# Weekly Calibration #7

**WEEKLY · CALIBRATION** · September 22, 2026 · MarketMania Research · Weekly series

Window: **Sep 14-20, 2026 UTC** · Cutoff: **Mon Sep 21, 2026, 16:00 UTC** · Methodology **v1.1 (2026-08-10)**, hash `e66c7e8c864a2233`

PDF: https://marketmania.ai/research/reports/weekly-calibration-2026-09-14.pdf · Open data (JSON): https://marketmania.ai/research/reports/weekly-calibration-2026-09-14.json

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

## Research Snapshot

| RESEARCH SNAPSHOT |  |  |  |
|---|---|---|---|
| **5,598** directional calls scored | **7** models (stable lineup) | of **9,308** mature forecasts | window **Sep 14-20, 2026 UTC** |
| hit = direction rule: tp1/tp2->hit, sl->miss, expiry->sign of gross pnl | FH gate **1h / 4h / 1d** | base field hit **50.7%** vs **62.8** stated | cutoff **Mon Sep 21, 2026, 16:00 UTC** |
| market: BTC net **+5.64%** (prior -4.36%) | ann vol **51.0%** (was 19.7%) | TOP5 volume **$18.74B**, +17% w/w | pairwise corr **0.93** (was 0.74) |

## Key finding

> **KEY FINDING · OBSERVATION (one weekly window)**
>
> ### The field's overconfidence gap moved +22.5pp -> +12.1pp week over week, 7 of 7 model lines narrowed, and field Brier went 0.2948 -> 0.2669.

## TL;DR

- **OBSERVATION -- The gap read +12.1pp this week.** Directional calls hit 50.7% against 62.8 stated mean confidence (issue #6: +22.5pp on 39.6% vs 62.1). Field Brier is 0.2669 (was 0.2948) and 0 of 7 model lines scored below 0.25, the score of an uninformative always-50% predictor. One window, descriptive.
- **7 of 7 models narrowed their gap.** Gap moves ran from claude-opus-5's -14.6pp (+24.5pp -> +9.9pp) to deepseek-v4-pro's -6.0pp (+18.4pp -> +12.4pp); 7 of 7 model lines narrowed week over week. claude-fable-5 is the best-calibrated line (Brier 0.2593); gemini-3.1-pro leads on hit-rate (52.9%).
- **The high-confidence buckets, N>=10 only.** deepseek-v4-pro 70-80 42.3% (n=26), gemini-3.1-pro 70-80 52.9% (n=342), gpt-5.6-sol 70-80 44.9% (n=267). 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 -$30.12 (claude-opus-5, best) to -$79.39 (deepseek-v4-pro, worst), field -$337.72 against issue #6's -$1018.93.

### 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 three windows back. "Was" values are the numbers published in issue #6 for the back-to-back window Sep 7-13.

## 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 #7 the series has seven 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 three 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 second 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 |
|---|---|---|---|---|---|---|---|
| claude-fable-5 | 799 | 60.1% | 52.3% | 61.4 | +9.1 | 0.2593 | 48.9%-55.8% |
| claude-opus-5 | 711 | 53.5% | 51.6% | 61.5 | +9.9 | 0.2603 | 47.9%-55.3% |
| qwen-3.8-max | 819 | 61.6% | 50.8% | 60.0 | +9.2 | 0.2604 | 47.4%-54.2% |
| grok-4.6 | 629 | 47.3% | 49.6% | 60.3 | +10.7 | 0.2620 | 45.7%-53.5% |
| deepseek-v4-pro | 936 | 70.4% | 48.4% | 60.8 | +12.4 | 0.2660 | 45.2%-51.6% |
| gemini-3.1-pro | 883 | 66.4% | 52.9% | 66.8 | +14.0 | 0.2701 | 49.6%-56.2% |
| gpt-5.6-sol | 821 | 61.7% | 49.7% | 68.1 | +18.4 | 0.2876 | 46.3%-53.1% |
| **Field (all models)** | **5,598** | - | **50.7%** | **62.8** | **+12.1** | **0.2669** | - |

