# Weekly Calibration #4

**WEEKLY · CALIBRATION** · September 2, 2026 · MarketMania Research · Weekly series

Window: **Aug 24-30, 2026 UTC** · Cutoff: **Mon Aug 31, 2026, 16:00 UTC** · Methodology **v1.1 (2026-08-10)**, hash `e66c7e8c864a2233`

PDF: https://marketmania.ai/research/reports/weekly-calibration-2026-08-24.pdf · Open data (JSON): https://marketmania.ai/research/reports/weekly-calibration-2026-08-24.json

> Research question: *"when a model says 70, does it hit 70% of the time?"*

## Research Snapshot

| RESEARCH SNAPSHOT |  |  |  |
|---|---|---|---|
| **4,576** directional calls scored | **7** models (grok line spliced) | of **9,301** mature forecasts | window **Aug 24-30, 2026 UTC** |
| hit = direction rule: tp1/tp2->hit, sl->miss, expiry->sign of gross pnl | FH gate **1h / 4h / 1d** | base field hit **42.9%** vs **62.6** stated | cutoff **Mon Aug 31, 2026, 16:00 UTC** |
| market: BTC net **-0.07%** (prior +23.58%) | ann vol **30.9%** (was 71.1%) | TOP5 volume **$19.72B**, -24% w/w | pairwise corr **0.85** (was 0.81) |

## Key finding

> **KEY FINDING · OBSERVATION (one weekly window)**
>
> ### The field's overconfidence gap moved +8.6pp -> +19.6pp week over week, 0 of 7 model lines narrowed, and field Brier went 0.2559 -> 0.2884.

## TL;DR

- **OBSERVATION -- The gap read +19.6pp this week.** Directional calls hit 42.9% against 62.6 stated mean confidence (issue #3: +8.6pp on 54.0% vs 62.7). Field Brier is 0.2884 (was 0.2559) and 0 of 7 model lines scored below 0.25, the score of an uninformative always-50% predictor. One window, descriptive.
- **No model narrowed its gap.** Gap moves ran from deepseek-v4-pro's +6.5pp (+9.4pp -> +15.9pp) to claude-opus-5's +16.6pp (+2.2pp -> +18.8pp); 0 of 7 model lines narrowed week over week. deepseek-v4-pro is both the best-calibrated line (Brier 0.2760) and the hit-rate leader (44.6%).
- **The high-confidence buckets, N>=10 only.** deepseek-v4-pro 70-80 27.3% (n=11), gemini-3.1-pro 70-80 43.3% (n=261), gemini-3.1-pro 80-100 36.4% (n=11), gpt-5.6-sol 70-80 35.2% (n=210). Cells below N=10 are marked insufficient and rank nothing.
- **Trading, kept separate from every calibration table.** 0 of 7 lines finished net-positive; net PnL ran -$58.75 (grok-4.6 (spliced), best) to -$88.36 (gpt-5.6-sol, worst), field -$473.59 against issue #3's +$889.37.

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

Overconfidence gap in percentage points (mean stated confidence minus hit-rate), week over week, sorted by this week's gap. 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. The grok column compares the spliced 4.5+4.6 line against issue #3's pure grok-4.5. "Was" values are the numbers published in issue #3 for the back-to-back window Aug 17-23.

## Why it matters

Every MarketMania forecast carries a model-stated confidence from 0 to 100 alongside its direction call. This report checks whether that number tracks reality: on a well-calibrated forecaster, calls made at 70% confidence should hit their direction about 70% of the time. With issue #4 the series has four points on every model's gap -- enough to see movement, not enough to claim a trend: calibration and market regime move together, and durability claims belong in the Monthly series, not here.

## How to read this

Gap pp = mean stated confidence minus hit-rate, in percentage points; positive means overconfident. Brier is the mean squared error of the stated probability against the realised outcome, so lower is better and 0.25 is what an uninformative always-50% predictor scores. Coverage = calls scored divided by the mature calls available to that model. Confidence buckets are the models' own natural breakpoints, not an arbitrary binning; cells below N=10 are marked insufficient and never used to rank anything. Prediction and trading metrics live in separate sections and are never combined. The grok line is the lineage splice grok-4.6 (incl. 4.5-era, Aug 24 00:00-09:22).

