# Weekly Model Watch #3

**WEEKLY · MODEL WATCH** · August 26, 2026 · MarketMania Research · Weekly series

Window: **Aug 17-23, 2026 UTC** · Cutoff: **Mon Aug 24, 2026, 16:00 UTC** · Methodology **v1.1 (2026-08-10)**, hash `e66c7e8c864a2233`

PDF: https://marketmania.ai/research/reports/model-watch-2026-08-17.pdf · Open data (JSON): https://marketmania.ai/research/reports/model-watch-2026-08-17.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 |  |  |  |
|---|---|---|---|
| **5,130** directional calls scored | **7** models (stable lineup) | of **9,260** mature forecasts | window **Aug 17-23, 2026 UTC** |
| hit = direction rule: tp1/tp2->hit, sl->miss, expiry->sign of gross pnl | FH gate **1h / 4h / 1d** | base field hit **54.1%** | cutoff **Mon Aug 24, 2026, 16:00 UTC** |
| market: BTC net **+23.58%** (prior -3.08%) | ann vol **71.1%** (was 10.0%) | TOP5 volume **$25.97B**, +235% w/w | pairwise corr **0.81** (was 0.52) |

## Key finding

> **KEY FINDING · OBSERVATION (one weekly window)**
>
> ### The market moved and the models kept up: field accuracy 43.8% -> 54.1% (+10.3pp) with 7 of 7 models improving -- and after two blind weeks, both ETH reversals finally found callers.

## TL;DR

- **OBSERVATION -- No weekly title again.** claude-opus-5 leads descriptively at 59.1% [55.3%, 62.7%] after the field's biggest jump (+17.5pp w/w), but the gap to #2 (claude-fable-5, 56.5%) is 2.6pp with overlapping CIs -- the title rule (>=5pp gap AND non-overlapping 95% CI) was not met for the third issue running. All 7 models improved; issue #2's leader gemini-3.1-pro slipped 1st -> 3rd on a +7.1pp gain.
- **Every ticker was readable.** BTC is the week's most readable coin (57.5%, n=1,118) and XRP the hardest at 51.2% -- the first week in the series where every one of the 5 assets cleared 51%. Issue #2's hardest ticker, ETH, rose 35.6% -> 52.8%; SOL bounced 38.9% -> 54.3%. Pair of the week: claude-opus-5 x BNB 60.3% (n=126); worst: grok-4.5 x XRP 45.2% (n=135).
- **The reversal blind spot broke.** Both qualifying reversals were ETH: the Aug 22 turn short (-3.46% over 8h) drew 1 caller, first gemini-3.1-pro; the Aug 23 turn long (+2.14%) drew 3, first claude-fable-5. Issues #1 and #2 had two reversals and zero callers between them.
- **Self-agreement turned mildly positive -- and the regime split widened.** Cross-TF/FH lifts ran -0.6 to +4.6pp (issue #2: -18.5 to +0.2pp), with the 1d cross-TF cell going from the week's disaster (17.9%) to its best agree bucket (75.9%, n=133). On the 5 trend days, 4h TF-agreement hit 64.8% vs 47.5% on the 2 flat days (n=455 vs 120).

### 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. Field base 43.8% -> 54.1%. "Was" values are the numbers published in issue #2; 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 models 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. Issue #3 is the series' first live-tape week -- 5 trend days against issue #2's 1 -- so every week-over-week movement here is a movement across regimes as well as across weeks.

## 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.

| Model | Issue #2 | Issue #3 | 95% CI (issue #3) |
|---|---|---|---|
| claude-opus-5 | 41.5% | 59.1% | 55.3%-62.7% |
| claude-fable-5 | 41.7% | 56.5% | 53.0%-59.9% |
| gemini-3.1-pro | 47.0% | 54.1% | 50.7%-57.6% |
| gpt-5.6-sol | 44.2% | 53.6% | 49.9%-57.3% |
| qwen-3.8-max | 44.1% | 52.3% | 48.6%-55.9% |
| deepseek-v4-pro | 42.5% | 51.6% | 48.1%-55.1% |
| grok-4.5 | 44.0% | 51.4% | 47.7%-55.1% |

Field base this week: 54.1% (issue #2: 43.8%). Week-over-week columns compare back-to-back windows: Aug 10-16 (issue #2, and week A of this issue's alive slice) vs Aug 17-23 (this issue). "Was" values are the numbers published in issue #2; deltas are computed on unrounded rates.

