# Weekly Calibration #5

**WEEKLY · CALIBRATION** · September 8, 2026 · MarketMania Research · Weekly series

Window: **Aug 31-Sep 6, 2026 UTC** · Cutoff: **Mon Sep 7, 2026, 16:00 UTC** · Methodology **v1.1 (2026-08-10)**, hash `e66c7e8c864a2233`

PDF: https://marketmania.ai/research/reports/weekly-calibration-2026-08-31.pdf · Open data (JSON): https://marketmania.ai/research/reports/weekly-calibration-2026-08-31.json

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

## Research Snapshot

| RESEARCH SNAPSHOT |  |  |  |
|---|---|---|---|
| **4,520** directional calls scored | **7** models (stable lineup) | of **9,264** mature forecasts | window **Aug 31-Sep 6, 2026 UTC** |
| hit = direction rule: tp1/tp2->hit, sl->miss, expiry->sign of gross pnl | FH gate **1h / 4h / 1d** | base field hit **42.8%** vs **62.4** stated | cutoff **Mon Sep 7, 2026, 16:00 UTC** |
| market: BTC net **+3.42%** (prior -0.07%) | ann vol **43.4%** (was 30.9%) | TOP5 volume **$16.24B**, -18% w/w | pairwise corr **0.78** (was 0.85) |

## Key finding

> **KEY FINDING · OBSERVATION (one weekly window)**
>
> ### The field's overconfidence gap moved +19.6pp -> +19.6pp week over week, 2 of 7 model lines narrowed, and field Brier went 0.2884 -> 0.2874.

## TL;DR

- **OBSERVATION -- The gap read +19.6pp this week.** Directional calls hit 42.8% against 62.4 stated mean confidence (issue #4: +19.6pp on 42.9% vs 62.6). Field Brier is 0.2874 (was 0.2884) and 0 of 7 model lines scored below 0.25, the score of an uninformative always-50% predictor. One window, descriptive.
- **2 of 7 models narrowed their gap.** Gap moves ran from gemini-3.1-pro's -2.8pp (+22.5pp -> +19.7pp) to grok-4.6's +2.0pp (+20.3pp -> +22.3pp); 2 of 7 model lines narrowed week over week. qwen-3.8-max is the best-calibrated line (Brier 0.2755); gemini-3.1-pro leads on hit-rate (46.5%).
- **The high-confidence buckets, N>=10 only.** deepseek-v4-pro 70-80 52.6% (n=19), gemini-3.1-pro 70-80 42.2% (n=251), gpt-5.6-sol 70-80 35.0% (n=217). 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 -$81.09 (gemini-3.1-pro, best) to -$112.37 (gpt-5.6-sol, worst), field -$669.03 against issue #4's -$473.59.

### 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 this window's pure 4.6 line against issue #4's spliced 4.5+4.6 line. "Was" values are the numbers published in issue #4 for the back-to-back window Aug 24-30.

## 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 #5 the series has five 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 4.5 -> 4.6 flip (Aug 24, 2026 09:22 UTC) sits inside the PREVIOUS window (week A, issue #4): every issue-#4 grok number is the lineage splice "grok-4.6 (incl. 4.5-era, Aug 24 00:00-09:22)"; this window is pure grok-4.6, so the grok week-over-week row compares a pure-4.6 week with a spliced week.

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

| Model | N | Cov. | Hit rate | Mean conf | Gap pp | Brier | 95% CI |
|---|---|---|---|---|---|---|---|
| qwen-3.8-max | 697 | 53.5% | 42.5% | 59.1 | +16.6 | 0.2755 | 38.9%-46.2% |
| deepseek-v4-pro | 724 | 54.6% | 44.2% | 60.6 | +16.5 | 0.2761 | 40.6%-47.8% |
| claude-fable-5 | 628 | 47.2% | 42.8% | 60.7 | +17.9 | 0.2795 | 39.0%-46.7% |
| claude-opus-5 | 545 | 41.1% | 42.4% | 61.2 | +18.8 | 0.2804 | 38.3%-46.6% |
| grok-4.6 | 476 | 36.0% | 37.6% | 59.9 | +22.3 | 0.2883 | 33.4%-42.0% |
| gemini-3.1-pro | 761 | 57.3% | 46.5% | 66.2 | +19.7 | 0.2922 | 43.0%-50.1% |
| gpt-5.6-sol | 689 | 51.8% | 41.6% | 67.8 | +26.2 | 0.3183 | 38.0%-45.4% |
| **Field (all models)** | **4,520** | - | **42.8%** | **62.4** | **+19.6** | **0.2874** | - |

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. The grok 4.5 -> 4.6 flip (Aug 24, 2026 09:22 UTC) sits inside the PREVIOUS window (week A, issue #4): every issue-#4 grok number is the lineage splice "grok-4.6 (incl. 4.5-era, Aug 24 00:00-09:22)"; this window is pure grok-4.6, so the grok week-over-week row compares a pure-4.6 week with a spliced week.