## Calibration by model (sorted by Brier, best first)

| Model | N | Cov. | Hit rate | Mean conf | Gap pp | Brier | 95% CI |
|---|---|---|---|---|---|---|---|
| deepseek-v4-pro | 729 | 54.8% | 44.6% | 60.5 | +15.9 | 0.2760 | 41.0%-48.2% |
| qwen-3.8-max | 672 | 50.5% | 42.6% | 59.4 | +16.9 | 0.2772 | 38.9%-46.3% |
| claude-fable-5 | 651 | 49.2% | 43.0% | 60.9 | +17.9 | 0.2799 | 39.3%-46.8% |
| claude-opus-5 | 561 | 42.2% | 42.4% | 61.2 | +18.8 | 0.2816 | 38.4%-46.6% |
| grok-4.6 (spliced) | 479 | 36.0% | 39.9% | 60.2 | +20.3 | 0.2825 | 35.6%-44.3% |
| gemini-3.1-pro | 755 | 56.9% | 44.2% | 66.7 | +22.5 | 0.3025 | 40.7%-47.8% |
| gpt-5.6-sol | 729 | 54.8% | 42.7% | 67.5 | +24.8 | 0.3133 | 39.1%-46.3% |
| **Field (all models)** | **4,576** | - | **42.9%** | **62.6** | **+19.6** | **0.2884** | - |

N = directional calls scored this window. Cov. = coverage, scored / mature calls available to that model. Gap pp = mean stated confidence minus hit-rate (positive = overconfident). 95% CI is the descriptive Wilson interval on hit-rate. grok-4.6 (spliced) = grok-4.6 (incl. 4.5-era, Aug 24 00:00-09:22).

## Calibration curve: hit-rate by stated-confidence bucket

| Model | 50-60 | 60-70 | 70-80 | 80-100 |
|---|---|---|---|---|
| claude-fable-5 | 44.7% (199) | 42.3% (452) | n/a (0) | n/a (0) |
| claude-opus-5 | 46.0% (100) | 41.6% (461) | n/a (0) | n/a (0) |
| deepseek-v4-pro | 48.9% (264) | 42.5% (454) | 27.3% (11) | n/a (0) |
| gemini-3.1-pro | 50.0% (6) * | 44.9% (477) | 43.3% (261) | 36.4% (11) |
| gpt-5.6-sol | 75.0% (4) * | 45.3% (514) | 35.2% (210) | 100.0% (1) * |
| grok-4.6 (spliced) | 39.8% (181) | 40.1% (297) | 0.0% (1) * | n/a (0) |
| qwen-3.8-max | 45.9% (325) | 39.2% (339) | 0.0% (2) * | n/a (0) |

n in parentheses. * = N<10 (insufficient) -- no conclusions drawn from these cells. "n/a (0)" = no calls landed in that bucket this week. Sub-50 bucket (n greater than 0 only): qwen-3.8-max 66.7% (n=6) *. Mid-scale ranking: 1 of the 5 models with N>=10 in both 50-60 and 60-70 out-hit 50-60 from 60-70 this week (issue #3: 4 of 5).

## Hit-rate by forecast horizon (FH)

| Model | 1h | 4h | 1d |
|---|---|---|---|
| claude-fable-5 | 42.3% (416) | 42.5% (188) | 51.1% (47) |
| claude-opus-5 | 42.1% (330) | 42.5% (186) | 44.4% (45) |
| deepseek-v4-pro | 43.9% (435) | 44.1% (236) | 51.7% (58) |
| gemini-3.1-pro | 44.0% (493) | 42.4% (212) | 54.0% (50) |
| gpt-5.6-sol | 43.9% (458) | 39.9% (233) | 44.7% (38) |
| grok-4.6 (spliced) | 38.3% (300) | 41.1% (151) | 50.0% (28) |
| qwen-3.8-max | 43.0% (430) | 40.6% (197) | 46.7% (45) |

n in parentheses; * = N<10 (insufficient). 7 of 7 model lines hit better on 1d than on 1h this window (issue #3: 7 of 7).