## Leaderboard of the week

The full 7-model field, ranked by this week's directional hit-rate; W/w pp = movement vs issue #2.

| Model | N | Coverage | Hit rate | 95% CI | W/w pp |
|---|---|---|---|---|---|
| claude-opus-5 | 672 | 50.5% | 59.1% | 55.3%-62.7% | +17.5 |
| claude-fable-5 | 779 | 58.6% | 56.5% | 53.0%-59.9% | +14.7 |
| gemini-3.1-pro | 798 | 60.1% | 54.1% | 50.7%-57.6% | +7.1 |
| gpt-5.6-sol | 688 | 51.7% | 53.6% | 49.9%-57.3% | +9.4 |
| qwen-3.8-max | 706 | 55.1% | 52.3% | 48.6%-55.9% | +8.2 |
| deepseek-v4-pro | 777 | 58.4% | 51.6% | 48.1%-55.1% | +9.1 |
| grok-4.5 | 710 | 53.4% | 51.4% | 47.7%-55.1% | +7.4 |

No weekly title is awarded this issue — for the third time. Title rule: **>=5pp gap AND non-overlapping 95% CI, N>=10** per side. claude-opus-5's lead over claude-fable-5 is 2.6pp and their CIs overlap, so neither condition is met. Rank shuffle vs issue #2: claude-opus-5 7th -> 1st, claude-fable-5 6th -> 2nd, gemini-3.1-pro 1st -> 3rd. Both Claude models made the field's two biggest jumps from the bottom two ranks, in a week when the base rate moved +10.3pp under everyone.

## Ticker of the week

This week's 5-ticker field, ranked by directional hit-rate.

| Symbol | N | Hit rate | 95% CI | Issue #2 |
|---|---|---|---|---|
| BTC | 1,118 | 57.5% | 54.6%-60.4% | 44.0% |
| SOL | 1,036 | 54.3% | 51.3%-57.4% | 38.9% |
| BNB | 932 | 54.0% | 50.8%-57.2% | 49.3% |
| ETH | 1,074 | 52.8% | 49.8%-55.8% | 35.6% |
| XRP | 970 | 51.2% | 48.1%-54.4% | 48.5% |

BTC was the field's most readable ticker (57.5%) and XRP the hardest (51.2%); their CIs are separated, while every adjacent pair in the middle of the list overlaps its neighbour. The story is the floor, not the top: this is the first week where all 5 assets cleared 51%, and both previous extremes reversed — issue #2's best (BNB) is 3rd and issue #2's worst (ETH, hardest for two issues running) is 4th. Three issues in, no asset has held a rank.

## 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>=8 calls.

**Top 6 pairs**

| Model | Symbol | N | Hit rate |
|---|---|---|---|
| claude-opus-5 | BNB | 126 | 60.3% |
| grok-4.5 | BTC | 141 | 60.3% |
| gemini-3.1-pro | BTC | 175 | 59.4% |
| claude-opus-5 | ETH | 138 | 59.4% |
| claude-opus-5 | SOL | 137 | 59.1% |
| claude-opus-5 | XRP | 119 | 58.8% |

**5 worst pairs**

| Model | Symbol | N | Hit rate |
|---|---|---|---|
| qwen-3.8-max | ETH | 140 | 49.3% |
| grok-4.5 | ETH | 159 | 49.1% |
| deepseek-v4-pro | ETH | 168 | 48.2% |
| qwen-3.8-max | XRP | 144 | 47.9% |
| grok-4.5 | XRP | 135 | 45.2% |

claude-opus-5 x BNB (60.3%, n=126) is the week's best pair — by 0.04pp over grok-4.5 x BTC, which rounds to the same 60.3% — and claude-opus-5 takes 4 of the top 6. At the other end every one of the 5 worst pairs still cleared 45%, and 3 of the 5 are ETH pairings (issue #2: 4 of 5). Pair history accumulates across issues — read this as the third data point, not a ranking.