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

| Model | 50-60 | 60-70 | 70-80 | 80-100 |
|---|---|---|---|---|
| claude-fable-5 | 43.2% (199) | 42.7% (429) | n/a (0) | n/a (0) |
| claude-opus-5 | 46.0% (113) | 41.4% (432) | n/a (0) | n/a (0) |
| deepseek-v4-pro | 49.2% (244) | 41.2% (461) | 52.6% (19) | n/a (0) |
| gemini-3.1-pro | 40.0% (5) * | 48.6% (498) | 42.2% (251) | 57.1% (7) * |
| gpt-5.6-sol | 40.0% (5) * | 44.9% (466) | 35.0% (217) | 0.0% (1) * |
| grok-4.6 | 46.2% (186) | 32.1% (290) | n/a (0) | n/a (0) |
| qwen-3.8-max | 44.4% (363) | 40.1% (324) | 50.0% (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 50.0% (n=8) *. Mid-scale ranking: 0 of the 5 models with N>=10 in both 50-60 and 60-70 out-hit 50-60 from 60-70 this week (issue #4: 1 of 5).

## Hit-rate by forecast horizon (FH)

| Model | 1h | 4h | 1d |
|---|---|---|---|
| claude-fable-5 | 45.1% (410) | 38.1% (181) | 40.5% (37) |
| claude-opus-5 | 45.0% (342) | 38.5% (174) | 34.5% (29) |
| deepseek-v4-pro | 46.9% (469) | 38.3% (209) | 43.5% (46) |
| gemini-3.1-pro | 48.4% (514) | 42.6% (209) | 42.1% (38) |
| gpt-5.6-sol | 43.5% (430) | 38.5% (221) | 39.5% (38) |
| grok-4.6 | 42.7% (314) | 29.3% (140) | 18.2% (22) |
| qwen-3.8-max | 43.5% (446) | 39.2% (209) | 47.6% (42) |

n in parentheses; * = N<10 (insufficient). 1 of 7 model lines hit better on 1d than on 1h this window (issue #4: 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 |
|---|---|---|---|---|---|
| gemini-3.1-pro | 761 | 37.5% | -$81.09 | -$4.99 | -$89.79 |
| claude-fable-5 | 628 | 36.3% | -$83.56 | -$20.76 | -$87.66 |
| claude-opus-5 | 545 | 34.7% | -$89.04 | -$34.54 | -$92.14 |
| qwen-3.8-max | 697 | 36.0% | -$92.91 | -$23.21 | -$98.00 |
| deepseek-v4-pro | 724 | 36.9% | -$104.50 | -$32.10 | -$109.34 |
| grok-4.6 | 476 | 32.1% | -$105.56 | -$57.96 | -$106.14 |
| gpt-5.6-sol | 689 | 34.5% | -$112.37 | -$43.47 | -$117.14 |

0 of 7 model lines finished net-positive and 0 gross-positive this week (issue #4: 0 of 7 net-positive, 3 gross-positive). 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: an up week on lower volume

> **MARKET CHECK · OBSERVATION (two adjacent calendar weeks)**
>
> OBSERVATION — mean |1d move| 2.35% -> 2.10% (-11% rel), BTC realized vol 37.9% -> 35.2% (ann., hourly); field directional accuracy 42.9% -> 42.8% (-0.1pp), 1 of 7 models improved, sim win-rate up for 1 of 7, field sim PnL -$473.59 -> -$669.03 (adjacent calendar weeks Aug 24-30 vs Aug 31-Sep 6; descriptive, one pair of weeks, not a claim).
>
> Robustness: raw price-sign accuracy 44.2% -> 51.6% — the opposite direction to the trade-based hit rule.

| Measure | Aug 24-30 (week A) | Aug 31-Sep 6 (week B) | Change |
|---|---|---|---|
| Mean \|1d move\| (5 assets) | 2.35% | 2.10% | -11% rel |
| BTC realized vol (ann., hourly) | 37.9% | 35.2% | -2.7pp |
| Field directional accuracy | 42.9% | 42.8% | -0.1pp |
| Models improving hit-rate | -- | 1 of 7 | -- |
| Raw price-sign accuracy | 44.2% | 51.6% | +7.4pp |
| Field sim win-rate | 37.0% | 35.6% | -1.4pp |
| Field sim net PnL | -$473.59 | -$669.03 | -- |