## Trading result (a different question)

A different question: is any of this profitable to trade. "Accurate" and "profitable" are not the same thing, and per methodology this section is never mixed into the prediction/calibration metrics above.

| Model | Trades | WR | Net PnL | Gross PnL | Max DD |
|---|---|---|---|---|---|
| grok-4.6 (spliced) | 479 | 35.1% | -$58.75 | -$10.85 | -$63.22 |
| deepseek-v4-pro | 729 | 39.4% | -$59.07 | +$13.83 | -$67.18 |
| claude-opus-5 | 561 | 36.9% | -$59.66 | -$3.56 | -$65.63 |
| claude-fable-5 | 651 | 36.9% | -$62.85 | +$2.25 | -$76.89 |
| gemini-3.1-pro | 755 | 37.2% | -$67.73 | +$7.77 | -$78.95 |
| qwen-3.8-max | 672 | 36.5% | -$77.17 | -$9.97 | -$77.17 |
| gpt-5.6-sol | 729 | 36.5% | -$88.36 | -$15.46 | -$88.36 |

0 of 7 model lines finished net-positive and 3 gross-positive this week (issue #3: 7 of 7 on both). Ranked best-to-worst net PnL. Per methodology this table is never combined with the calibration tables above; a week's PnL is regime as much as skill.

## Platform note: TP/SL calibration

> **Context, not a finding.** Since Aug 13 the public sandbox can trade with model-specific TP/SL multipliers learned from this same weekly history (v1); since Aug 19, v2 adds per-ticker and confidence-bucket (50-70 / 70-100) resolution. That feature consumes calibration history; it does not feed back into any table in this report, which measures stated confidence vs. direction-hit only. No effect size is claimed for v2. The toggle lives at marketmania.ai/indices.

## Market check: the tape cooled off

> **MARKET CHECK · OBSERVATION (two adjacent calendar weeks)**
>
> OBSERVATION — mean |1d move| 4.28% -> 2.35% (-45% rel), BTC realized vol 58.7% -> 37.9% (ann., hourly); field directional accuracy 54.0% -> 42.9% (-11.1pp), 0 of 7 models improved, sim win-rate up for 0 of 7, field sim PnL +$889.37 -> -$473.59 (adjacent calendar weeks Aug 17-23 vs Aug 24-30; descriptive, one pair of weeks, not a claim).
>
> Robustness: raw price-sign accuracy 53.5% -> 44.2% — the same direction as the trade-based hit rule.

| Measure | Aug 17-23 (week A) | Aug 24-30 (week B) | Change |
|---|---|---|---|
| Mean \|1d move\| (5 assets) | 4.28% | 2.35% | -45% rel |
| BTC realized vol (ann., hourly) | 58.7% | 37.9% | -20.8pp |
| Field directional accuracy | 54.0% | 42.9% | -11.1pp |
| Models improving hit-rate | -- | 0 of 7 | -- |
| Raw price-sign accuracy | 53.5% | 44.2% | -9.3pp |
| Field sim win-rate | 46.3% | 37.0% | -9.2pp |
| Field sim net PnL | +$889.37 | -$473.59 | -- |

Week A is exactly the issue-#3 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, inside week B). Week A (2026-08-17..2026-08-23) is pure grok-4.5 and reproduces the published issue #3; week B mixes the 4.5-era (Aug 24 00:00-09:22) with 4.6.

## Practical implications

- Stated confidence is still a ranking hint at best, not a probability: the field's gap is +19.6pp this week against +8.6pp in issue #3.
- Gap and regime move together. A field that hits 42.9% instead of 54.0% closes or opens a confidence gap without a single stated number changing.
- Trading direction this week (0 of 7 lines net-positive) is a regime read, not a strategy result; the same lines printed the opposite sign inside three weeks.

