# Weekly Model Watch #7

**WEEKLY · MODEL WATCH** · September 22, 2026 · MarketMania Research · Weekly series

Window: **Sep 14-20, 2026 UTC** · Cutoff: **Mon Sep 21, 2026, 16:00 UTC** · Methodology **v1.1 (2026-08-10)**, hash `e66c7e8c864a2233`

PDF: https://marketmania.ai/research/reports/model-watch-2026-09-14.pdf · Open data (JSON): https://marketmania.ai/research/reports/model-watch-2026-09-14.json

> Research question: *"who led the week, which tickers were readable, who saw the turn first, and does a model agreeing with itself across timeframes/horizons make its call stronger?"*

## Research Snapshot

| RESEARCH SNAPSHOT |  |  |  |
|---|---|---|---|
| **5,598** directional calls scored | **7** models (stable lineup) | of **9,308** mature forecasts | window **Sep 14-20, 2026 UTC** |
| hit = direction rule: tp1/tp2->hit, sl->miss, expiry->sign of gross pnl | FH gate **1h / 4h / 1d** | base field hit **50.7%** | cutoff **Mon Sep 21, 2026, 16:00 UTC** |
| market: BTC net **+5.64%** (prior -4.36%) | ann vol **51.0%** (was 19.7%) | TOP5 volume **$18.74B**, +17% w/w | pairwise corr **0.93** (was 0.74) |

## Key finding

> **KEY FINDING · OBSERVATION (one weekly window)**
>
> ### Field accuracy went 39.6% -> 50.7% (+11.2pp) with 7 of 7 models improving, and no weekly title was awarded, for the seventh issue running.

## TL;DR

- **OBSERVATION -- No weekly title again.** gemini-3.1-pro leads descriptively at 52.9% [49.6%, 56.2%]; the gap to #2 (claude-fable-5, 52.3%) is 0.6pp. 7 of 7 models improved week over week; issue #6's leader gemini-3.1-pro is now 1 of 7.
- **1 of 5 tickers were readable.** ETH is the week's most readable coin (57.2%, n=1,116) and BNB the hardest at 46.3%; 1 of the 5 assets cleared 51%. Pair of the week: gemini-3.1-pro x ETH 57.9% (n=164); worst: qwen-3.8-max x BNB 43.6% (n=149).
- **1 of 1 qualifying turns found a caller.** ETH long +2.22%, 3 caller(s), first gemini-3.1-pro.
- **Self-agreement read as a lift in every cell.** Cross-TF/FH lifts ran +1.4pp to +6.5pp (issue #6: -6.0pp to +0.3pp). On trend days, 4h TF-agreement hit 62.9% (n=353) against 41.4% (n=304) on flat days.

### Week-over-week chart (see PDF for the grouped bar chart)

Directional hit-rate (%) by model, week over week, ranked by this week's rate. Short names: opus-5 = claude-opus-5, fable-5 = claude-fable-5, gemini-3.1 = gemini-3.1-pro, gpt-5.6 = gpt-5.6-sol, qwen-3.8 = qwen-3.8-max, deepseek = deepseek-v4-pro. Field base 39.6% -> 50.7%. The grok column compares two pure grok-4.6 weeks: the 4.5 -> 4.6 flip sits three windows back. "Was" values are the numbers published in issue #6; the leaderboard's W/w column is computed on unrounded rates and can differ by 0.1pp from the difference of the rounded columns.

## Why it matters

MarketMania scores 7 model lines against the same market, hour after hour. Weekly Model Watch asks four practical questions about that week of calls: who led the week, which tickers were readable, who saw the turn first, and does a model agreeing with itself across timeframes/horizons make its call stronger? Each question gets its own block below, with its own table, its own N, and its own honesty caveats. The grok row is a pure grok-4.6 week in this issue and was one in issue #6 as well -- the 4.5 -> 4.6 flip sits three windows back -- so read the grok week-over-week row as a line, not a single model build.

