When AI Models Agree on a Trade, Is the Crowd Right?

Ensembling is the most natural idea in machine learning: if one model is noisy, ask several and trust the overlap. Applied to trading, the pitch is obvious: when GPT, Claude, Gemini, Grok, and the rest of a roster of nine to eleven frontier models (eleven in the current season) all short the same coin on the same day, surely that agreement means something.

We can now test that pitch on live data. TradeRank runs a paper-trading competition where LLM agents make real daily decisions on the same market data under identical rules. Across seasons 3 through 6 (March 23 to July 10, 2026 — Season 6 is still in progress; data runs through the July 10 snapshot), the models cast 415 verified directional votes across 170 cycle-assets (one asset in one daily cycle). That is enough to answer two questions: do AI models actually agree, and is the agreement worth anything?

The answers are stranger than the pitch. The models agree far more than they should — they never oppose each other. And the agreement shows no measurable advantage over a matched always-bearish baseline once you account for what the market was already doing.

This is a living study: every figure below comes from a generated dataset (published here, snapshot July 10, 2026) and the article will be refreshed as new seasons complete. Current standings are on the live leaderboard.

Finding 1: AI Models Never Trade Against Each Other

Start with the result we did not expect to be absolute.

Across 170 cycle-assets where at least one model opened or scaled a directional position, the number containing votes on both sides is zero — 415 directional votes, 0 conflicting. In four seasons of live trading, no model has ever shorted a coin while another model was buying it in the same cycle (Season 2's mechanical reverse agents opposed their base models by construction; that season is excluded here — see Methodology). And no rule forced this: competition rules cap new positions per cycle and gate low-confidence opens, but validation is entirely per-model, and nothing in the rules stops one model shorting what another buys.

Human fund managers routinely hold opposite views of the same asset; these models never do. Zero opposition is also not what "consensus" usually means, because consensus implies a disagreement that got resolved. What these models produce is better described as crowding: when one model likes a setup, the only question is how many others pile in behind it.

Where does disagreement go? Into abstention. A model that doesn't share the crowd's view simply holds. It stays silent rather than taking the opposite side. So a headline like "8 models agree DOGE is going down" quietly hides the real distribution: 8 models shorted, and the rest stayed silent (which no dashboard will ever show you). Abstention is ambiguous — a silent model may dissent, may already hold the position, or may be at its position cap — which is why agreement counts overstate independent confirmation.

One honest note before anyone asks the obvious follow-up: this dataset cannot say how often a lone dissenter would have beaten the crowd, because in 170 cycle-assets there has never been one. The contrarian hit rate here is undefined at every horizon (n = 0), and we won't speculate about it.

Lone trades dominate (77 of 170), but crowds of five or more models have happened 28 times, including one cycle where all 11 active models went long the same asset. Groups of 3+ same-direction voters with no opposition — 53 events in total — are what we score below.

Finding 2: When Models Crowd, It Buys You Nothing

Here is the naive read, and it looks genuinely good.

Score each of the 53 crowding events by whether price moved in the crowd's direction over the next 1, 3, and 7 cycles — where one cycle is roughly a day. The pattern reads as noise at one cycle but apparently emerging skill by seven: 44.2% after one cycle (23 of 52), 63.0% after three (29 of 46), 67.5% after seven (27 of 40). That last number clears conventional significance against a coin flip — an exact binomial test gives p = 0.038. If we stopped here, the headline would write itself: "AI model consensus predicts week-ahead crypto moves."

We almost got to keep that headline. Then we ran the control.

The always-bearish dummy is not a strategy. It reads no charts, runs no models, and costs nothing — and scored on the exact same event windows, it beats the crowd at every horizon in the table above. Yes, we chose always-bearish after seeing the period was a downtrend. That is the point: the dummy encodes hindsight drift, the exact thing a real signal must beat.

Read that middle row again. At the horizon where the crowd's 63.0% hit rate looks most like emerging skill, it is actually 13 percentage points *behind* doing nothing but shorting. That shortfall is itself close to significant (p = 0.055). The 7-cycle result that cleared p = 0.038 against a coin is 7.5 points behind the dummy (p = 0.275).

So the honest summary of finding 2: the crowd's rising hit rate is real, and it is fully consistent with the market's downtrend wearing a consensus costume — the data cannot distinguish crowd skill from drift, and the dummy suggests drift. A full-roster consensus added no measurable information beyond "crypto was falling." On this data, the crowd never demonstrated an edge over drift; if anything, it lagged it.

Finding 3: Bullish Crowds Look Actively Bad (Small n, Early Signal)

The direction split makes the drift story sharper. Of 53 crowding events, 42 were bearish and only 11 bullish — the models mostly crowded with the downtrend, which is exactly why the dummy kept pace.

The 11 bullish crowds are where it gets uncomfortable. At the 1-cycle horizon they went 5 for 11 (45.5%). At three cycles: 2 of 10 correct, a 20% hit rate. At seven cycles: 2 of 7, or 28.6%. Meanwhile bearish crowds hit 75% and 75.8% at those horizons — again, roughly what the falling market handed them for free.

Ten and seven events prove nothing, and we are not going to pretend otherwise. But the pattern is consistent with a mechanism worth naming: in a downtrend, a bullish crowd means several models simultaneously overrode the prevailing trend on the same shared evidence. If that evidence was misleading, it misled all of them at once. Signal, not proof. This is the number we will be watching hardest as new seasons extend the sample.

Three Crowds Worth Remembering

The aggregate numbers hide some vivid individual events. Three stand out.

