# Weekly Model Watch #2

**WEEKLY · MODEL WATCH** · 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/model-watch-2026-08-10.pdf · Open data (JSON): https://marketmania.ai/research/reports/model-watch-2026-08-10.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 |  |  |  |
|---|---|---|---|
| window **Aug 10-16, 2026 UTC** | **7** models (stable lineup) | **3,670** directional calls scored of **9,307** mature | FH gate **1h / 4h / 1d** |
| hit = direction rule: tp1/tp2->hit, sl->miss, expiry->sign of gross pnl | base field hit **43.8%** | cutoff **Mon Aug 17, 2026, 16:00 UTC** | methodology **v1.1 (2026-08-10)**, 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 reversal blind spot repeated: zero of 7 models caught ETH's -2.65% Monday turn -- two issues, two qualifying reversals, zero callers.

## TL;DR

- OBSERVATION -- **No weekly title again.** gemini-3.1-pro leads descriptively at **47.0%** [42.9%, 51.2%] after the field's biggest jump (+3.3pp w/w), but the gap to #2 (gpt-5.6-sol, 44.2%) is 2.8pp with overlapping CIs -- the title rule (>=5pp gap AND non-overlapping 95% CI) was not met. Six of seven models improved week-over-week; last week's leader qwen-3.8-max held flat at 44.1% and slipped 1st -> 3rd.
- **Ticker readability reshuffled hard.** BNB is the week's most readable coin (49.3%, n=816) and last week's best, SOL, collapsed 48.2% -> 38.9% -- a -9.3pp swing to 4th of 5. ETH is the hardest for the second week running (35.6%). Pair of the week: gemini-3.1-pro x BNB 55.7% (n=122); worst: deepseek-v4-pro x ETH 29.9% (n=87).
- **The reversal blind spot repeated.** The week's only qualifying reversal -- ETH turning short at 08:00 UTC on Monday Aug 10, an 8-hour move of -2.65% -- drew zero callers from all 7 models. Two issues, two reversals (both on the window's Monday), first caller: NOBODY, twice.
- **Self-agreement still added nothing pooled -- but the first regime split landed.** Cross-TF/FH lifts ran -18.5 to +0.2pp, and 1d confirmation was a disaster (17.9-18.5% hit). On the window's single trend day, though, 4h TF-agreement hit 54.3% vs 38.3% on the six flat days (n=46 vs 264) -- the regime control's first non-degenerate read; a hint, not a result.

### Weekly leaderboard chart (see PDF for the week-over-week bar chart)

| Model | Issue #1 | Issue #2 | 95% CI (issue #2) |
|---|---|---|---|
| gemini-3.1-pro | 43.7% | 47.0% | 42.9%-51.2% |
| gpt-5.6-sol | 43.0% | 44.2% | 40.4%-48.1% |
| qwen-3.8-max | 44.1% | 44.1% | 40.0%-48.1% |
| grok-4.5 | 42.2% | 44.0% | 40.2%-48.0% |
| deepseek-v4-pro | 41.6% | 42.5% | 38.0%-47.1% |
| claude-fable-5 | 41.4% | 41.7% | 37.4%-46.2% |
| claude-opus-5 | 40.9% | 41.5% | 36.6%-46.7% |

Field base this week: 43.8% (issue #1: 42.4%).

## 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. With issue #2 the series gains its first week-over-week movement column -- and its first non-degenerate regime control.

