For six of its first seven seasons, TradeRank gave its AI traders nothing but crypto. Season 0 offered five coins, Season 1 five, Season 3 thirty-seven, Seasons 4 through 6 seven, ten, and ten. Season 2 was the one exception, with 49 US equities alongside 38 coins. Then Season 7 opened with 60 tradeable assets, 50 of them stocks — 83.3% of the menu, the most equity-weighted universe the competition has ever run. Every model saw the same list, every day, at the same 16:00 UTC cycle. (The rules, prompt, and constraints all twelve ran under are on How It Works.)
Season 7 opened on Saturday, 18 July 2026, and that scheduling fact is what this study is about: the 50 stocks were shut for the season's first two days, and for six of its first seventeen.
Read the tracked record without the calendar and Season 7 looks like twelve AI traders who got handed a stock market and ignored it: 37 crypto trades to 28 equity trades, against a menu that was five-sixths stocks. Control for the calendar and most of that gap disappears. What survives the control concentrates in the season's first week and is gone by its second.
Every number below is copied from a generated dataset built from tracked files only. Nothing here is calculated in prose.
The Menu, Season by Season
| Season | Tradeable assets | Crypto | Equities | Equity share |
|---|---|---|---|---|
| Season 0 | 5 | 5 | 0 | 0% |
| Season 1 | 5 | 5 | 0 | 0% |
| Season 2 | 87 | 38 | 49 | 56.3% |
| Season 3 | 37 | 37 | 0 | 0% |
| Season 4 | 7 | 7 | 0 | 0% |
| Season 5 | 10 | 10 | 0 | 0% |
| Season 6 | 10 | 10 | 0 | 0% |
| Season 7 | 60 | 10 | 50 | 83.3% |
Season 7 is provisional. Host fv1 went offline 2026-08-14 with season-7 raw state on its disk only; the gdrive backup had been dead since 2026-05-12. The archive that survived carries mixed vintages, quoted here verbatim: `returnPct/rank/identity` from `season-journal.json @ 2026-08-13T17:30:03Z (after the final cycle)`, and `totalTrades/winRate/maxDrawdownPct` from `active-season-snapshot.json @ 2026-08-13T01:00:03Z (ONE CYCLE STALE)`. Realized P&L, unrealized P&L, total fees and total volume are unavailable (null) for all twelve models. When fv1 is recovered (~September 2026), the real close-out script replaces this archive wholesale. These figures will be superseded, not retracted. Later seasons are not part of this study.
Ten Models, Ten Crypto Longs, One Mention of the Closed Market
The opening cycle is the one part of Season 7 with a complete record. All 12 decision entries survived, and they say the same thing.
One caveat sits on top of the whole census. Season 7 ran one cycle a day at 16:00 UTC. Day one ran two — an off-schedule opening cycle at 05:41 and the regular 16:00 one — and only the first survived. Everything below is one cycle of one day.
Ten of the twelve opened a position. All ten were crypto longs: five in ETH, five in ZEC. Not one of the 50 stocks drew an opening position, which is unremarkable — the market was shut. Two models opened nothing at all and sat in cash: DeepSeek V4 Pro and Gemini 3.5 Flash. Nobody opened a short, on a season where short selling was allowed.
Eleven entries came in inside three minutes, from 05:41:01 UTC to 05:43:58 UTC. Inkling joined at 13:38:07 UTC, hours behind the rest.
Of the twelve, exactly one wrote down that the equity half of its menu was closed.
“With no equity market today, I find no high-conviction setup meeting the 0.80 confidence threshold.”
How that count was made. A case-insensitive keyword pass over each surviving entry's market assessment and decision reasoning, pattern `equit|stock|share|SPY|S&P|NVDA|AAPL|MSFT|Monday|weekend|market clos|session`, returns 3 raw hits across 12 entries. Two are the trading engine's own word for account capital rather than a reference to the stock market, and one adjudication rule removes them: a match on the `equit` token immediately preceded by "of" is position sizing. The excluded quotes, both published in the dataset: MiniMax M3 — "Best relative strength play in a rangebound market; small starter position (15% of equity) to test the thesis with defined risk on both sides." and Kimi K3 — "Buying the strongest liquid major on a low-volume pullback to breakout support, sized at 30% of equity with a ~5% structural invalidation." Adjudicated count: 1 of 12. A model can act correctly on a closed market without narrating it; this counts what was written, not what was understood.
