ARIMA_SAX_Hybrid_Forecaster/results/R2-H/R2_H_PROTOCOL.md

R2-H — HORIZON-BOUNDARY EXPERIMENT (FROZEN PROTOCOL)

Experiment: R2-H Protocol version: r2-h-v0.1.0-frozen Status: DESIGN FROZEN — EXECUTION NOT AUTHORIZED. Branch: main Anchor commit (R1-R, immutable): 67be79aa75ae98eace7c67b74c93797f4327d51a Freeze session: governance-controlled formal design-freeze (2026-08-26)

This document is a formal, self-contained, frozen protocol. It defines the R2-H research family design. It freezes origin lists, model policy, cost model, metrics, inference, multiplicity control, no-lookahead gates, stop conditions, and the outcome matrix. No forecasting model is executed, no performance is computed, and no prediction/evidence artifact is produced in this session. Execution requires a separate, explicit EXECUTION AUTHORIZATION gate.

---

A. Governance status

• Owner decision: FORMAL R2-H FREEZE AUTHORIZED (2026-08-26).
• Scope authorized: checkpoint identity resolution; protocol authoring; protocol consistency checks; creation of R2-H design artifacts; committing the frozen protocol and associated governance documentation.
• Explicitly NOT authorized: experiment execution; model fitting; forecasting; backtesting; parameter optimization; statistical inference from newly generated experimental results; economic evaluation; creation of prediction/evidence artifacts.
• The repository checkpoint discrepancy is resolved: authoritative HEAD is 67be79aa75ae98eace7c67b74c93797f4327d51a (commit research: finalize R1-R robustness synthesis). All other quoted hashes are transcription errors and are invalid checkpoints.
• This session creates ONLY the formal R2-H design artifacts necessary for the freeze (origin generator, frozen origin lists, protocol, manifest). No other repository content is modified.

B. R1-R anchor

R1-R is a CLOSED and IMMUTABLE scientific anchor. Within its defined scope it established, on XAUUSDc M1/M5/M15:

• approximately 15-minute elapsed directional forecasting on XAUUSD;
• no useful gross directional signal;
• direction accuracy approximately 0.48–0.51;
• no non-naive model beat the naive baseline on the tested target;
• gross return approximately zero and mixed-sign;
• economics decisively negative (cost ≈ 0.8–3.4 ATR/trade vs gross ≈ 0–0.15 ATR/trade);
• no-lookahead, forensic, tick-validation, and reproducibility gates passed;
• deterministic origin selection repaired the historical R1-F.8 provenance problem.

R2-H does not reopen, retune, reinterpret, or amend R1-R. R1-R evidence, protocol, origin lists, predictions, and manifests remain untouched.

C. Research question

Does the R1-R approximately 15-minute directional null persist, weaken, or invert at materially longer elapsed prediction horizons on XAUUSD when transaction-cost scaling is explicitly accounted for?

The experiment is not framed as a search for a positive result. Valid outcomes are: NULL, BOUNDARY, PARTIAL / ECONOMICALLY NOT VIABLE, INCONCLUSIVE, INVALID.

D. Hypotheses

• Global null (FWER family {60, 240}): At the 60-minute and 240-minute elapsed horizons, the Drift model exhibits no gross directional signal (direction accuracy ≤ 0.50), no positive gross return per trade, and no error improvement over the naive baseline.
• Global alternative: At least one member of the FWER family exhibits a gross directional signal that satisfies all three Q1 conditions (A, B, C).
• Economic hypothesis: Even if a gross signal exists, the net return under the frozen 7.57 bps round-trip cost has a dependence-aware 95% CI whose lower bound is ≤ 0 (Q2 fails → PARTIAL / ECONOMICALLY NOT VIABLE).
• The 15-minute H0 horizon is an in-study replication/control of R1-R and provides no independent evidence for the new horizon-boundary claim.

E. Horizon grid

label	horizon (minutes)	base grid	role	FWER family
H0	15	M1	in-study R1-R control	outside (no independent claim)
H1	60	M1	primary	in family
H2	240	M1	primary	in family
H9	960	M1	diagnostic-only	never enters family

Rules:

• H0/H1/H2 are the ONLY horizons in the primary multiplicity-controlled family.
• 960 minutes is DIAGNOSTIC-ONLY. It must NOT enter the FWER family, determine success/failure, qualify as a boundary, justify reopening, or be adaptively promoted after results are observed.
• No horizon may be added, removed, or substituted.