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 three 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 second 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 | 54.3% (186) | 51.7% (613) | n/a (0) | n/a (0) |
| claude-opus-5 | 53.8% (117) | 51.2% (594) | n/a (0) | n/a (0) |
| deepseek-v4-pro | 47.2% (288) | 49.2% (622) | 42.3% (26) | n/a (0) |
| gemini-3.1-pro | 33.3% (6) * | 53.2% (530) | 52.9% (342) | 40.0% (5) * |
| gpt-5.6-sol | 75.0% (4) * | 51.9% (547) | 44.9% (267) | 33.3% (3) * |
| grok-4.6 | 49.1% (216) | 50.0% (412) | 0.0% (1) * | n/a (0) |
| qwen-3.8-max | 51.3% (349) | 50.5% (461) | 37.5% (8) * | 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 100.0% (n=1) *. Mid-scale ranking: 2 of the 5 models with N>=10 in both 50-60 and 60-70 out-hit 50-60 from 60-70 this week (issue #6: 1 of 6).

## Hit-rate by forecast horizon (FH)

| Model | 1h | 4h | 1d |
|---|---|---|---|
| claude-fable-5 | 53.2% (496) | 54.5% (253) | 32.0% (50) |
| claude-opus-5 | 51.2% (422) | 56.1% (244) | 31.1% (45) |
| deepseek-v4-pro | 48.8% (586) | 50.5% (297) | 32.1% (53) |
| gemini-3.1-pro | 53.8% (561) | 52.9% (272) | 42.0% (50) |
| gpt-5.6-sol | 49.7% (505) | 52.6% (266) | 34.0% (50) |
| grok-4.6 | 49.1% (385) | 53.2% (205) | 35.9% (39) |
| qwen-3.8-max | 51.9% (505) | 51.7% (269) | 33.3% (45) |

n in parentheses; * = N<10 (insufficient). 0 of 7 model lines hit better on 1d than on 1h this window (issue #6: 0 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 |
|---|---|---|---|---|---|
| claude-opus-5 | 711 | 43.5% | -$30.12 | +$40.98 | -$60.18 |
| claude-fable-5 | 799 | 45.2% | -$31.40 | +$48.50 | -$63.40 |
| gemini-3.1-pro | 883 | 44.4% | -$34.19 | +$54.11 | -$77.86 |
| grok-4.6 | 629 | 42.1% | -$44.11 | +$18.79 | -$54.55 |
| qwen-3.8-max | 819 | 43.4% | -$56.30 | +$25.60 | -$63.65 |
| gpt-5.6-sol | 821 | 42.5% | -$62.22 | +$19.88 | -$92.02 |
| deepseek-v4-pro | 936 | 42.4% | -$79.39 | +$14.21 | -$80.25 |

0 of 7 model lines finished net-positive and 7 gross-positive this week (issue #6: 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: an up week on a louder tape

> **MARKET CHECK · OBSERVATION (two adjacent calendar weeks)**
>
> OBSERVATION — mean |1d move| 1.35% -> 2.69% (+100% rel), BTC realized vol 31.4% -> 35.1% (ann., hourly); field directional accuracy 39.6% -> 50.7% (+11.2pp), 7 of 7 models improved, sim win-rate up for 7 of 7, field sim PnL -$1018.93 -> -$337.72 (adjacent calendar weeks Sep 7-13 vs Sep 14-20; descriptive, one pair of weeks, not a claim).
>
> Robustness: raw price-sign accuracy 38.7% -> 54.2% — the same direction as the trade-based hit rule.

| Measure | Sep 7-13 (week A) | Sep 14-20 (week B) | Change |
|---|---|---|---|
| Mean \|1d move\| (5 assets) | 1.35% | 2.69% | +100% rel |
| BTC realized vol (ann., hourly) | 31.4% | 35.1% | +3.7pp |
| Field directional accuracy | 39.6% | 50.7% | +11.2pp |
| Models improving hit-rate | -- | 7 of 7 | -- |
| Raw price-sign accuracy | 38.7% | 54.2% | +15.4pp |
| Field sim win-rate | 32.1% | 43.4% | +11.2pp |
| Field sim net PnL | -$1018.93 | -$337.72 | -- |

