# Consensus Watch #2

**WEEKLY · CONSENSUS** · August 19, 2026 · MarketMania Research · Weekly series

Window: **Aug 10-16, 2026 UTC** · Cutoff: **Mon Aug 17, 2026, 16:00 UTC** · Methodology **v1.1 (2026-08-10)**, hash `e66c7e8c864a2233`

PDF: https://marketmania.ai/research/reports/consensus-watch-2026-08-10.pdf · Open data (JSON): https://marketmania.ai/research/reports/consensus-watch-2026-08-10.json

> Research question: *"if a model's forecast matches the consensus of the other models, is the forecast more reliable?"*

## Research Snapshot

| RESEARCH SNAPSHOT |  |  |  |
|---|---|---|---|
| **2,929** LOO-scoreable calls | **7** models (stable lineup) | **741** thin/tied excluded | window **Aug 10-16, 2026 UTC** |
| method = **leave-one-out** majority, 3+ directional peers | base field hit **43.8%** | cutoff **Mon Aug 17, 16:00 UTC** | methodology **v1.1**, hash **e66c7e8c864a2233** |
| market: BTC net **-3.08%** (prior +2.09%) | ann. vol **10.0%** (was 11.4%) | TOP5 volume **$7.75B**, -8.9% w/w | pairwise corr **0.52** (was 0.40) |

## Key finding

> **KEY FINDING · OBSERVATION (one weekly window)**
>
> ### The margin curve inverted: the thinnest majorities (3 peers) were the only tier to beat the field, and near-unanimity stayed below base for a second week.

## TL;DR

- OBSERVATION -- Herding intensified, and still did not pay. Of 2,929 LOO-scoreable directional calls, 2,923 (99.8%) sided with peers' majority (issue #1: 99.4%); agree-hit was 43.1% vs. 50.0% for the 6 disagreers, a pooled lift of -6.9pp on n=6 with a Wilson 95% CI of [18.8%, 81.2%] that swallows the agree bucket's own CI [41.3%, 44.9%]. Descriptive, one week, not a claim of a durable edge.
- The margin curve flipped upside-down vs. issue #1: 3 peers 50.2% (was 30.5%), 4 peers 39.4% (was 43.3%), 5 peers 41.0% (was 49.9%), full unanimity 43.3% (was 33.6%). Only the thinnest majority beat the 43.8% base; two weeks, two opposite shapes.
- Daily-horizon herds were the week's disaster: 1d/1d margin-5/6 hit 19.4%/17.1%; 1d/4h near-unanimous went 0 for 28.
- Herd failures were real: 299 cells with 6+ on one side; the 5 worst -- all 7 models long at ~70 mean confidence -- all resolved 0%, in a week where 'sideways' was 60.6% of mature forecasts. All five landed on Aug 11-12.

## Pooled result: agree vs. disagree with the LOO consensus

| Group | n | Hit rate | Wilson 95% CI | Lift, pp |
|---|---|---|---|---|
| Agree with LOO peers | 2,923 | 43.1% | [41.3%, 44.9%] | -6.9 (vs. disagree) |
| Disagree with LOO peers | 6 | 50.0% | [18.8%, 81.2%] | ref. |

Read with caution: lift_pp rests on disagree n=6; Wilson 95% CI [18.8%, 81.2%] swallows the agree bucket's CI [41.3%, 44.9%]; descriptive, one week only, not a durability claim.

## Hit rate by leave-one-out margin (week over week)

| Margin | Issue #1 (Aug 3-9) | Issue #2 (Aug 10-16) | Issue #2 n |
|---|---|---|---|
| 3 peers | 30.5% | 50.2% | 428 |
| 4 peers | 43.3% | 39.4% | 525 |
| 5 peers | 49.9% | 41.0% | 792 |
| 6 peers (unanimity) | 33.6% | 43.3% | 1,169 |

A 2-peer bucket exists (n=9, 66.7%) — footnote only. Field base this week: 43.8%.

## By forecast horizon / timeframe

| FH / TF | Agree n | Agree hit | Dis. n | Dis. hit | Excluded |
|---|---|---|---|---|---|
| 1h / 1h | 1,865 | 45.0% | 6 | 50.0% | 475 |
| 4h / 4h | 435 | 40.9% | 0 | n/a | 142 |
| 4h / 1h | 463 | 40.8% | 0 | n/a | 90 |
| 1d / 1d | 88 | 34.1% | 0 | n/a | 11 |
| 1d / 4h | 72 | 30.6% | 0 | n/a | 23 |

"Excluded" = thin (fewer than 3 directional peers) or tied peer splits; not folded into agree or disagree. All 6 disagreements sat in the 1h/1h cell — the second week running.