## Limitations

- Prediction metrics (this report) and trading metrics are kept in separate sections per methodology -- they are never combined into a single score.
- 95% Wilson CIs shown are descriptive, not inferential: observations inside one window are dependent, so read them as a range, not a formal coverage guarantee.
- All 4,576 scored calls sit inside one market regime. This window ran 3 trend days of 7 (issue #3: 5 of 7); every comparison with issue #3 is a comparison across regimes as well as across weeks. Four week-over-week points cannot separate drift from regime; no durability claim is made.
- Models report confidence at discrete levels, not a continuous scale; the buckets reflect those natural breakpoints. Cells below N=10 are marked insufficient.
- Underlying price series are reconstructed from trade entry prices (median per symbol-slot), not an independent tick feed. Series density and lineage notes (wave 4): (1) rolling-1w forecasts emit daily since Aug 22, 2026 and rolling-1M daily since Aug 23, 2026, so inside this window (Aug 24-30) daily coverage is FULL for the first time in the series: 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 490). (2) The grok line flipped 4.5 -> 4.6 at Aug 24, 2026 09:22 UTC, INSIDE this window: every grok row in this issue is the lineage splice "grok-4.6 (incl. 4.5-era, Aug 24 00:00-09:22)" (110 rows from the 4.5 era, 1,360 from 4.6); issue #3 was pure grok-4.5, so the grok week-over-week row joins a spliced line to a single-model line. (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 #3 after the audit that found the raw table mixes two exchanges; issue #3 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 (Aug 31 16:00 UTC) -- the definition pinned in issue #3, which printed 33,809 at the Aug 24 16:00 cutoff. The pack reproduces that pin at the previous cutoff (control OK), so this issue's 38,385 is an additive step under one definition; counter deltas against issue #2 and earlier remain definitional.

## Counters & lineage

| Counter | Value | Counter | Value |
|---|---|---|---|
| Forecasts total | 10,290 | Mature (scored pool) | 9,301 |
| OK in gate | 9,301 | -- of them directional | 4,576 |
| Out of gate (1w / 1M) | 488 / 490 | -- of them sideways | 4,725 |
| Invalid | 11 | Pending (next issue) | 0 |
| Late closes | 0 | Uptime, all grids | 266 / 266 slots |
| Source file | weekly_metrics_2026-08-24.json | Report cutoff | Mon Aug 31, 2026, 16:00 UTC |
| Generated at (pipeline) | 2026-09-02 06:14 UTC | Methodology | v1.1 (2026-08-10) · hash e66c7e8c864a2233 |

**THIS REPORT: 4,576 scored observations.** MARKETMANIA RESEARCH TO DATE (as of Aug 31, 2026 cutoff): 38,385 directional forecasts resolved since Jul 11 · 7 models tracked (current line-up; earlier versions folded into their successors' lineage) · 5 assets · 5 horizons · hourly · 16 published reports.

> **Issue #4.** Weekly Calibration is a living series. Issue #3's open questions -- does the halved gap survive another week, and does gpt-5.6-sol's 70-80 bucket stay below its own 60-70? -- read this issue as: field gap +8.6pp -> +19.6pp (0 of 7 lines narrowed), and gpt-5.6-sol's 70-80 bucket read 35.2% (n=210) against its own 60-70 at 45.3% (n=514). The grok line is spliced this issue (grok-4.6 (incl. 4.5-era, Aug 24 00:00-09:22)) because the 4.5 -> 4.6 flip happened inside the window. 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 the field gap stay inside one band across two regimes, or does it track the base rate one-for-one?
- Do any two models keep the same Brier ordering for a third issue running?
- Monthly series: is the overconfidence gap stable across market regimes (trend vs flat) at monthly n?

## Related research

| Report | Direct PDF link |
|---|---|
| Consensus Watch #4 | https://marketmania.ai/research/reports/consensus-watch-2026-08-24.pdf |
| Weekly Model Watch #4 | https://marketmania.ai/research/reports/model-watch-2026-08-24.pdf |
| Weekly Calibration #3 | https://marketmania.ai/research/reports/weekly-calibration-2026-08-17.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_calibration_2026w35,
  title  = {Weekly Calibration #4: confidence vs. direction-hit, Aug 24-30 2026},
  author = {{MarketMania Research}},
  year   = {2026}, month = {September}, day = {2},
  url    = {https://marketmania.ai/research/reports/weekly-calibration-2026-08-24.pdf},
  note   = {Methodology v1.1, hash e66c7e8c864a2233; source weekly_metrics_2026-08-24.json}
}
```