Week A is exactly the issue-#4 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 A). Week A (2026-08-24..2026-08-30) mixes the 4.5-era (Aug 24 00:00-09:22) with 4.6 and reproduces the published issue #4; week B is pure grok-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 +19.6pp in issue #4.
- Gap and regime move together. The field hit 42.8% against 42.9% in issue #4 and stated mean confidence read 62.4 against 62.6, so the gap held at +19.6pp: a field that hits 42.8% instead of 42.9% 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 4,520 scored calls sit inside one market regime. This window ran 3 trend days of 7 (issue #4: 3 of 7); every comparison with issue #4 is a comparison across regimes as well as across weeks. Five 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 5): (1) rolling-1w forecasts emit daily since Aug 22, 2026 and rolling-1M daily since Aug 23, 2026, so inside this window (Aug 31-Sep 6) daily coverage is FULL, as it was in issue #4 (the first full week of 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 490, 1M 490). (2) The grok line flipped 4.5 -> 4.6 at Aug 24, 2026 09:22 UTC, inside the PREVIOUS window (week A, issue #4): this window carries 1,470 grok rows and none from the 4.5 era, so every grok number in this issue is a pure grok-4.6 line, while every issue-#4 grok number is the lineage splice "grok-4.6 (incl. 4.5-era, Aug 24 00:00-09:22)" -- the grok week-over-week row compares a pure-4.6 week with a spliced week. (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 #4 after the audit that found the raw table mixes two exchanges; issue #4 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 7 16:00 UTC) -- the definition pinned in issue #3 and carried by issue #4, which printed 38,385 at the Aug 31 16:00 cutoff. The pack reproduces that pin at the previous cutoff (control OK), so this issue's 42,905 is an additive step under one definition; counter deltas against issue #3 and earlier remain definitional.

## Counters & lineage

| Counter | Value | Counter | Value |
|---|---|---|---|
| Forecasts total | 10,290 | Mature (scored pool) | 9,264 |
| OK in gate | 9,264 | -- of them directional | 4,520 |
| Out of gate (1w / 1M) | 490 / 490 | -- of them sideways | 4,744 |
| Invalid | 46 | Pending (next issue) | 0 |
| Late closes | 0 | Uptime, all grids | 266 / 266 slots |
| Source file | weekly_metrics_2026-08-31.json | Report cutoff | Mon Sep 7, 2026, 16:00 UTC |
| Generated at (pipeline) | 2026-09-08 08:08 UTC | Methodology | v1.1 (2026-08-10) · hash e66c7e8c864a2233 |

**THIS REPORT: 4,520 scored observations.** MARKETMANIA RESEARCH TO DATE (as of Sep 7, 2026 cutoff): 42,905 directional forecasts resolved since Jul 11 · 7 models tracked (current line-up; earlier versions folded into their successors' lineage) · 5 assets · 5 horizons · hourly · 24 published reports.

> **Issue #5.** Weekly Calibration is a living series. Issue #4's open questions -- does the field gap stay inside one band or track the base rate one-for-one, and do any two models keep the same Brier ordering for a third issue running? -- read this issue as: field gap +19.6pp -> +19.6pp (2 of 7 lines narrowed) on a field base that moved -0.1pp, and a new best-Brier line (qwen-3.8-max at 0.2755; issue #4: deepseek-v4-pro at 0.2760), so the three-issue ordering question stays open. Also on the record: gpt-5.6-sol's 70-80 bucket read 35.0% (n=217) against its own 60-70 at 44.9% (n=466). The grok row is pure grok-4.6 this issue; issue #4's grok row was the lineage splice "grok-4.6 (incl. 4.5-era, Aug 24 00:00-09:22)". 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 hold near +19.6pp for a third issue, or does it track the base rate one-for-one?
- Does the best-Brier line change hands again, or does one model hold it two issues 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 #5 | https://marketmania.ai/research/reports/consensus-watch-2026-08-31.pdf |
| Weekly Model Watch #5 | https://marketmania.ai/research/reports/model-watch-2026-08-31.pdf |
| Weekly Calibration #4 | https://marketmania.ai/research/reports/weekly-calibration-2026-08-24.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_2026w36,
  title  = {Weekly Calibration #5: confidence vs. direction-hit, Aug 31-Sep 6 2026},
  author = {{MarketMania Research}},
  year   = {2026}, month = {September}, day = {8},
  url    = {https://marketmania.ai/research/reports/weekly-calibration-2026-08-31.pdf},
  note   = {Methodology v1.1, hash e66c7e8c864a2233; source weekly_metrics_2026-08-31.json}
}
```