## Leaderboard of the week

The full 7-model field, ranked by this week's directional hit-rate; W/w pp = movement vs issue #1 (begins this issue, as promised).

| Model | N | Coverage | Hit rate | 95% CI | W/w pp |
|---|---|---|---|---|---|
| **gemini-3.1-pro** | 555 | 41.7% | 47.0% | 42.9%-51.2% | +3.3 |
| gpt-5.6-sol | 636 | 47.8% | 44.2% | 40.4%-48.1% | +1.2 |
| qwen-3.8-max | 572 | 43.0% | 44.1% | 40.0%-48.1% | 0.0 |
| grok-4.5 | 620 | 46.6% | 44.0% | 40.2%-48.0% | +1.8 |
| deepseek-v4-pro | 454 | 34.2% | 42.5% | 38.0%-47.1% | +0.9 |
| claude-fable-5 | 472 | 35.5% | 41.7% | 37.4%-46.2% | +0.3 |
| claude-opus-5 | 361 | 27.1% | 41.5% | 36.6%-46.7% | +0.6 |

No weekly title is awarded this issue -- for the second time. Title rule: **>=5pp gap AND non-overlapping 95% CI, N>=10** (N>=10 per side). gemini-3.1-pro's lead over gpt-5.6-sol is 2.8pp (47.0% vs 44.2%) and their CIs overlap, so neither condition is met -- the lead is descriptive. Rank shuffle vs issue #1: gemini-3.1-pro 2nd -> 1st, qwen-3.8-max 1st -> 3rd (flat hit-rate in a rising field). Hit-rate rank and Brier (calibration) rank still diverge: the hit-rate leader gemini-3.1-pro is only 5th of 7 by Brier -- see the calibration bridge in Counters, and the full picture in Weekly Calibration #2.

## Ticker of the week

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

| Symbol | N | Hit rate | 95% CI |
|---|---|---|---|
| **BNB** | 816 | 49.3% | 45.9%-52.7% |
| XRP | 824 | 48.5% | 45.1%-52.0% |
| BTC | 727 | 44.0% | 40.5%-47.6% |
| SOL | 659 | 38.9% | 35.2%-42.6% |
| **ETH** | 644 | 35.6% | 32.0%-39.3% |

BNB was the field's most readable ticker this week (49.3%, n=816) and ETH the hardest for the second issue running (35.6%, n=644); their 95% CIs are well separated (45.9-52.7% vs 32.0-39.3%), while every adjacent pair in the ranked list overlaps its neighbor. The reshuffle is the story: last week's best, SOL (48.2%), fell to 4th at 38.9% -- a -9.3pp swing -- and last week's 3rd, BNB, took the top. Read per-ticker readability as weekly weather, not a durable per-asset skill claim; ETH-hardest is the only repeat so far.

## Pair of the week

Top 6 of the 12 model x ticker pairs tracked this week (best hit-rate), and the 5 worst. Pairs shown require N>=8 calls.

**Top 6 pairs**

| Model | Symbol | N | Hit rate |
|---|---|---|---|
| gemini-3.1-pro | BNB | 122 | 55.7% |
| qwen-3.8-max | BNB | 124 | 50.8% |
| gemini-3.1-pro | XRP | 133 | 50.4% |
| gpt-5.6-sol | BNB | 145 | 49.7% |
| qwen-3.8-max | XRP | 127 | 49.6% |
| deepseek-v4-pro | XRP | 97 | 49.5% |

**5 worst pairs**

| Model | Symbol | N | Hit rate |
|---|---|---|---|
| deepseek-v4-pro | SOL | 78 | 35.9% |
| gpt-5.6-sol | ETH | 113 | 35.4% |
| claude-opus-5 | ETH | 65 | 35.4% |
| claude-fable-5 | ETH | 84 | 33.3% |
| deepseek-v4-pro | ETH | 87 | 29.9% |

gemini-3.1-pro x BNB (55.7%, n=122) is the week's best pair -- the only pairing above 55% -- and the same model tops the leaderboard. deepseek-v4-pro sits at both ends: x XRP makes the top 6 (49.5%) while x SOL and x ETH (29.9%, the week's worst) anchor the bottom. Four of the five worst pairs are ETH pairings. Pair history accumulates across issues -- read this as the second 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** | Mon Aug 10, 08:00 | Short | -2.65% | 0 | **NOBODY** |

This week's only qualifying reversal was ETH's turn to short at 08:00 UTC on Monday, Aug 10 -- an 8-hour move of -2.65%. Zero of the 7 active models had a matured hit call in the new (short) direction inside the window; first caller: **NOBODY** -- the same outcome as issue #1's BTC turn. Two issues, two reversals, zero callers; both events landed on the window's Monday. Still a two-event sample: reported descriptively, not as a claim about turn-detection skill. Scope in v1 remains BTC/ETH only.

