# Weekly Calibration #6

**WEEKLY · CALIBRATION** · September 15, 2026 · MarketMania Research · Weekly series

Window: **Sep 7-13, 2026 UTC** · Cutoff: **Mon Sep 14, 2026, 16:00 UTC** · Methodology **v1.1 (2026-08-10)**, hash `e66c7e8c864a2233`

PDF: https://marketmania.ai/research/reports/weekly-calibration-2026-09-07.pdf · Open data (JSON): https://marketmania.ai/research/reports/weekly-calibration-2026-09-07.json

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

## Research Snapshot

| RESEARCH SNAPSHOT |  |  |  |
|---|---|---|---|
| **4,468** directional calls scored | **7** models (stable lineup) | of **9,301** mature forecasts | window **Sep 7-13, 2026 UTC** |
| hit = direction rule: tp1/tp2->hit, sl->miss, expiry->sign of gross pnl | FH gate **1h / 4h / 1d** | base field hit **39.6%** vs **62.1** stated | cutoff **Mon Sep 14, 2026, 16:00 UTC** |
| market: BTC net **-4.36%** (prior +3.42%) | ann vol **19.7%** (was 43.4%) | TOP5 volume **$16.05B**, -1% w/w | pairwise corr **0.74** (was 0.78) |

## Key finding

> **KEY FINDING · OBSERVATION (one weekly window)**
>
> ### The field's overconfidence gap moved +19.6pp -> +22.5pp week over week, 1 of 7 model lines narrowed, and field Brier went 0.2874 -> 0.2948.

## TL;DR

- **OBSERVATION -- The gap read +22.5pp this week.** Directional calls hit 39.6% against 62.1 stated mean confidence (issue #5: +19.6pp on 42.8% vs 62.4). Field Brier is 0.2948 (was 0.2874) and 0 of 7 model lines scored below 0.25, the score of an uninformative always-50% predictor. One window, descriptive.
- **1 of 7 models narrowed their gap.** Gap moves ran from grok-4.6's -1.8pp (+22.3pp -> +20.5pp) to claude-fable-5's +5.7pp (+17.9pp -> +23.6pp); 1 of 7 model lines narrowed week over week. qwen-3.8-max is the best-calibrated line (Brier 0.2794); gemini-3.1-pro leads on hit-rate (41.8%).
- **The high-confidence buckets, N>=10 only.** gemini-3.1-pro 70-80 38.3% (n=193), gpt-5.6-sol 70-80 42.0% (n=269). 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 -$114.37 (grok-4.6, best) to -$163.27 (gpt-5.6-sol, worst), field -$1018.93 against issue #5's -$669.03.

### 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 two pure grok-4.6 weeks: the 4.5 -> 4.6 flip sits two windows back. "Was" values are the numbers published in issue #5 for the back-to-back window Aug 31-Sep 6.

## 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 #6 the series has six 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 two 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 first pure-4.6 against pure-4.6 comparison of the series.

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

| Model | N | Cov. | Hit rate | Mean conf | Gap pp | Brier | 95% CI |
|---|---|---|---|---|---|---|---|
| qwen-3.8-max | 707 | 53.4% | 40.0% | 58.8 | +18.8 | 0.2794 | 36.5%-43.7% |
| deepseek-v4-pro | 727 | 54.7% | 41.4% | 59.8 | +18.4 | 0.2825 | 37.9%-45.0% |
| grok-4.6 | 518 | 39.0% | 39.6% | 60.0 | +20.5 | 0.2847 | 35.5%-43.9% |
| claude-fable-5 | 624 | 46.9% | 36.5% | 60.1 | +23.6 | 0.2921 | 32.9%-40.4% |
| claude-opus-5 | 510 | 38.4% | 36.5% | 61.0 | +24.5 | 0.2932 | 32.4%-40.7% |
| gemini-3.1-pro | 673 | 50.6% | 41.8% | 65.4 | +23.6 | 0.3057 | 38.1%-45.5% |
| gpt-5.6-sol | 709 | 53.4% | 40.1% | 68.5 | +28.4 | 0.3230 | 36.5%-43.7% |
| **Field (all models)** | **4,468** | - | **39.6%** | **62.1** | **+22.5** | **0.2948** | - |

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 two 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 first pure-4.6 against pure-4.6 comparison of the series.