## Who saw the reversal first

Reversal definition: 4h grid; trend = sign of prior 24h; counter-move >=2.0% (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** | Sat Aug 22, 00:00 | Short | -3.46% | 1 | **gemini-3.1-pro** |
| **ETH** | Sun Aug 23, 08:00 | Long | +2.14% | 3 | **claude-fable-5** |

Two qualifying reversals, both ETH, and both found callers — the first non-zero rows in the series. The Aug 22 turn drew exactly one caller: gemini-3.1-pro, on a 4h/1h call placed at 00:01 with confidence 60. The Aug 23 turn drew three — claude-fable-5 (first, 1d/1d slot at 00:01, confidence 62), claude-opus-5 and deepseek-v4-pro. Two weeks of NOBODY became one week of 1 and 3 callers; that is four qualifying events across three issues, still far too few to characterise turn-detection skill either way. Scope in v1 remains BTC/ETH only.

## 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 | 575 | 61.2% | 41 | 41.5% | 60.6% (840) | +0.6 |
| cross-TF | 1d | TF 1d vs 4h | 133 | 75.9% | 5 \* | 40.0% | 71.4% (171) | +4.6 |
| cross-FH | 4h | by FH 1h | 508 | 60.0% | 42 | 47.6% | 60.6% (840) | -0.6 |
| cross-FH | 1d | by FH 4h | 100 | 74.0% | 11 | 72.7% | 71.4% (171) | +2.7 |

`\*` = N<10 (insufficient); insufficient cells are never used to rank models. Mixed pairs (a model with only one side of the comparison present): 223 / 33 / 288 / 60. The 1d cross-TF row is the mirror image of issue #2 — a 1d call confirmed by the model's own 4h-TF read hit 75.9%, 4.6pp above the 1d base, where the same cell hit 17.9% and -18.5pp last week. Pooled across all four cell families the lifts now run -0.6 to +4.6pp: no longer a penalty, not yet an edge.

**Regime control.** Methodology calls for a trend-vs-flat split on this table. Issue #1 had zero trend days, issue #2 exactly one; this window has five: 08-17 +2.55% (trend), 08-18 +0.21%, 08-19 +7.82% (trend), 08-20 +5.07% (trend), 08-21 +7.26% (trend), 08-22 -1.67% (trend), 08-23 +0.69%. On the 5 trend days, 4h TF-agreement hit **64.8%** (n=455) vs **47.5%** on the 2 flat days (n=120); the disagree cells ran 41.0% (n=39) vs 50.0% (n=2, insufficient). The 1d cell splits the other way (74.3% trend, n=109, vs 83.3% flat, n=24). Issue #2's single-trend-day hint (54.3% vs 38.3%) reappears on the 4h cell at a readable n — the first time the regime control has had two usable sides. Still one window; the monthly block-bootstrap test is where this belongs.

## Side-mix and the week's failure

| Model | N | Long | Short | Sideways | L:S |
|---|---|---|---|---|---|
| claude-fable-5 | 1,330 | 52.3% | 6.3% | 41.4% | 8.27 |
| claude-opus-5 | 1,330 | 45.9% | 4.6% | 49.5% | 10.01 |
| deepseek-v4-pro | 1,330 | 46.6% | 11.8% | 41.6% | 3.95 |
| gemini-3.1-pro | 1,329 | 45.8% | 14.3% | 40.0% | 3.20 |
| gpt-5.6-sol | 1,330 | 38.4% | 13.4% | 48.3% | 2.87 |
| grok-4.5 | 1,330 | 35.2% | 18.2% | 46.6% | 1.93 |
| qwen-3.8-max | 1,281 | 45.4% | 9.7% | 44.9% | 4.69 |

L:S = each model's own long-share divided by its short-share. The field flipped its skew back and then some: issue #1 ran ~2.2:1 long-over-short, issue #2 ~0.7:1 short-over-long, and this week the average of the 7 per-model ratios is ~5.0:1 long-over-short in a week of 5 trend days. claude-fable-5 carried the highest long share (52.3%) and claude-opus-5 the highest ratio (10.01); grok-4.5 is the most balanced (1.93) and the field's weakest hit-rate. 'Sideways' fell to 44.6% of mature forecasts (4,130 of 9,260) from 60.6%, which is why the directional pool grew from 3,670 to 5,130. Consensus skew: 593 symbol/FH/TF cells had 5 or more models on the same side and hit 54.4% (n=3,683) — just above the 54.1% base, the first issue where a 5-of-7+ skew did not underperform the field, though 0.4pp is well inside noise.