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

Within one forecast horizon (FH), does a model's call from a longer input timeframe (TF) agree with its call from a shorter one -- and if so, does the call do any better?

| FH | vs. | Agree n | Agree hit | Dis n | Dis hit | Mixed | Base hit | Lift pp |
|---|---|---|---|---|---|---|---|---|
| 4h | TF 4h vs 1h | 310 | 40.6% [35.3%-46.2%] | 20 | 35.0% [18.1%-56.7%] | 247 | 41.2% (n=577) | -0.6 |
| 1d | TF 1d vs 4h | 56 | 17.9% [10.0%-29.8%] | 0 | n/a (0) | 43 | 36.4% (n=99) | -18.5 |

n in Base hit column is the base-rate call count. `\*` = N<10 (insufficient); `#` = N<20 (thin); "n/a (0)" = no disagreeing calls. Insufficient/thin cells are never used to rank models. The 1d row is the week's ugliest number: when a model's 1d call was confirmed by its own 4h-TF read, it hit 17.9% -- 18.5pp BELOW the 1d base.

**Regime control (first non-degenerate split).** Methodology calls for a trend-vs-flat split on this table. Issue #1 had zero trend days; this window has exactly one -- Mon Aug 10 (-1.63%), the other six days were flat: 08-10 -1.63% (trend) · 08-11 -0.56% · 08-12 -0.02% · 08-13 -0.25% · 08-14 -0.67% · 08-15 +0.07% · 08-16 -0.35%. On that single trend day, 4h TF-agreement hit **54.3%** (n=46) vs **38.3%** on flat days (n=264); the disagree cells ran 16.7% (n=6) vs 42.9% (n=14). One trend day is a hint that self-agreement may be regime-dependent -- exactly what the monthly block-bootstrap test is for -- not a result to trade.

## Cross-FH confirmation: does a shorter horizon confirm a longer one?

Does a model's call at one forecast horizon (FH) agree with its call at a shorter FH on the same timeframe -- and if so, does the call do any better?

| FH | vs. | Agree n | Agree hit | Dis n | Dis hit | Mixed | Base hit | Lift pp |
|---|---|---|---|---|---|---|---|---|
| 4h | by FH 1h | 299 | 41.5% [36.0%-47.1%] | 31 | 29.0% [16.1%-46.6%] | 246 | 41.2% (n=577) | +0.2 |
| 1d | by FH 4h | 54 | 18.5% [10.4%-30.8%] | 2 | 100.0% [34.2%-100.0%] \* | 43 | 36.4% (n=99) | -17.8 |

The 4h FH-confirmation is the only cell family near breakeven this week (+0.2pp pooled; 48.9% on the trend day vs 40.2% flat). The 1d rows repeat the cross-TF story: 18.5% agree-hit, -17.8pp lift. Disagree cells stay tiny (n=0, 2, 20, 31); the 1d disagree pair (n=2, both hits) is insufficient and never used to rank models.