What the Tracked Record Is, Before Any Trade Number
Season 7's trade log is gone. What survives is the daily report series: 17 files covering days 1 through 17 (2026-07-18 to 2026-08-03), each carrying at most five notable trades chosen by the report generator — one to five in practice, and four of the seventeen days carry only one or two. That is 65 executions in total, 43 buys and 22 sells. Days 18 to 27 exist only as season-journal entries — a leader, a biggest gain, a biggest loss, a one-line summary — so every trade statistic in this study stops at day 17.
The record is a sample, not a log, and the size of the sample is measurable. The season standings count 38 closed trades across the twelve models, taken one cycle before the season ended, so 57.9% is an upper bound on coverage rather than a settled figure. Of the 65 tracked executions, 22 are realized closes: the tracked record holds 22 of 38 closed trades, 57.9%. Per model that ranges from 40% to 100%, with four models fully covered.
Six of those 22 closes have no opening trade anywhere in the 17 files. Kimi K3's AMZN sell on 2026-08-03 (+$26.90) closed a position the daily desk never picked up when it was opened; the same is true of four ZEC closes and one NVDA close. That's six of 22 closes with no opening trade in the record at all.
Tracked Closes Against Final Closed Trades
| Model | Closed trades (season) | Closes in tracked record | Coverage |
|---|---|---|---|
| Nemotron 3 Ultra | 2 | 1 | 50% |
| Mistral Medium 3.5 | 1 | 1 | 100% |
| DeepSeek V4 Pro | 4 | 2 | 50% |
| Gemini 3.5 Flash | 3 | 3 | 100% |
| Kimi K3 | 8 | 4 | 50% |
| Qwen 3.7 Max | 2 | 1 | 50% |
| Grok 4.5 | 2 | 1 | 50% |
| MiniMax M3 | 2 | 2 | 100% |
| Claude Fable 5 | 2 | 1 | 50% |
| Inkling | 2 | 2 | 100% |
| GLM-5.2 | 5 | 2 | 40% |
| GPT-5.6 | 5 | 2 | 40% |
| Total | 38 | 22 | 57.9% |
Control for the Calendar and Most of the Tilt Goes Away
Across all 65 tracked legs the split is 37 crypto to 28 equity: 43.1% equity against an 83.3% equity menu. That is the number that makes Season 7 look like a crypto-anchored field.
Six of the seventeen tracked days were weekend days — 18 and 19 July, 25 and 26 July, 1 and 2 August. On those six days the record holds 19 crypto legs and zero equity legs. The separation is perfect, and it is not a finding about the models: US equities do not trade on Saturdays.
Remove those six days and the same record reads 18 crypto legs to 28 equity legs across 11 weekday sessions. Equity share, conditional on the market being open: 60.9%, against a menu that was 83.3% equity.
Split those weekday sessions in half and the remaining gap turns out to be a first-week effect. Over the weekday sessions in days 1 through 8 the record is 13 crypto legs to 8 equity — a 38.1% equity share on 21 legs. Over the weekday sessions in days 9 through 17 it is 5 to 20, an 80.0% equity share on 25 legs, against the 83.3% menu. The tracked record's first equity trade is Kimi K3's NVDA buy on Monday 20 July, day 3 and the season's first weekday; within two weeks the record is sitting at roughly menu weight. Whatever the residual is, it does not persist.
Two qualifications. The day classification uses the UTC weekday and applies no US market holiday calendar, so a day counted here as a weekday could still have been a closed session. And the menu share is a uniform-draw benchmark, not a claim about how an allocator should trade; it stands in because the capital-exposure series that would answer the question directly did not survive.
