agent's grade threshold did not change. The Macro Agent's veto logic did not change. What compounded across 121 trades was the cumulative R-multiple — +3.02R at January-end, +9.66R at February-end, +8.64R at March-end, +11.63R at April-end, +19.99R at May-end — and the cumulative simulated balance, $106,040 at January-end climbing to $139,996 at May-end.

Inside those 121 canonical trades, US30 carried the volume contribution at 31 trades for +5.39R, US500 banked +5.13R on 23 trades at 60.9 percent win rate, GBPUSD ran cleanest at 6 trades and 83.3 percent win rate for +4.32R, EURUSD ran 17 trades for +4.16R at 58.8 percent, NAS100 ran the most volume at 36 trades for +1.13R, and USDJPY ran 8 trades for -0.14R as the only red column on the year. The 70 winners contributed +65.40R combined at the TP1 baseline; the 51 losers cost -45.40R combined (including the partial structural exits on USDJPY at -0.5R apiece). The net at +19.99R is the arithmetic of a 57.9 percent win rate with an average winner of +0.93R and an average loser of -0.91R.

How a year banks +19.99R on a flat $1R distribution

Seventy winners summed to +65.40R: the cumulative payoff of a hit rate just above 50 percent on a system whose winners average +0.93R per trade. Fifty-one losers cost -45.40R: the cumulative cost of a stop-loss policy that holds each invalidation near a fixed -1R. The net at +19.99R is the cumulative arithmetic of expectancy math. The compounded simulated balance at $145,328 versus the static balance at $139,996 — a roughly $5,300 spread — is the visible footprint of letting risk-per-trade scale with the running balance over 121 trades. Every winner reinvests at a slightly larger base, every loser reduces it. The spread will widen if the next 120 trades print on the same expectancy band.

Decision highlights

The end-of-March decision to consolidate Claude Opus 4.7 as the desk's primary execution model is the year's most significant operating choice. Through January and February the architecture allowed multiple model families to evaluate the same setups simultaneously; the Risk Agent treated each model's confluence read as an independent draw. By end-March the team had a 62-trade sample showing Claude's discipline held more consistently across regime shifts on the FX and NAS100 sides. The consolidation did not remove GPT from the architecture; it moved Claude to the front of the execution queue. The Apr-May period validated the choice: Claude has carried +6.89R of the +8.33R May print on 26 of the 47 May trades, and the year-to-date Claude netR sits at +14.24R against GPT's +5.76R.

The May 13 nine-trade session is the per-trade isolation read of the year. Nine entries inside one afternoon, across both Claude and GPT lineages, on four different instruments, including six D-grade GPT entries under pre-rebrand snapshot attribution. The architecture's evaluation logic treated each entry as an independent confluence draw regardless of grade or model family. The cluster produced six winners and three losers and contributed roughly +2.86R to the YTD column. The decision logic at trigger did not collapse correlated entries into a portfolio position; each was sized independently against equity-at-trigger. The May 13 print is the cleanest demonstration of the architecture's per-trade independence — a design choice that has compounded across 121 trades without smearing attribution.

The May 18 AI Trader formalization is the year's structural-architecture decision. The six Claude automations and six GPT automations went live as a single 12-instrument lineup with disjoint family attribution; the legacy model-snapshot leakage on Claude-side automations (the pre-rebrand GPT-5 and GPT-5.4 trades that bypassed strict family attribution via legacy mode) retired in favor of clean automation-ID-to-model-family mapping. The formalization did not change the entry logic, the sizing rule, or the confluence floor. It cleaned up the attribution layer so the Claude-vs-GPT comparison going forward runs on disjoint automation IDs rather than fuzzy model-snapshot lookups.

Loss worth learning from

The year's most-instructive single loss came from the May 27 GPT EURUSD long, which the May 25-31 weekly losses article tears down at length. The entry triggered at 14:42 UTC on a pullback-to-VWAP setup at B grade — the highest grade on any May trade. The Trend Agent saw a genuine confluence card; the Macro Agent did not veto; the Risk Agent sized at the standard 2 percent risk on equity at trigger. Confluence said the level was good; the runner the trade needed did not extend inside the entry window. The stop printed at a clean -1R at the documented invalidation.