## How to read this

Every block below is scored on the same pool: matured directional calls inside the 1h/4h/1d FH gate, hit by the methodology v1.1 direction rule. The weekly title is awarded only on a >=5pp gap AND non-overlapping 95% Wilson CIs with N>=10 per side -- a descriptive lead is not a title. Reversals use the 4h grid with a >=2.0% counter-move sustained 8h, BTC/ETH only in v1. Cross-TF/FH blocks ask whether a model agreeing with itself does any better, and are split trend vs flat wherever the week supplies both sides. The grok 4.5 -> 4.6 flip (Aug 24, 2026 09:22 UTC) sits three windows back (issue #4): both weeks of this issue are pure grok-4.6, with no 4.5-era rows in either window, so the grok week-over-week row is the second pure-4.6 against pure-4.6 comparison of the series.

## Leaderboard of the week

| Model | N | Cov. | Hit rate | 95% CI | W/w pp |
|---|---|---|---|---|---|
| gemini-3.1-pro | 883 | 66.4% | 52.9% | 49.6%-56.2% | +11.1 |
| claude-fable-5 | 799 | 60.1% | 52.3% | 48.9%-55.8% | +15.8 |
| claude-opus-5 | 711 | 53.5% | 51.6% | 47.9%-55.3% | +15.1 |
| qwen-3.8-max | 819 | 61.6% | 50.8% | 47.4%-54.2% | +10.8 |
| gpt-5.6-sol | 821 | 61.7% | 49.7% | 46.3%-53.1% | +9.6 |
| grok-4.6 | 629 | 47.3% | 49.6% | 45.7%-53.5% | +10.0 |
| deepseek-v4-pro | 936 | 70.4% | 48.4% | 45.2%-51.6% | +7.0 |

No weekly title is awarded this issue -- for the seventh time. Title rule: **>=5pp gap AND non-overlapping 95% CI, N>=10** per side. gemini-3.1-pro's lead over claude-fable-5 is 0.6pp and their CIs overlap, so neither condition is met. The grok 4.5 -> 4.6 flip (Aug 24, 2026 09:22 UTC) sits three windows back (issue #4): both weeks of this issue are pure grok-4.6, with no 4.5-era rows in either window, so the grok week-over-week row is the second pure-4.6 against pure-4.6 comparison of the series.

## Ticker of the week

| Symbol | N | Hit rate | 95% CI | Issue #6 |
|---|---|---|---|---|
| ETH | 1,116 | 57.2% | 54.2%-60.0% | 29.5% |
| XRP | 1,108 | 50.4% | 47.4%-53.3% | 42.2% |
| BTC | 1,163 | 50.3% | 47.4%-53.2% | 41.4% |
| SOL | 1,156 | 49.5% | 46.6%-52.4% | 36.0% |
| BNB | 1,055 | 46.3% | 43.3%-49.3% | 46.9% |

ETH was the field's most readable ticker (57.2%) and BNB the hardest (46.3%); 1 of 5 assets cleared 51% this week (issue #6: 0 of 5). Issue #6's best ticker was BNB and its hardest ETH -- this issue's ordering is ETH > XRP > BTC > SOL > BNB.

## Pair of the week

Top 6 of the 12 model x ticker pairs tracked this week (best hit-rate), then the 5 worst in their own table -- the pair list is split rather than paged so neither table breaks across a page. Pairs shown require N of 8 or more calls.

| Model (top 6) | Symbol | N | Hit rate |
|---|---|---|---|
| gemini-3.1-pro | ETH | 164 | 57.9% |
| claude-fable-5 | ETH | 152 | 57.9% |
| claude-opus-5 | ETH | 147 | 57.8% |
| grok-4.6 | ETH | 135 | 57.8% |
| qwen-3.8-max | ETH | 160 | 56.9% |
| gpt-5.6-sol | ETH | 167 | 56.3% |

| Model (5 worst) | Symbol | N | Hit rate |
|---|---|---|---|
| gpt-5.6-sol | BNB | 152 | 45.4% |
| deepseek-v4-pro | SOL | 192 | 45.3% |
| grok-4.6 | SOL | 129 | 45.0% |
| deepseek-v4-pro | BNB | 171 | 43.9% |
| qwen-3.8-max | BNB | 149 | 43.6% |

gemini-3.1-pro x ETH (57.9%, n=164) is the week's best pair and qwen-3.8-max x BNB (43.6%, n=149) the weakest. Pair history accumulates across issues -- read this as the seventh data point, not a ranking.

