CHANGELOG.md

All notable changes to this research repository are recorded here.

The format follows a light research-change convention:

[phase] date — description

Goals

Add frozen E1–E8 pre-registered experiment runner (scripts/run_e1_e8.py).
Execute the frozen E1–E8 protocol on the synthetic benchmark dataset.
Produce reproducible evidence artifacts (git-ignored results/).

Executed (2026-08-24, pre-experiment commit `3fbd268`, config hash `ef1e3fd5...`)

Ran E1–E8 zero-optimization measurement on the frozen synthetic 3000-bar XAUUSD M15 series.
Artifacts remain under results/ (E1-E8_MANIFEST.json, E1-E8_RESULTS.csv, E1-E8_SUMMARY.md, per-experiment JSON/CSV, cost-sensitivity and segment tables).

Result status (abbreviated)

ARIMA vs naive: FAIL (MASE 1.078, negative net expectancy).
SAX vs naive: FAIL (MASE 1.353, neg net).
Hybrid vs ARIMA / vs SAX: degenerate paired diff (direction shared), aggregate net negative.
Agreement vs disagreement: inconclusive (both non-positive).
Robustness: not stable across 3 chrono segments.
Scientific conclusion on synthetic benchmark: INCONCLUSIVE (no incremental information shown).

Goal of this change: build the reproducible ARIMA-vs-SAX-vs-Hybrid research harness with strict chronological validation and no-lookahead guarantees.

Added

Repository governance and protocol documents (README, RESEARCH_PROTOCOL, ARCHITECTURE, VALIDATION_PROTOCOL, LICENSE).
Initial Python source tree (research harness, no live trading):
- common target (Forward Return / ATR), forecast record interface
- baselines (naive + drift)
- SAX component (z-norm, PAA, SAX encoding, analog search, MINDIST)
- ARIMA component (statsmodels-backed, configurable (p,d,q))
- hybrid evidence layer (deterministic state classification)
- data integrity checks
- forecast/distribution/economic metric evaluators
- chronological walk-forward validation driver
- pipeline orchestration -> structured records
Initial test tree (unit, no-lookahead, integration).
Default frozen configuration (configs/default.json).
Research log schema (see src/forecasting/record.py).

Notes

No performance claims are made.
No live trading / execution code is included at this stage.
No parameter optimization was performed.

2.3 KiB Raw Permalink Blame History