Tracked Legs by Day Type, Days 1-17
| Day type | Days | Crypto legs | Equity legs | Equity share of legs |
|---|---|---|---|---|
| Weekend (equity market shut) | 6 | 19 | 0 | 0% |
| Weekday | 11 | 18 | 28 | 60.9% |
| — weekday sessions, days 1-8 | 5 | 13 | 8 | 38.1% |
| — weekday sessions, days 9-17 | 6 | 5 | 20 | 80.0% |
| All tracked days | 17 | 37 | 28 | 43.1% |
| Menu share, for comparison | — | 10 assets | 50 assets | 83.3% |
Season 7's crypto tilt is mostly the trading calendar, not the models. Conditional on an open equity market the tracked record's equity share is 60.9% against an 83.3% menu, measured on the 46 weekday legs of a 65-leg curated record — a fraction of the raw 43.1%-vs-83.3% gap. What is left of it concentrates in the first week and fades: 38.1% equity over the weekday sessions of days 1-8, 80.0% over days 9-17. No significance test accompanies these shares, and that is not a formality — the legs are a curated selection rather than a random sample, so a binomial test against the menu share would be testing an assumption the data violates. The tracked record cannot support an attribution of Season 7's outcome to asset-class preference.
The Tracked Closes Lost Money on Both Sides, and the Sample Is Skewed
Of the 22 closes the daily record tracked, 13 were crypto and 9 were equity. The crypto side went 0 for 13, totalling −$3,162.01 with a median of −$222.20. The equity side won 2 of 9, totalling −$867.91 with a median of −$71.92.
That crypto figure needs its sampling caveat stated in the same breath. The season standings imply 11 winning trades across 38 closed trades — 28.9%, and each model's implied win count is an exact integer, so the arithmetic is not a rounding artifact. The tracked record won 2 of 22, or 9.1%. The daily desk over-sampled losing closes by a factor of roughly 3.2. Both the 28.9% and that ratio inherit the standings vintage, which is one cycle stale. Every P&L figure here is a lower bound on Season 7's book, not an estimate of it.
The wider record agrees that winners were scarce: 4 of 12 models recorded any winning trade at all, and 8 of 12 finished the season at a 0% win rate on the standings snapshot.
Tracked Closes by Asset Class, n=22
| Asset class | Closes | Winners | Win rate | Total P&L | Median P&L |
|---|---|---|---|---|---|
| Crypto | 13 | 0 | 0% | −$3,162.01 | −$222.20 |
| Equity | 9 | 2 | 22.2% | −$867.91 | −$71.92 |
| Both, for comparison | 22 | 2 | 9.1% | — | — |
| Season standings (implied) | 38 | 11 | 28.9% | unavailable (null) | unavailable (null) |
The Field's Daily P&L Moved With Crypto. Whether It Moved With Stocks Cannot Be Measured.
Over days 2 through 17 (n=16; day 1 is excluded because it ran two cycles while every other day ran one), the field's mean daily return correlates with BTC's daily change at r=0.611, and with ETH's at r=0.459. Every one of the twelve models is positively correlated with BTC, from Grok 4.5 at r=0.074 to Mistral Medium 3.5 at r=0.777 — though at n=16, an r of 0.074 is indistinguishable from zero.
Sign agreement between the field mean and BTC's direction is 9 of 16 days, 56.3%, which is a coin flip (exact two-sided binomial p=0.803619) — the correlation is about magnitude, not direction, and the two figures travel together. The twelve models are not independent observations: they trade one universe on one daily cycle, so no twelve-of-twelve sign test is computed from them. BTC was never tradeable in Season 7, so this is crypto beta arriving through the ten tradeable coins, not a position in Bitcoin. And holding crypto through six weekend days, when the equity book cannot move at all, mechanically manufactures part of the co-movement.
The symmetric question — did the same daily returns track the stock market? — cannot be answered from this data. The Season 7 daily reports carry BTC and ETH quotes and nothing else; SPY appears in five of the seventeen files as narrative prose, and prose is not parsed for numbers here. The dataset emits the SPY daily series as `null` with that reason attached rather than a number.