## Who saw the reversal first

Reversal definition: 4h grid; trend = sign of prior 24h; counter-move of 2.0% or more (BTC/ETH) sustained 8h. A model 'calls' the reversal if it has a matured directional hit call in the new direction, FH 4h or 1d, in the slot window [T-12h, T+FH].

| Symbol | Time (UTC) | New side | Move (8h) | Callers | First caller |
|---|---|---|---|---|---|
| ETH | Sun Sep 20, 12:00 | Long | +2.22% | 3 | gemini-3.1-pro |

The ETH long turn drew 3 caller(s), first gemini-3.1-pro on a 1d/1d call at Sep 20, 00:01 with confidence 75. Scope in v1 remains BTC/ETH only; the event count across seven issues is still far too small to characterise turn-detection skill either way.

## Cross-TF and cross-FH confirmation: does agreeing with yourself help?

| Block | FH | vs. | Agree n | Agree hit | Dis n | Dis hit | Base hit | Lift |
|---|---|---|---|---|---|---|---|---|
| cross-TF | 4h | TF 4h vs 1h | 657 | 53.0% | 27 | 48.1% | 51.6% (898) | +1.4 |
| cross-TF | 1d | TF 1d vs 4h | 133 | 39.1% | 3 * | 100.0% | 36.8% (166) | +2.4 |
| cross-FH | 4h | by FH 1h | 584 | 53.1% | 36 | 44.4% | 51.6% (898) | +1.5 |
| cross-FH | 1d | by FH 4h | 104 | 43.3% | 7 * | 57.1% | 36.8% (166) | +6.5 |

* = N<10 (insufficient); insufficient cells are never used to rank models. Mixed pairs (a model with only one side of the comparison present): 214 / 30 / 278 / 55. Pooled across all four cell families the lifts run +1.4pp to +6.5pp this week (issue #6: -6.0pp to +0.3pp).

**Regime control.** Methodology calls for a trend-vs-flat split on this table. This window has 3 trend days of 7: 09-14 +1.79% (trend), 09-15 -3.35% (trend), 09-16 +0.86%, 09-17 +0.21%, 09-18 +5.82% (trend), 09-19 +0.46%, 09-20 +0.13%. On trend days, 4h TF-agreement hit 62.9% (n=353) against 41.4% (n=304) on flat days.

## Side-mix and the week's failure

| Model | N | Long | Short | Sideways | L:S |
|---|---|---|---|---|---|
| claude-fable-5 | 1,330 | 44.0% | 16.1% | 39.9% | 2.73 |
| claude-opus-5 | 1,330 | 40.0% | 13.5% | 46.5% | 2.97 |
| deepseek-v4-pro | 1,330 | 43.1% | 27.3% | 29.6% | 1.58 |
| gemini-3.1-pro | 1,330 | 41.3% | 25.1% | 33.6% | 1.64 |
| gpt-5.6-sol | 1,330 | 34.7% | 27.1% | 38.3% | 1.28 |
| grok-4.6 | 1,329 | 26.3% | 21.1% | 52.7% | 1.25 |
| qwen-3.8-max | 1,329 | 40.7% | 20.9% | 38.4% | 1.95 |

L:S = each model's own long-share divided by its short-share; the field average of the per-model ratios is 1.91 this week. 'Sideways' was 39.9% of mature forecasts (3,710 of 9,308). Consensus skew: 685 symbol/FH/TF cells had 5 or more models on the same side and hit 50.5% (n=4,356) against the 50.7% field base.

| Model | Symbol | Side | Conf | FH / TF | Slot (UTC) | Exit | Net PnL |
|---|---|---|---|---|---|---|---|
| gpt-5.6-sol | BNB | Short | 84 | 1h / 1h | Sep 16, 09:01 | EXPIRY | -$0.67 |

**Fail of the week.** The week's single highest-confidence individual miss, exit reason expiry. Listed as a single card, never as a model ranking.

## Calibration bridge

The hit-rate leader this week is gemini-3.1-pro (52.9%); the best-calibrated line is claude-fable-5 (Brier 0.2593, gap +9.1pp). Per-model ok-rate this window ran 99.46% to 100.00%. The full confidence-bucket analysis lives in Weekly Calibration #7.