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

| Model | 50-60 | 60-70 | 70-80 | 80-100 |
|---|---|---|---|---|
| claude-fable-5 | 41.0% (234) | 33.9% (390) | n/a (0) | n/a (0) |
| claude-opus-5 | 36.8% (117) | 36.4% (393) | n/a (0) | n/a (0) |
| deepseek-v4-pro | 47.7% (308) | 37.1% (410) | 12.5% (8) * | n/a (0) |
| gemini-3.1-pro | 16.7% (6) * | 43.3% (473) | 38.3% (193) | n/a (0) |
| gpt-5.6-sol | 36.4% (11) | 39.0% (428) | 42.0% (269) | 0.0% (1) * |
| grok-4.6 | 43.2% (199) | 37.2% (317) | 0.0% (1) * | n/a (0) |
| qwen-3.8-max | 43.5% (386) | 35.7% (308) | 0.0% (1) * | 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 41.7% (n=12); deepseek-v4-pro 100.0% (n=1) *; grok-4.6 100.0% (n=1) *; gemini-3.1-pro 100.0% (n=1) *. Mid-scale ranking: 1 of the 6 models with N>=10 in both 50-60 and 60-70 out-hit 50-60 from 60-70 this week (issue #5: 0 of 5).

## Hit-rate by forecast horizon (FH)

| Model | 1h | 4h | 1d |
|---|---|---|---|
| claude-fable-5 | 41.1% (387) | 30.5% (203) | 20.6% (34) |
| claude-opus-5 | 42.1% (299) | 29.9% (177) | 20.6% (34) |
| deepseek-v4-pro | 45.3% (439) | 39.7% (237) | 15.7% (51) |
| gemini-3.1-pro | 46.0% (437) | 36.3% (201) | 20.0% (35) |
| gpt-5.6-sol | 44.5% (434) | 34.9% (235) | 22.5% (40) |
| grok-4.6 | 42.5% (327) | 36.1% (169) | 22.7% (22) |
| qwen-3.8-max | 43.7% (439) | 36.8% (231) | 16.2% (37) |

n in parentheses; * = N<10 (insufficient). 0 of 7 model lines hit better on 1d than on 1h this window (issue #5: 1 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 | 518 | 32.4% | -$114.37 | -$62.57 | -$114.37 |
| gemini-3.1-pro | 673 | 33.7% | -$131.95 | -$64.65 | -$134.16 |
| claude-opus-5 | 510 | 29.0% | -$137.47 | -$86.47 | -$137.47 |
| deepseek-v4-pro | 727 | 34.2% | -$152.22 | -$79.52 | -$152.76 |
| qwen-3.8-max | 707 | 32.7% | -$158.82 | -$88.12 | -$159.24 |
| claude-fable-5 | 624 | 29.3% | -$160.85 | -$98.45 | -$160.85 |
| gpt-5.6-sol | 709 | 32.3% | -$163.27 | -$92.37 | -$163.27 |

0 of 7 model lines finished net-positive and 0 gross-positive this week (issue #5: 0 of 7 net-positive, 0 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: a down week on a quieter tape

> **MARKET CHECK · OBSERVATION (two adjacent calendar weeks)**
>
> OBSERVATION — mean |1d move| 2.10% -> 1.35% (-36% rel), BTC realized vol 35.2% -> 31.4% (ann., hourly); field directional accuracy 42.8% -> 39.6% (-3.3pp), 1 of 7 models improved, sim win-rate up for 1 of 7, field sim PnL -$669.03 -> -$1018.93 (adjacent calendar weeks Aug 31-Sep 6 vs Sep 7-13; descriptive, one pair of weeks, not a claim).
>
> Robustness: raw price-sign accuracy 51.6% -> 38.7% — the same direction as the trade-based hit rule.

| Measure | Aug 31-Sep 6 (week A) | Sep 7-13 (week B) | Change |
|---|---|---|---|
| Mean \|1d move\| (5 assets) | 2.10% | 1.35% | -36% rel |
| BTC realized vol (ann., hourly) | 35.2% | 31.4% | -3.8pp |
| Field directional accuracy | 42.8% | 39.6% | -3.3pp |
| Models improving hit-rate | -- | 1 of 7 | -- |
| Raw price-sign accuracy | 51.6% | 38.7% | -12.8pp |
| Field sim win-rate | 35.6% | 32.1% | -3.5pp |
| Field sim net PnL | -$669.03 | -$1018.93 | -- |