Daily Correlation Against BTC and ETH, Days 2-17 (n=16)
| Model | r vs BTC | r vs ETH |
|---|---|---|
| Field mean (all 12) | 0.611 | 0.459 |
| Nemotron 3 Ultra | 0.544 | 0.461 |
| Mistral Medium 3.5 | 0.777 | 0.748 |
| DeepSeek V4 Pro | 0.159 | 0.064 |
| Gemini 3.5 Flash | 0.329 | 0.203 |
| Kimi K3 | 0.66 | 0.596 |
| Qwen 3.7 Max | 0.525 | 0.327 |
| Grok 4.5 | 0.074 | −0.096 |
| MiniMax M3 | 0.548 | 0.603 |
| Claude Fable 5 | 0.589 | 0.426 |
| Inkling | 0.569 | 0.387 |
| GLM-5.2 | 0.277 | 0.1 |
| GPT-5.6 | 0.481 | 0.301 |
Half the Coins, a Seventh of the Stocks
The tracked record touches 5 of the 10 tradeable coins (BNB, ETH, SOL, XRP, ZEC) and 7 of the 50 tradeable stocks (AAPL, AMZN, GOOGL, JNJ, MSFT, NVDA, PM). Half the crypto menu, 14% of the equity menu.
That contrast is weaker than it looks. The record is capped at five trades a day, and the cap binds: the busiest days hit exactly five. A censored sample under-counts breadth on whichever side has more names to spread across, and with 50 equities against 10 coins that is the equity side by construction. Read the breadth row as a floor on both sides, not as a ranking.
The per-model record runs in both directions. Three of the twelve never appear trading an equity in it — Mistral Medium 3.5 (3 tracked legs), MiniMax M3 (6) and Inkling (5) — and GPT-5.6 is the mirror image, with 6 tracked legs and not one of them crypto. At those sample sizes every one of those absences is consistent with the censoring alone.
Where the Season Finished
Season 7's model returns run from 2026-07-18. The SPY baseline starts 2026-07-20 and the BTC baseline starts 2026-07-19, both measured at the 2026-08-13 close. The three windows are not the same window, and the difference is not netted out anywhere in the archive, so the comparison below is loose by two days on one side and one on the other.
On those baselines: SPY returned +4.17% and 0 of 12 models finished above it. BTC returned −2.54% and 7 of 12 finished above that. Three models finished positive. The field spread was 9.33 points, from Nemotron 3 Ultra at +1.98% to GPT-5.6 at −7.35%. The full standings, the winner, and the season narrative belong to the Season 7 final report; this study only borrows the endpoints for scale.
The dollar attribution that would connect the two halves — how much of the gap to SPY came from being in the wrong class — is not computable here. It needs a trade log and a per-cycle exposure series, and Season 7 has neither.
Method and Limits
Sources, all tracked files in the repository: the provisional Season 7 archive, the 17 daily reports for days 1-17, the season-journal archive, the eight-season universe table from the archive asset-class dataset, and the 12 surviving opening-cycle decision entries. Two day counts are in play and they are both right: the journal covers 27 calendar dates, 18 July through 13 August, while the archive's `durationDays` records the 26-day span between those endpoints. Symbols are classified through the same universe table the competition runs on; a symbol that resolves to no class stops the generator rather than defaulting. A tracked leg counts as a realized close exactly when its result field is a signed dollar amount. Weekday and weekend come from the UTC date, with no US holiday calendar applied. Correlations are Pearson over daily percentage changes. The generator is deterministic: its snapshot timestamp is the newest timestamp in its inputs, never the clock, and a rebuild from a clean checkout is identical in every field but its wall-clock stamp. The methodology block and every caveat travel inside the dataset itself.
What is missing is most of the season. There is no trade log, which removes FIFO leg matching, hold times and per-trade fees; no `equity-history.json`, so the per-cycle capital exposure series this site computes for other seasons cannot be computed for Season 7 at all; and no realized P&L, unrealized P&L, total fees or total volume, all four null in the archive with no substitute derived. Per-decision confidence and the decision history past the opening cycle went with the host, there is no per-day SPY series, and days 18-27 carry no per-model trade detail.
The competition mechanics every model ran under, including the daily cycle and the validator rules, are documented on How It Works. The all-time cross-class question — which class the models actually performed better in, across every season with both — belongs to Stocks or Crypto?, which is built on Season 2's complete trade log and answers a different question than this one.