## Market check: an up week on a louder tape

> **MARKET CHECK · OBSERVATION (two adjacent calendar weeks)**
>
> OBSERVATION — mean |1d move| 1.35% -> 2.69% (+100% rel), BTC realized vol 31.4% -> 35.1% (ann., hourly); field directional accuracy 39.6% -> 50.7% (+11.2pp), 7 of 7 models improved, sim win-rate up for 7 of 7, field sim PnL -$1018.93 -> -$337.72 (adjacent calendar weeks Sep 7-13 vs Sep 14-20; descriptive, one pair of weeks, not a claim).
>
> Robustness: raw price-sign accuracy 38.7% -> 54.2% — the same direction as the trade-based hit rule.

| Measure | Sep 7-13 (week A) | Sep 14-20 (week B) | Change |
|---|---|---|---|
| Mean \|1d move\| (5 assets) | 1.35% | 2.69% | +100% rel |
| BTC realized vol (ann., hourly) | 31.4% | 35.1% | +3.7pp |
| Field directional accuracy | 39.6% | 50.7% | +11.2pp |
| Models improving hit-rate | -- | 7 of 7 | -- |
| Raw price-sign accuracy | 38.7% | 54.2% | +15.4pp |
| Field sim win-rate | 32.1% | 43.4% | +11.2pp |
| Field sim net PnL | -$1018.93 | -$337.72 | -- |

Week A is exactly the issue-#6 window, so this slice and every week-over-week column in this report compare the same two back-to-back weeks. Descriptive, one pair of weeks; the hit rule is methodology v1.1 (trade-based) and the raw price-sign row is the robustness check (hourly closes at :00 against slots at :01). BTC realized vol here is the hourly-return series from the alive slice; the snapshot row's figure is the audited daily-candle estimate -- different estimators, both reported as measured. Lineage splice ACTIVE: grok-4.5 rows are aggregated into grok-4.6 (flip 2026-08-24T09:22:00+00:00, three windows back, inside issue #4's window). Week A (2026-09-07..2026-09-13) is pure grok-4.6 (it reproduces the published issue #6); week B is pure grok-4.6.

## Practical implications

- Do not read this week's leaderboard as a skill ranking: no title was awarded and the top CIs overlap.
- Treat per-ticker readability as weekly weather. Seven issues, and the best ticker changed hands in every one except issue #6 (SOL -> BNB -> BTC -> SOL -> BNB -> BNB -> ETH): ETH leads this week at 57.2%, after two issues of BNB.
- Self-agreement lifts run +1.4pp to +6.5pp with small disagree cells; carry the regime split into the monthly test, do not act on it.

## Limitations

- Single week (Sep 14-20, 2026 UTC). Descriptive for this window only; no claim about next week or any model's underlying skill. This window ran 3 trend days of 7 (issue #6: 2 of 7); every comparison with issue #6 is a comparison across regimes as well as across weeks.
- Observations inside one window are not independent, and the 95% Wilson CIs shown throughout are descriptive, not inferential.
- The reversal detector remains v1 and BTC/ETH-only, on price series reconstructed from trade entry prices. 13 events across seven issues cannot characterise detector performance either way.
- Cross-TF and cross-FH disagreement cells are small (n=27, 3, 36, 7); cells under N=10 are insufficient and never used to rank models. Week-over-week deltas are computed on unrounded rates and may differ by 0.1pp from the difference of the rounded columns.
- Model lines in this window: claude-fable-5, claude-opus-5, deepseek-v4-pro, gemini-3.1-pro, gpt-5.6-sol, grok-4.6, qwen-3.8-max. DeepSeek announced the retirement of V4 Pro in favour of V4.1 Flash from Sep 14 (Beijing noon); as of Sep 22 the API still returns deepseek-v4-pro as the served model for our requests and the per-answer latency, token and cost series of Sep 1-22 show no regime change, so this issue keeps the label; it will change the day the served id changes. Series density and lineage notes (wave 7): (1) rolling-1w forecasts emit daily since Aug 22, 2026 and rolling-1M daily since Aug 23, 2026, so inside this window (Sep 14-20) daily coverage is FULL, as it was in issues #4, #5 and #6: the 1w series has slots on 7 of 7 days and the 1M series on 7 of 7 days; both sit outside this report's FH gate (1h/4h/1d) and appear only in the exclusions counter (1w 488, 1M 485). (2) The grok line flipped 4.5 -> 4.6 at Aug 24, 2026 09:22 UTC, three windows back (inside the issue-#4 window): this window carries 1,470 grok rows and none from the 4.5 era, and neither does week A (the issue-#6 window), so every grok number in this issue and in issue #6 is a pure grok-4.6 line -- the grok week-over-week row is the second pure-4.6 against pure-4.6 comparison of the series. (3) qwen-3.8-max succeeded qwen-3.7-max on Aug 5, 2026 (fully before this window).
- Market-state row is computed on a single-exchange (binance) daily candle series, as in issue #6 after the audit that found the raw table mixes two exchanges; issue #6 values are as published.
- Research-to-date counter: directional forecasts resolved inside the FH gate (1h/4h/1d) since Jul 11, counted at the issue cutoff (Sep 21 16:00 UTC) -- the definition pinned in issue #3 and carried by issues #4, #5 and #6, which printed 47,373 at the Sep 14 16:00 cutoff. The pack reproduces that pin at the previous cutoff (control OK), so this issue's 52,971 is an additive step under one definition; counter deltas against issue #5 and earlier remain definitional.

