SniperGold_ML/docs/P3_S24_NEXT_RESEARCH_PRIORITY.md

# P3-S24 — NEXT RESEARCH PRIORITY

```text
Date       : 2026-08-26
Session    : P3-S24 — RESEARCH VALUE ASSESSMENT (ANALYSIS ONLY)
Result     : D — CURRENT EDGE NOT DEMONSTRATED
Priority   : 1) EXTERNAL DATA VALIDATION  2) DATASET EXPANSION (power targets)
            3) MECHANISM WORK (deferred)   4) ADVANCED ML (deferred)
Next       : P3-S25 = NOT STARTED (owner authorization required)
```

## Executive summary

The corrected-population experiments (P3-S22.3 → P3-S23) show that the weak
logistic ranking signal observed on the frozen population did NOT survive the
correction of the research M30 construction (pooled OOS ROC-AUC 0.5792 →
0.5235; 95% CI [0.450, 0.592] includes 0.5; fold 3 collapses to 0.464;
LogLoss/Brier worse than the constant prior). The record therefore does not
demonstrate a stable edge for the current setup definition. The highest-value
next investment is NOT a better model — it is **validating the data basis
independently, then expanding the validated sample to the computed power
targets**.

## Highest-value next action — Option 2: EXTERNAL DATA VALIDATION

**What:** independently validate the broker-feed / UTC-clock M30 basis-parity
of the corrected research construction and the corrected Candidate Setup
population (707 in-scope; 610 leads) using external or second-source tick/M15
data (Tickstory / Dukascopy / independent broker feed at MQL5/Files).

**Why (evidence):**
- P3-S22.3 replaced the research M30 construction with a fixed UTC 30-min grid,
  changing ~19.5% of gate bars and 202/223 setups — a material population
  factor. This is a *self-feed* (same M15 source) correction; its basis-parity
  vs the broker/terminal feed was **inherited, not re-established** (S23 §G).
- It is the only step that can distinguish "genuine absence of edge" (**D**)
  from "residual feed/construction artifact" (**E**) — the two states that the
  current record cannot separate.
- External validation also stress-tests population robustness: if the corrected
  construction on an independent feed reproduces the same 707/610/579
  accounting, the basis is confirmed; if not, a new E-type factor is exposed
  before any further modelling.
- Cost is LOW–MEDIUM; information gain is HIGH.

**Governance:** NOT imported or used in P3-S24 (analysis-only). Requires a
separate, explicit owner authorization to import external data.

## Second action — Option 1: DATASET EXPANSION (to power targets)

Only **after** the basis is validated. Sample power analysis (Hanley-McNeil,
α=0.05):

| Target AUC | total N @ current 1:2.3 ratio (80%/90% power) | total N balanced |
|---|---|---|
| 0.55 | ≈ 1,261 / 1,687 | ≈ 1,036 / 1,386 |
| 0.58 | ≈ 493 / 658 | ≈ 398 / 532 |
| 0.60 | ≈ 314 / 420 | ≈ 250 / 336 |

Practical targets: **≈1,300–1,700 independent leads** to resolve a 0.55-level
edge; **≈500–660** for 0.58; **≈1,700** for a ±0.03 CI near noise. The current
579-row population is roughly ½–⅓ of the 0.55 requirement.

## Deferred options (rejected for now, with reasons)

| Option | Verdict | Reason |
|---|---|---|
| 3 — Setup mechanism redesign | DEFERRED | No structural family heterogeneity on the corrected population (direction/zone/CHoCH/liquidity/volatility splits all p>0.2); only temporal year drift is significant (p=0.024). Redesign is premature without a replicable basis signal. |
| 4 — Advanced ML | DEFERRED | Nonlinear models overfit on this sample (P3-S18), calibration added nothing (P3-S21.2), and the corrected baseline contains no demonstrated signal (P3-S23). Adding capacity now risks fitting feed noise. |
| A — dataset-only classification | REJECTED as primary | Sample size is a second-order limit, but the corrected point (0.5235) is below the detectable threshold; more rows on an unvalidated, edge-free basis will not create information. |
| C — label/contract review | REJECTED | Label contract v1 semantically validated; censoring ≈3%, ambiguity ≈1% — not the limiting factor. |

## Stop condition

```text
P3-S25 = NOT STARTED.
No calibration, nonlinear model, feature/label/TP-SL change, external-data
import, or production change is begun or authorized by this phase.
An explicit owner decision authorizing (1) external data validation and/or
(2) dataset expansion is required before any next phase.
```

*End of P3-S24 next research priority.*