## Side-mix and the week's failure

How each model split its calls this week across long / short / sideways, and the week's single costliest miss.

| Model | N | Long | Short | Sideways | L:S |
|---|---|---|---|---|---|
| claude-fable-5 | 1,330 | 19.6% | 15.9% | 64.5% | 1.23 |
| claude-opus-5 | 1,330 | 14.5% | 12.6% | 72.9% | 1.15 |
| deepseek-v4-pro | 1,328 | 10.5% | 23.7% | 65.8% | 0.44 |
| gemini-3.1-pro | 1,330 | 16.2% | 25.5% | 58.3% | 0.64 |
| gpt-5.6-sol | 1,330 | 14.7% | 33.1% | 52.2% | 0.45 |
| grok-4.5 | 1,330 | 15.9% | 30.8% | 53.4% | 0.52 |
| qwen-3.8-max | 1,329 | 17.4% | 25.7% | 57.0% | 0.68 |

L:S = each model's own long-share divided by its short-share. The field flipped its skew: issue #1 ran ~2.2:1 long-over-short; this week the average of the 7 per-model ratios is **~0.7:1** -- short-over-long -- in a week BTC fell -3.08%. gpt-5.6-sol is the most short-committed (33.1% short vs 14.7% long); claude-fable-5 and claude-opus-5 are the only models still leaning long. 'Sideways' stayed the field's most common call: 60.6% of mature forecasts (5,637 of 9,307), up from 55.7%.

**Consensus skew.** 404 symbol/FH/TF cells this week had 5 or more of the (up to 7) active models on the same side. Those skewed cells' mature calls hit 41.8% (n=2,486) -- below the week's 43.8% field base rate, so leaning with a 5-of-7+ skew did not outperform the field for the second week running. (Strict unanimity is Consensus Watch's own metric -- see that report for the full herding analysis.)

| Model | Symbol | Side | Conf | FH / TF | Slot (UTC) | Exit | Net PnL |
|---|---|---|---|---|---|---|---|
| **gpt-5.6-sol** | ETH | Short | 82 | 4h / 1h | Aug 10, 16:01 | Expiry | **-$0.36** |

**Fail of the week.** The week's single highest-confidence individual miss: gpt-5.6-sol, ETH, short, confidence 82, FH 4h / TF 1h, slot Aug 10 16:01 UTC, exit by expiry, net -$0.36. Cross-reference: ETH's only qualifying reversal (Block 4) had turned short at 08:00 that same day -- this conf-82 short entered at 16:01, after the qualifying 8-hour counter-move had already completed, and expired without reaching TP. ETH was also the week's hardest ticker (35.6%), and 4 of the 5 worst pairs are ETH pairings.

## Counters and the calibration bridge

Gate totals, uptime and data-quality for this window, plus a bridge to this week's calibration read.

| Metric | Value |
|---|---|
| Forecasts total | 9,380 |
| OK in gate | 9,307 |
| Invalid | 4 |
| Out of gate (1w) | 69 |
| Mature (scored pool) | 9,307 |
|   - directional | 3,670 |
|   - sideways | 5,637 |
| Pending (next issue) | 0 |
| Late closes | 0 |
| Uptime, 1h slots (tf=1h) | 168 / 168 |
| Uptime, 4h slots (tf=4h) | 42 / 42 |
| Uptime, 4h slots (tf=1h) | 42 / 42 |
| Uptime, 1d slots (tf=1d) | 7 / 7 |
| Uptime, 1d slots (tf=4h) | 7 / 7 |
| Source file | weekly_metrics_2026-08-10.json |
| Generated at (metrics pipeline) | 2026-08-18 08:37 UTC |
| Methodology | v1.1 (2026-08-10), hash e66c7e8c864a2233 |
| Report cutoff | 2026-08-17 16:00 UTC (frozen) |

**Data quality.** First perfect-uptime week of the series: every 1h, 4h and 1d slot grid ran full (issue #1: 165/168 on the hourly grid). Per-model ok-rate was 99.85% for qwen-3.8-max and deepseek-v4-pro and 100.0% for the other five (claude-fable-5, claude-opus-5, gemini-3.1-pro, gpt-5.6-sol, grok-4.5).

**Calibration bridge.** qwen-3.8-max is again the week's best-calibrated model by Brier score (0.2688, lowest of the 7) -- but this issue it is only 3rd by hit-rate, while the hit-rate leader gemini-3.1-pro ranks 5th of 7 by Brier (0.2848) with a +18.4pp overconfidence gap. Hit-rate and calibration keep rewarding different behavior; the full confidence-bucket analysis (including the field's first working 70-80 bucket) lives in Weekly Calibration #2.