Week A is exactly the issue-#6 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, three windows back, inside issue #4's window). Week A (2026-09-07..2026-09-13) is pure grok-4.6 (it reproduces the published issue #6); 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 +12.1pp this week against +22.5pp in issue #6.
- Gap and regime move together. The field hit 50.7% against 39.6% in issue #6 and stated mean confidence read 62.8 against 62.1, so the gap came out at +12.1pp: a field that hits 50.7% instead of 39.6% 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 five 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 5,598 scored calls sit inside one market regime. This window ran 3 trend days of 7 (issue #6: 2 of 7); every comparison with issue #6 is a comparison across regimes as well as across weeks. Seven 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 7): (1) rolling-1w forecasts emit daily since Aug 22, 2026 and rolling-1M daily since Aug 23, 2026, so inside this window (Sep 14-20) daily coverage is FULL, as it was in issues #4, #5 and #6: 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 485). (2) The grok line flipped 4.5 -> 4.6 at Aug 24, 2026 09:22 UTC, three 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-#6 window), so every grok number in this issue and in issue #6 is a pure grok-4.6 line -- the grok week-over-week row is the second 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 #6 after the audit that found the raw table mixes two exchanges; issue #6 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 21 16:00 UTC) -- the definition pinned in issue #3 and carried by issues #4, #5 and #6, which printed 47,373 at the Sep 14 16:00 cutoff. The pack reproduces that pin at the previous cutoff (control OK), so this issue's 52,971 is an additive step under one definition; counter deltas against issue #5 and earlier remain definitional.

## Counters & lineage

| Counter | Value | Counter | Value |
|---|---|---|---|
| Forecasts total | 10,290 | Mature (scored pool) | 9,308 |
| OK in gate | 9,308 | -- of them directional | 5,598 |
| Out of gate (1w / 1M) | 488 / 485 | -- of them sideways | 3,710 |
| Invalid | 9 | Pending (next issue) | 0 |
| Late closes | 0 | Uptime, all grids | 266 / 266 slots |
| Source file | weekly_metrics_2026-09-14.json | Report cutoff | Mon Sep 21, 2026, 16:00 UTC |
| Generated at (pipeline) | 2026-09-22 06:56 UTC | Methodology | v1.1 (2026-08-10) · hash e66c7e8c864a2233 |

**THIS REPORT: 5,598 scored observations.** MARKETMANIA RESEARCH TO DATE (as of Sep 21, 2026 cutoff): 52,971 directional forecasts resolved since Jul 11 · 7 models tracked (current line-up; earlier versions folded into their successors' lineage) · 5 assets · 5 horizons · hourly · 32 published reports.

> **Issue #7.** Weekly Calibration is a living series. Issue #6's open questions -- does the gap keep tracking the field base one-for-one or hold a level of its own, and does qwen-3.8-max hold the best-Brier line for a third issue? -- read this issue as: field gap +22.5pp -> +12.1pp (7 of 7 lines narrowed) on a field base that moved +11.2pp, and the best-Brier line changed hands: claude-fable-5 at 0.2593 (issue #6: qwen-3.8-max at 0.2794). Also on the record: gpt-5.6-sol's 70-80 bucket read 44.9% (n=267) against its own 60-70 at 51.9% (n=547). The grok row is pure grok-4.6 this issue and was pure grok-4.6 in issue #6; the 4.5 -> 4.6 flip (Aug 24, 2026 09:22 UTC) sits three 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 -10.4pp while the field base moved +11.2pp -- the second issue running in which the two move nearly one-for-one in opposite directions; does that hold for a third?
- The best-Brier line changed hands to claude-fable-5 -- does it hold for a second issue, or change hands again?
- Monthly series: is the overconfidence gap stable across market regimes (trend vs flat) at monthly n?

## Related research

| Report | Direct PDF link |
|---|---|
| Consensus Watch #7 | https://marketmania.ai/research/reports/consensus-watch-2026-09-14.pdf |
| Weekly Model Watch #7 | https://marketmania.ai/research/reports/model-watch-2026-09-14.pdf |
| Weekly Calibration #6 | https://marketmania.ai/research/reports/weekly-calibration-2026-09-07.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_2026w38,
  title  = {Weekly Calibration #7: confidence vs. direction-hit, Sep 14-20 2026},
  author = {{MarketMania Research}},
  year   = {2026}, month = {September}, day = {22},
  url    = {https://marketmania.ai/research/reports/weekly-calibration-2026-09-14.pdf},
  note   = {Methodology v1.1, hash e66c7e8c864a2233; source weekly_metrics_2026-09-14.json}
}
```