| Model | Symbol | Side | Conf | FH / TF | Slot (UTC) | Exit | Net PnL |
|---|---|---|---|---|---|---|---|
| **gemini-3.1-pro** | XRP | Long | 85 | 4h / 1h | Aug 21, 16:01 | SL | **-$3.10** |

**Fail of the week.** The week's single highest-confidence individual miss. Cross-reference: XRP was the week's hardest ticker (51.2%) and the asset with the largest mean |1d| move (6.65%, against BTC's 3.34%) — an 85-confidence directional call into the week's noisiest tape. Unlike issue #2's expiry-based fail, this one hit its stop, which is what the calibration v2 TP/SL work is aimed at.

## Calibration bridge

For the first time in the series the hit-rate leader and the best-calibrated model are the same: claude-opus-5 tops the leaderboard at 59.1% and posts the field's lowest Brier (0.2419) with a +2.2pp overconfidence gap. Issue #2's best-calibrated model, qwen-3.8-max, is 3rd by Brier (0.2540) and 5th by hit-rate. Per-model ok-rate this window was 100.0% for five models, 99.93% for gemini-3.1-pro and 96.20% for qwen-3.8-max. The full confidence-bucket analysis lives in Weekly Calibration #3.

## Market check: the tape woke up

> **MARKET CHECK · OBSERVATION (two adjacent calendar weeks)**
>
> OBSERVATION — the tape sped up and accuracy followed: mean |1d move| 0.73% -> 4.28% (483% rel), BTC realized vol 19.3% -> 58.7% (ann., hourly); field directional accuracy 43.8% -> 54.1% (+10.3pp), 7 of 7 models improved, sim win-rate up for 7 of 7, field sim PnL -$527 -> +$889 (adjacent calendar weeks Aug 10-16 vs Aug 17-23; descriptive, one pair of weeks, not a claim).
>
> Robustness: raw price-sign accuracy 40.1% -> 53.5% — consistent with the trade-based hit rule.

| Measure | Aug 10-16 (week A) | Aug 17-23 (week B) | Change |
|---|---|---|---|
| Mean \\|1d move\\| (5 assets) | 0.73% | 4.28% | +483% rel |
| BTC realized vol (ann., hourly) | 19.3% | 58.7% | +39.4pp |
| Field directional accuracy | 43.8% | 54.1% | +10.3pp |
| Models improving hit-rate | — | 7 of 7 | — |
| Raw price-sign accuracy | 40.1% | 53.5% | +13.4pp |
| Field sim win-rate | 29.7% | 46.3% | +16.6pp |
| Field sim net PnL | -$527.07 | +$889.37 | — |

Week A is exactly the issue-#2 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 71.1% is the audited daily-candle figure — different estimators, both reported as measured.

## Practical implications

- Do not read this week's leaderboard as a skill ranking: no title was awarded, the top CIs overlap, and the two biggest jumps (+17.5pp, +14.7pp) came from the two models that were last and second-to-last a week earlier.
- Treat per-ticker readability as weekly weather. Three issues, three different best tickers (SOL -> BNB -> BTC) and the two-week 'ETH is hardest' pattern broke.
- Self-agreement stopped being a penalty but is not yet a booster: pooled lifts -0.6 to +4.6pp with overlapping CIs, and the reversal result is two events. Carry the regime split into the monthly test, do not act on it.