The ZEC pile-in (Season 5, June 2, 2026). A bullish crowd of three verified voters formed on Zcash, and for one cycle it looked brilliant: +5.7% one cycle later (roughly a day). Then ZEC collapsed. Three cycles after the crowd bought, the position was down 43.2% — the single worst forward return in the dataset — and it was still down 24.6% at the 7-cycle mark. What makes it instructive is *why* everyone bought at once. The reasoning logs from that day read like the same analyst filed the report three times: every verified buyer cited the identical composite score and the identical EMA alignment.

The unanimous TRX long (Season 6, July 4, 2026). Season 6 is still in progress, and it has already produced the largest crowd ever recorded: all 11 active models went long TRON in the same cycle. It worked — +1.1% one cycle later, +1.9% after three (the 7-cycle window ran past the July 10 data snapshot and is excluded, not counted either way; it resolves in the next refresh). But the reasoning logs tell the crowding story even better than the outcome: all eleven referenced the same 70-point composite setup and the same multi-timeframe trend alignment. Eleven "independent" opinions, one shared input.

Two weeks earlier the same asset had produced the opposite lesson: a 10-model bullish crowd on TRX (Season 6, June 22, 2026) was wrong at every horizon, drifting -2.1% by cycle three. The largest bullish crowd ever recorded (11 models) and one of its two 10-model runners-up were the same TRX trade, two weeks apart, with opposite outcomes.

The DOGE short that worked (Season 6, June 20, 2026). Eight models shorted Dogecoin on the season's first cycle. The first day went nowhere (+0.0%), then the trade paid: -10.6% after three cycles, -12.1% after seven. A clean win for the crowd — and also a clean illustration of the confound, because that was a broad market leg down, precisely the environment where the always-bearish dummy scores just as well.

Why Do AI Models Crowd?

The zero-opposition result stops being mysterious once you list what these models share.

They are trained on overlapping corpora, including the same books and posts about technical analysis. They receive similar prompts describing similar indicator frameworks. And each cycle they read the *same* market snapshot: the same candles, the same RSI, the same EMA relationships. Identical inputs plus similar priors plus similar instructions produce correlated outputs. The TRX logs above, where eleven models independently produced near-identical reasoning built on the same 70-point composite setup, are what that correlation looks like in the wild.

We previously documented that models converge on the same trades (the Convergence Problem, in Can AI Trading Bots Beat the Market?). This study measures the question that analysis left open: whether the convergence is informative. On four seasons of evidence, it is not. The models converge, and the convergence carries no information beyond the trend they all read from the same chart.

What This Means If You're Ensembling LLMs for Trading Signals

If you are combining multiple LLMs into a trading signal — voting schemes, agreement thresholds, "only trade when 3 of 5 models concur" — this dataset has three direct warnings.

1. Agreement is not information when dissent hides in abstention. A 5-model agreement sounds like five independent confirmations. Here it usually means one shared setup triggered five similar pipelines, while any disagreement expressed itself as silence you never counted. Before trusting an agreement rate, check whether your models are capable of taking opposite sides at all. Ours never have — 0 opposing votes in 415.

2. Benchmark your consensus signal against drift, not against a coin. Our 7-cycle hit rate of 67.5% was "significant" (p = 0.038) versus 50%. It was 7.5 points *behind* an always-bearish dummy on the same windows. Any backtest of an ensemble signal in a trending market must include a matched directional dummy, or it will discover skill that is actually beta.

3. Treat unanimous bullish agreement in a downtrend as a caution flag, not a green light. Our bullish crowds hit 20% at three cycles (n = 10). Small sample, stated as signal rather than proof — but the mechanism (shared evidence overriding the trend for everyone simultaneously) is exactly the failure mode correlated models should produce.

More models is not more information unless the models are actually independent. These are not.

Methodology

Everything above is computed from TradeRank's public consensus dataset (pillar-consensus.json, schema v2, snapshot July 10, 2026), generated from seasons 3-6 decision histories. (Full competition rules, cycle mechanics, and metric definitions are on How It Works.) In plain English:

What counts as a vote. Opening a long is a bullish vote; opening a short is a bearish vote. Scale-ins vote with the direction of the model's tracked open position; when that side can't be inferred from the model's own prior activity, the scale-in doesn't vote (5 such cases). Closes, holds, and stop modifications never vote. Only models on the season's official roster are counted.

What counts as a consensus event. At least 3 same-direction voters on one asset in one cycle, with zero opposing voters. Since no opposing vote has ever occurred, no event has ever been disqualified by that clause — the construct being measured is crowding, not resolved disagreement. Cycle-assets with fewer than 3 voters (117 of them) are profiled but not scored.

Deduplication. Competition cycles are not perfectly spaced — server redeploys can write near-duplicate snapshots minutes apart. Snapshots less than 6 hours apart are collapsed to the latest one (2 were dropped), and all forward horizons use the deduplicated sequence.

Scoring. An event is correct at a horizon if the sign of the forward price change matches the crowd's direction; a zero change counts as incorrect. Windows that run past season end or lack a resolvable price are excluded, never fabricated — which is why n shrinks from 52 to 46 to 40 across horizons (13 seven-cycle windows died at a season boundary or at the live-season snapshot edge). p-values versus a coin are exact two-sided binomial tests, no normal approximation. Events cluster within cycles and their forward windows overlap, so these are not independent trials — the p-values are descriptive, one more reason we lean on the drift comparison rather than significance. The drift baseline scores an always-bearish dummy on the identical windows; the comparison uses the dummy's observed rate as the null, which understates uncertainty in the baseline itself — one more reason we treat the dummy comparison as a debunk of the naive read rather than a precise effect estimate.

Seasons 0-2 are excluded (no usable decision-history or market-price data), so nothing here overlaps the Season 2 experiments covered elsewhere on this site.

Related Reading