Week A is exactly the issue-#5 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, two windows back, inside issue #4's window). Week A (2026-08-31..2026-09-06) is pure grok-4.6 (it reproduces the published issue #5); 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 +22.5pp this week against +19.6pp in issue #5.
- Gap and regime move together. The field hit 39.6% against 42.8% in issue #5 and stated mean confidence read 62.1 against 62.4, so the gap held at +22.5pp: a field that hits 39.6% instead of 42.8% 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 four 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 4,468 scored calls sit inside one market regime. This window ran 2 trend days of 7 (issue #5: 3 of 7); every comparison with issue #5 is a comparison across regimes as well as across weeks. Six 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 6): (1) rolling-1w forecasts emit daily since Aug 22, 2026 and rolling-1M daily since Aug 23, 2026, so inside this window (Sep 7-13) daily coverage is FULL, as it was in issues #4 and #5: 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, two 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-#5 window), so every grok number in this issue and in issue #5 is a pure grok-4.6 line -- the grok week-over-week row is the first 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 #5 after the audit that found the raw table mixes two exchanges; issue #5 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 14 16:00 UTC) -- the definition pinned in issue #3 and carried by issues #4 and #5, which printed 42,905 at the Sep 7 16:00 cutoff. The pack reproduces that pin at the previous cutoff (control OK), so this issue's 47,373 is an additive step under one definition; counter deltas against issue #4 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 | 4,468 |
| Out of gate (1w / 1M) | 490 / 490 | -- of them sideways | 4,833 |
| Invalid | 9 | Pending (next issue) | 0 |
| Late closes | 0 | Uptime, all grids | 266 / 266 slots |
| Source file | weekly_metrics_2026-09-07.json | Report cutoff | Mon Sep 14, 2026, 16:00 UTC |
| Generated at (pipeline) | 2026-09-15 15:24 UTC | Methodology | v1.1 (2026-08-10) · hash e66c7e8c864a2233 |

**THIS REPORT: 4,468 scored observations.** MARKETMANIA RESEARCH TO DATE (as of Sep 14, 2026 cutoff): 47,373 directional forecasts resolved since Jul 11 · 7 models tracked (current line-up; earlier versions folded into their successors' lineage) · 5 assets · 5 horizons · hourly · 28 published reports.

> **Issue #6.** Weekly Calibration is a living series. Issue #5's open questions -- does the field gap hold near +19.6pp for a third issue or track the base rate one-for-one, and does the best-Brier line change hands again? -- read this issue as: field gap +19.6pp -> +22.5pp (1 of 7 lines narrowed) on a field base that moved -3.3pp, and the best-Brier line did not change hands: qwen-3.8-max holds it two issues running (0.2794, issue #5: qwen-3.8-max 0.2755). Also on the record: gpt-5.6-sol's 70-80 bucket read 42.0% (n=269) against its own 60-70 at 39.0% (n=428). The grok row is pure grok-4.6 this issue and was pure grok-4.6 in issue #5; the 4.5 -> 4.6 flip (Aug 24, 2026 09:22 UTC) sits two 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: the gap moved +2.9pp while the field base moved -3.3pp -- does the gap keep tracking the base one-for-one, or does it hold a level of its own?
- Does qwen-3.8-max hold the best-Brier line for a third issue, or does it change hands?
- Monthly series: is the overconfidence gap stable across market regimes (trend vs flat) at monthly n?

## Related research

| Report | Direct PDF link |
|---|---|
| Consensus Watch #6 | https://marketmania.ai/research/reports/consensus-watch-2026-09-07.pdf |
| Weekly Model Watch #6 | https://marketmania.ai/research/reports/model-watch-2026-09-07.pdf |
| Weekly Calibration #5 | https://marketmania.ai/research/reports/weekly-calibration-2026-08-31.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_2026w37,
  title  = {Weekly Calibration #6: confidence vs. direction-hit, Sep 7-13 2026},
  author = {{MarketMania Research}},
  year   = {2026}, month = {September}, day = {15},
  url    = {https://marketmania.ai/research/reports/weekly-calibration-2026-09-07.pdf},
  note   = {Methodology v1.1, hash e66c7e8c864a2233; source weekly_metrics_2026-09-07.json}
}
```