## Limitations

- Single week (Aug 17-23, 2026 UTC). Descriptive for this window only; no claim about next week or any model's underlying skill. This window's regime is the opposite of issue #2's: 5 trend days of 7 against 1 of 7.
- 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. Four events across three issues cannot characterise detector performance either way.
- Cross-TF and cross-FH disagreement cells are still small (n=5, 11, 41, 42); the n=5 cell is 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.
- qwen-3.8-max ran a 96.20% ok-rate this window; its mature pool is 1,281 against 1,330 for the other six models. Models in this window: claude-fable-5, claude-opus-5, deepseek-v4-pro, gemini-3.1-pro, gpt-5.6-sol, grok-4.5, qwen-3.8-max. Series density and lineage notes (wave 3): (1) rolling-1w forecasts emit daily since Aug 22, 2026 and rolling-1M daily since Aug 23, 2026, so inside this window (Aug 17-23) daily coverage is PARTIAL by design: the 1w series has the Mon Aug 17 anchor plus daily slots on Aug 22-23 only, and the 1M series has a daily slot on Aug 23 only; both sit outside this report's FH gate (1h/4h/1d) and appear only in the exclusions counter (1w 209, 1M 68). (2) After the report window — from Aug 24, 2026 — the grok line runs Grok 4.6; every grok forecast in this window and in the week-2 comparison is grok-4.5. (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 this issue, after an audit found the raw table mixes two exchanges; issue #2 row is as published.
- Research-to-date counter pinned from this issue: directional forecasts resolved inside the FH gate (1h/4h/1d) since Jul 11, counted at the issue cutoff (Aug 24 16:00 UTC). Issue #2 printed 23,458 under an earlier, unpinned definition; treat cross-issue counter deltas across the pin as definitional, not additive.

## Counters & lineage

| Counter | Value | Counter | Value |
|---|---|---|---|
| Forecasts total | 9,590 | Mature (scored pool) | 9,260 |
| OK in gate | 9,260 | -- of them directional | 5,130 |
| Out of gate (1w / 1M) | 209 / 68 | -- of them sideways | 4,130 |
| Invalid | 53 | Pending (next issue) | 0 |
| Late closes | 0 | Uptime, all grids | 266 / 266 slots |
| Source file | weekly_metrics_2026-08-17.json | Report cutoff | Mon Aug 24, 2026, 16:00 UTC |
| Generated at (pipeline) | 2026-08-26 05:22 UTC | Methodology | v1.1 (2026-08-10) · hash e66c7e8c864a2233 |

**THIS REPORT: 5,130 scored observations.** MARKETMANIA RESEARCH TO DATE (as of Aug 24, 2026 cutoff): 33,809 directional forecasts resolved since Jul 11 · 11 models tracked (7 current + 4 archived legacy) · 5 assets · 5 horizons · hourly · 13 published reports.

> **Issue #3.** Weekly Model Watch is a living series. Issue #2's open questions -- would a third reversal repeat the blind spot, does SOL's readability bounce back, and does the trend-day agreement premium reappear? -- closed 'no' (both ETH turns found callers), 'yes' (38.9% -> 54.3%) and 'yes' (64.8% trend vs 47.5% flat on 4h TF-agreement, now on a readable split). After the report window — from Aug 24, 2026 — the grok line runs Grok 4.6; every grok forecast in this window and in the week-2 comparison is grok-4.5. Engine 1.1 has powered the sandbox since Aug 18, i.e. from day 2 of this window; no cross-engine PnL comparisons are claimed.

## What we're testing next

- Next issue: does the leaderboard reshuffle again when the tape calms, or is claude-opus-5's +17.5pp the start of something? Does the reversal detector find callers on a non-ETH turn?
- Does the every-ticker-above-51% floor survive a flat week?
- Monthly test (Sep 2): cross-TF/FH lifts with regime control across mixed weeks (trend vs. flat), block bootstrap.

## Related research

| Report | Direct PDF link |
|---|---|
| Consensus Watch #3 | https://marketmania.ai/research/reports/consensus-watch-2026-08-17.pdf |
| Weekly Calibration #3 | https://marketmania.ai/research/reports/weekly-calibration-2026-08-17.pdf |
| Weekly Model Watch #2 | https://marketmania.ai/research/reports/model-watch-2026-08-10.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_2026w34,
  title  = {Weekly Model Watch #3: weekly leaderboard, ticker/pair reads and cross-confirmation, Aug 17-23 2026},
  author = {{MarketMania Research}},
  year   = {2026}, month = {August}, day = {26},
  url    = {https://marketmania.ai/research/reports/model-watch-2026-08-17.pdf},
  note   = {Methodology v1.1 (2026-08-10), hash e66c7e8c864a2233; window Aug 17-23, 2026 UTC; source weekly_metrics_2026-08-17.json}
}
```