## Herd failures: the 5 worst unanimous misses

| Slot (UTC) | Symbol | FH / TF | Side | Models | Mean conf | Hit rate |
|---|---|---|---|---|---|---|
| Aug 11, 12:01 | BNB | 4h / 1h | long | 7 | 71.0 | 0.0% |
| Aug 12, 12:01 | SOL | 1h / 1h | long | 7 | 69.0 | 0.0% |
| Aug 11, 13:01 | BNB | 1h / 1h | long | 7 | 68.1 | 0.0% |
| Aug 11, 22:01 | SOL | 1h / 1h | long | 7 | 68.0 | 0.0% |
| Aug 12, 12:01 | ETH | 4h / 4h | long | 7 | 67.9 | 0.0% |

All 5 worst herds were long calls, in a week where 'sideways' was 60.6% of mature forecasts; all five landed inside Aug 11-12.

## Per-model: agree vs. disagree

| Model | Agree n | Agree hit | Disagree n | Disagree hit |
|---|---|---|---|---|
| claude-fable-5 | 401 | 41.9% | 0 | n/a |
| claude-opus-5 | 334 | 41.3% | 0 | n/a |
| deepseek-v4-pro | 371 | 42.0% | 2 | 0.0% |
| gemini-3.1-pro | 406 | 44.6% | 4 | 75.0% |
| gpt-5.6-sol | 477 | 40.9% | 0 | n/a |
| grok-4.5 | 482 | 45.4% | 0 | n/a |
| qwen-3.8-max | 452 | 44.7% | 0 | n/a |

## Practical implications

- Last week unanimity behaved like a crowding signal; this week it merely matched base while the whole tier ordering flipped. The practical read is not 'fade the herd' -- it is that NO consensus tier has yet shown a stable meaning, so a consensus filter should not carry weight in a config on two weeks of evidence.
- Treat any agree-vs-disagree lift as regime-dependent description, not strategy: this week's -6.9pp rests on 6 disagreeing calls whose CI swallows the agree bucket's entirely.
- Consensus Watch is a living series: each issue adds one more week to the LOO comparison and one more point on the margin curve. Two points cannot separate shape from noise; several issues can.

## Limitations

- Single week (Aug 10-16, 2026 UTC). Every finding is descriptive for this window only.
- Observations are not independent: overlapping symbol/FH/TF cells hour after hour; Wilson intervals are descriptive, not inferential.
- Correlational, not causal: shared inputs mean shared blind spots; LOO removes self-match bias, not this confound.
- The pooled lift rests on 6 disagreeing calls, all in FH 1h; per-model disagree counts run 0-4.
- 741 calls (20.2% of the 3,670 mature directional universe) were thin or tied and are excluded from every consensus table.

## Counters & lineage

| Counter | Value |
|---|---|
| Forecasts total | 9,380 |
| OK in gate | 9,307 |
| Out of gate (1w) | 69 |
| Invalid | 4 |
| Mature (scored pool) | 9,307 |
| — of them directional | 3,670 |
| Pending (next issue) | 0 |
| Late closes | 0 |
| Source file | weekly_metrics_2026-08-10.json |
| Methodology | v1.1 (2026-08-10), hash e66c7e8c864a2233 |
| Report cutoff | Mon Aug 17, 2026, 16:00 UTC |

**THIS REPORT:** 2,929 LOO-scored observations. **MARKETMANIA RESEARCH TO DATE** (as of Aug 17, 2026 cutoff): 23,458 resolved forecasts since Jul 11 · 9 models tracked (7 current frontier + 2 archived legacy) · 5 assets · 5 forecast horizons · hourly cadence · 9 published research reports incl. this wave.

> Issue #2. Consensus Watch is a living weekly comparison; each issue appends one more week of leave-one-out agreement-vs-hit data. Issue #1's open question -- does full unanimity keep underperforming? -- closed as 'yes, but barely'; the sharper result is the inverted margin curve. Engine 1.1 powers the sandbox since Aug 18 (after this window closed); no cross-engine PnL comparisons are claimed.

## What we're testing next

- Next issue: does the margin-curve inversion persist -- or is tier ordering week-to-week noise? Three points make a trend.
- Monthly test (Sep 2): the agreement curve across regimes at monthly n.

## Related research

- Weekly Calibration #2 — https://marketmania.ai/research/reports/weekly-calibration-2026-08-10.pdf
- Weekly Model Watch #2 — https://marketmania.ai/research/reports/model-watch-2026-08-10.pdf
- Config Watch #2 — https://marketmania.ai/research/reports/config-watch-2026-08-12.pdf
- Consensus Watch #1 — https://marketmania.ai/research/reports/consensus-watch-2026-08-03.pdf

The three weekly reports publish together as one issue each week; Config Watch follows on its own cycle. Direct links are the posting rule from this wave on.

## Cite this report

```
@misc{mm_consensus_watch_2026w33,
  title  = {Consensus Watch #2: does agreeing with the crowd make an LLM's market call safer?},
  author = {{MarketMania Research}},
  year   = {2026}, month = {August}, day = {19},
  url    = {https://marketmania.ai/research/reports/consensus-watch-2026-08-10.pdf},
  note   = {Methodology v1.1; window Aug 10-16, 2026 UTC; source weekly_metrics_2026-08-10.json}
}
```