## Counters & lineage

| Counter | Value | Counter | Value |
|---|---|---|---|
| Forecasts total | 10,290 | Mature (scored pool) | 9,308 |
| OK in gate | 9,308 | -- of them directional | 5,598 |
| Out of gate (1w / 1M) | 488 / 485 | -- of them sideways | 3,710 |
| Invalid | 9 | Pending (next issue) | 0 |
| Late closes | 0 | Uptime, all grids | 266 / 266 slots |
| Source file | weekly_metrics_2026-09-14.json | Report cutoff | Mon Sep 21, 2026, 16:00 UTC |
| Generated at (pipeline) | 2026-09-22 06:56 UTC | Methodology | v1.1 (2026-08-10) · hash e66c7e8c864a2233 |

**THIS REPORT: 5,598 scored observations.** MARKETMANIA RESEARCH TO DATE (as of Sep 21, 2026 cutoff): 52,971 directional forecasts resolved since Jul 11 · 7 models tracked (current line-up; earlier versions folded into their successors' lineage) · 5 assets · 5 horizons · hourly · 33 published reports.

> **Issue #7.** Weekly Model Watch is a living series. Issue #6's open questions -- does gemini-3.1-pro hold the top rank for a third issue and does the title rule ever fire, and does a qualifying reversal ever fire outside ETH or a second turn in the same pair draw a caller? -- read this issue as: gemini-3.1-pro at the top (52.9%), the line that also led issue #6 (gemini-3.1-pro), so the rank held for a third issue; no weekly title was awarded, for the seventh issue running, and 1 of 1 qualifying turns found a caller, on turns in ETH only. The grok row is pure grok-4.6 this issue and was pure grok-4.6 in issue #6; the 4.5 -> 4.6 flip (Aug 24, 2026 09:22 UTC) sits three windows back. Engine 1.1 has powered the sandbox since Aug 18, i.e. before this window; no cross-engine PnL comparisons are claimed.

## What we're testing next

- Next issue: does gemini-3.1-pro hold the top rank for a fourth issue, and does the title rule ever fire?
- Does a qualifying reversal ever fire outside ETH, and does a second turn in the same pair ever draw a caller?
- Monthly series: cross-TF/FH lifts with regime control across mixed weeks (trend vs. flat), block bootstrap.

## Related research

| Report | Direct PDF link |
|---|---|
| Consensus Watch #7 | https://marketmania.ai/research/reports/consensus-watch-2026-09-14.pdf |
| Weekly Calibration #7 | https://marketmania.ai/research/reports/weekly-calibration-2026-09-14.pdf |
| Weekly Model Watch #6 | https://marketmania.ai/research/reports/model-watch-2026-09-07.pdf |

The three weekly reports publish together as one issue each week; each links straight to the others' PDF and to its own previous issue. Direct links are the posting rule from wave 2 on.

## Cite this report

```bibtex
@misc{mm_modelwatch_2026w38,
  title  = {Weekly Model Watch #7: weekly leaderboard, ticker/pair reads and cross-confirmation, Sep 14-20 2026},
  author = {{MarketMania Research}},
  year   = {2026}, month = {September}, day = {22},
  url    = {https://marketmania.ai/research/reports/model-watch-2026-09-14.pdf},
  note   = {Methodology v1.1 (2026-08-10), hash e66c7e8c864a2233; window Sep 14-20, 2026 UTC; source weekly_metrics_2026-09-14.json}
}
```