The teardown is the cleanest illustration of the year's central theme: confluence at entry is a necessary condition for a winner, but never a sufficient one. Across 70 winners and 51 losers, the architecture made the case repeatedly that the runner is the second draw, and that no entry-side grade — including B grade — guarantees it. The single most-instructive drawdown stretch of the year came in the May 18-24 window: 18 trades for 7 winners and 11 losers and -2.82R net on the same architecture that printed +8.83R the week before. The 24-hour gap between a +5.96R week and a -2.82R week is the visible footprint of expected variance on a 57.9 percent win rate system; the published expectancy lives on both sides of its mean, and the year's most-instructive lesson is that the architecture's discipline is what compounds across that variance, not the variance itself.

From the desk

2026 year-to-date closes May at +19.99R across 121 canonical trades, 70 winners and 51 losers, 57.9 percent win rate. A $100,000 simulated account at 2% risk per trade sits at $139,996 (static) or $145,328 (compounded). The compounded figure is what an operator running the same trade ledger banks if they let the risk-per-trade scale with the running balance instead of fixing it at $2,000 against the original $100,000 base. The roughly $5,300 spread between the two figures is the cumulative footprint of disciplined fixed-fractional sizing across 121 trades — every winner reinvests at a slightly larger base, every loser reduces it. Over the next 120 trades, that spread will widen if the system continues printing on its current expectancy band.

The structural reading of the year is that the architecture has held its shape across four meaningfully different cadences. January's three-trade soft launch, February's first real 24-trade ramp, March's full-volume 35-trade stress test that consolidated Claude as the primary execution model, and the April-May 65-trade settled cadence that banked +11.32R of the +19.99R YTD print — all four ran the same Trend Agent confluence floor, the same Risk Agent fixed-R sizing, and the same Macro Agent veto logic. The architecture's design choice to let multiple model families evaluate the same setups independently produced the data that informed the end-March Claude consolidation; the May 18 AI Trader formalization cleaned up the attribution layer so the comparison going forward runs on disjoint automation IDs. None of these decisions changed the entry logic. They changed the cadence.

What an operator evaluating SkyAnalyst should take from the YTD column is bounded: a 57.9 percent win rate with an average winner near +1R and an average loser near -1R produces a positive expectancy on the published methodology, and 121 trades is a sample large enough to confirm the directional read without yet settling the longer-run band. The next several months will retest the Claude-vs-GPT split with cleaner attribution, the same-instrument repeat-exposure scope will progress, and the year's cadence will produce another 100-200 trades by end-of-year on which the conclusions will sharpen. The fixed-R policy and the published confluence threshold will not change. The architecture is unchanged across all 121 entries and stays unchanged going forward.

From the desk, the SkyAnalyst Team. Next monthly recap in early July; next quarterly check-in mid-July; next year-to-date refresh after every monthly close.

What we're tuning

The year's three concrete tuning candidates are all narrow, all scoped from specific multi-trade patterns, and all unshipped through May. First, the same-instrument repeat-exposure check the May 18-24 weekly losses report scoped and the May 25-31 weekly losses report extended — the Risk Agent currently evaluates each entry independently and does not apply a clustered-exposure discount when a second entry hits the same instrument inside a rolling 24-hour window. We are scoping this narrowly rather than retuning the Risk Agent globally; the validation runs against the full 121-trade record before any change ships.

Second, the D-grade GPT attribution review. The pre-rebrand GPT-lineage entries (the May 13 D-grade cluster, the legacy gpt-5.4 attribution that bypassed family checks via legacy mode in YTD walks) deserve cleaner separation in published reports going forward. The fix is not to retire the entries — they cleared the confluence card and contributed +1.55R on May 13 alone — but to surface them with explicit pre-rebrand attribution rather than rolling them into the formal AI Trader GPT-5.5 column. The formalized May 18 lineup makes this cleaner going forward; the legacy attribution will hold for the YTD aggregate until the historical entries roll out of the sample.