## Practical implications

- Do not read this week's leaderboard as a skill ranking: no title was awarded, the top CIs overlap, and gemini-3.1-pro's +3.3pp jump is one week of movement in a field where last week's leader just went flat.
- Treat per-ticker readability as weekly weather. The best coin flipped (SOL -> BNB, a -9.3pp swing for SOL); the only repeat so far is ETH-hardest, two weeks running -- and the field missed both Monday reversals outright, which still points at turns as the general weak spot.
- Do not use within-model cross-TF/FH agreement as a confidence booster: pooled lifts ran -18.5 to +0.2pp, and 1d confirmation was the week's worst cell family (~18% hit). The one trend day's 54.3% agree-hit is the thing to watch at monthly n, not a signal to act on.

## Limitations

- Single week (Aug 10-16, 2026 UTC). Every finding is descriptive for this window only; no claim is made about next week or any model's underlying skill.
- The regime control rests on ONE trend day (Mon Aug 10, n=46/45 agree calls); the other six days were flat. The split is reported because it is the series' first non-degenerate one, not because one day can establish anything.
- Observations inside one window are not independent: the same models watch overlapping symbol/FH/TF cells hour after hour; agree/disagree calls made close in time are correlated, not i.i.d. draws.
- 95% Wilson CIs shown throughout are descriptive, not inferential, for the reason above -- read them as a range, not a formal coverage guarantee.
- The reversal detector remains v1 and BTC/ETH-only, on price series reconstructed from trade entry prices (median per symbol-slot), not an independent tick feed. Two events in two issues cannot characterize detector performance either way.
- Cross-TF and cross-FH disagreement cells are tiny again (n=0, 2, 20, 31) -- cells below N=10 are marked insufficient and are never used to rank models.
- Week-over-week movement begins this issue on two data points; rank shuffles at these CI widths are expected noise until several issues accumulate.

> **THIS REPORT: 3,670 scored observations**
>
> MARKETMANIA RESEARCH TO DATE (as of Aug 17, 2026 cutoff): 23,458 resolved forecasts since Jul 11, 2026 · 9 models tracked (7 current frontier + 2 archived legacy generations) · 5 assets · 5 forecast horizons · hourly cadence · 9 published research reports incl. this wave

**Snapshot principle.** Every figure in this report is frozen at the Monday 16:00 UTC cutoff for the Aug 10-16 window. Any later data correction is handled as a note in a future issue, not by silently editing this one.

> **Issue #2.** Weekly Model Watch is a living series; the movement column and the regime control went live this issue. Issue #1's open questions -- does the reversal blind spot repeat, and does cross-TF agreement stay flat-to-negative? -- both closed 'yes': a second Monday reversal drew zero callers, and pooled self-agreement lifts stayed at -18.5 to +0.2pp (with the first trend-day hint of regime dependence). 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:** a third Monday reversal would make the blind spot a pattern; does SOL's readability bounce back or was issue #1 the outlier; does the trend-day agreement premium reappear?
- **Monthly test (Sep 2):** cross-TF/FH lifts with regime control across mixed weeks (trend vs. flat), block bootstrap.

## Related research

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

```bibtex
@misc{mm_model_watch_2026w33,
  title  = {Weekly Model Watch #2: weekly leaderboard, ticker/pair reads and cross-confirmation, Aug 10-16 2026},
  author = {{MarketMania Research}},
  year   = {2026}, month = {August}, day = {19},
  url    = {https://marketmania.ai/research/reports/model-watch-2026-08-10.pdf},
  note   = {Methodology v1.1 (2026-08-10), hash e66c7e8c864a2233; window Aug 10-16, 2026 UTC; source weekly_metrics_2026-08-10.json}
}
```