F. Target definition

All horizons use the M1 base grid. The horizon is defined in ELAPSED TIME, not bar count.

tau_H(o) = [ P(t_end) - P(o) ] / ATR20(o)

t_end = first valid M1 bar timestamp >= timestamp(o) + H minutes
P(x)  = close price of M1 bar x
ATR20(o) = R1-compatible M1 close-only ATR(20) at origin o

• ATR20 is the deterministic close-change-based ATR proxy retained for cross-family comparability: simple rolling mean of |close[i] − close[i−1]| over the trailing 20 M1 bars (see src/forecasting/target.py, rolling_atr with high/low absent). It is NOT a Wilder ATR. This label is frozen; no silent ATR methodology change is permitted.

G. Data provenance

• Source: results/R1_real_data/XAUUSDc_M1_raw.json (MT5 terminal history, broker symbol XAUUSDc, Gold Spot Cent; M1; 397,558 bars; 2025-06-02 01:00 → 2026-08-24 15:19 server trade time).
• Fresh SHA-256 verification (R2-H provenance, computed in the freeze session):

file: results/R1_real_data/XAUUSDc_M1_raw.json
sha256: 5472357e67f70f7d1bfef8f9de67fcce91ee6f26208072d7a6766b06c158ee6e

• This hash matches the R1 manifest record (R1_REAL_DATASET_MANIFEST.json, dataset_hash_raw_json) but is re-verified here as new R2-H evidence; it is not silently reused. The R1-R close-only hash 80e4b52b0df0c6e348ce2002348279eee9f6e8a61c3109efe82fe475ed3cbc19 remains an R1-R property and is not re-used as R2-H evidence.
• The M1 file is used directly; no higher-timeframe aggregation is performed in R2-H.

H. Origin-selection rule (frozen, deterministic, model-independent)

Generator: scripts/run_r2h_origins.py (design artifact committed with this protocol). Parameters (frozen):

parameter	value
base grid	M1
warmup	2000 M1 bars (sufficient for ARIMA fit window 300, SAX window 24, drift lookback 40, ATR 20)
endpoint rule	t_end = first valid M1 bar timestamp ≥ timestamp(o) + H
gap tolerance	5 minutes — any inter-bar delta > 5 min within (o, t_end] excludes the origin
origin spacing	2 × H elapsed minutes between accepted origins
overlap	none (spacing 2×H with window H ⇒ non-overlapping target windows)
cap	none (no short-horizon cap applied)
selection	no human-selected origins; no model-dependent origins

Gap/session policy (audited, frozen):

• An origin is usable only when the required elapsed-time target can be resolved under the frozen gap rule.
• Any materially missing bar / gap that violates the frozen tolerance (inter-bar delta > 5 minutes within (o, t_end]) excludes the origin.
• Weekend/session boundaries are never silently converted into ordinary contiguous observations.
• All exclusions are deterministic and recorded (see manifest).
• No manual origin removal is permitted.

Generated origin lists (frozen, generated twice, byte-identical):

file	horizon	origins	SHA-256
origins_H015.csv	15 min	13,193	fb5a36e93c8f9c7b0fee638d35d551778f1a27cbc50e0727556cf6a1d8c5c7df
origins_H060.csv	60 min	3,169	e7b5132c321f5350ec1e231d49d68dbc29468cdf9ee6a5014a2764aaf36eb9af
origins_H240.csv	240 min	863	3ad65e94f46731e65b8d40cf2b7746a6c0bd2eafe43216178d9e6e6c6ebcb910

Two-run identity: the generator was run twice into separate directories; all origin CSVs and the params manifest are byte-identical between the runs (identical ordering, identical LF newline convention, identical contents, identical SHA-256). Verified with fc /b and SHA-256. Mismatch = HARD STOP — none occurred.

Recorded runtime parameters (per §9): warmup=2000; H ∈ {15, 60, 240}; step (spacing)=2×H elapsed minutes; cap=none; gap rule tolerance=5 minutes; dataset identity=results/R1_real_data/XAUUSDc_M1_raw.json (sha256 5472357e…); software/config identity = generator file scripts/run_r2h_origins.py, Python 3.14.6.

I. Model policy

• PRIMARY MODEL (frozen): Drift — the sole model used for the primary Q1/Q2 decision. No per-horizon "best model" selection exists.
• Secondary robustness/diagnostic models (never selected retrospectively as "best"): ARIMA(1,0,0) (rolling fit window 300), SAX (word 8, alphabet 5, window 24, top_k 10, min_analogs 5, neutral 0.25), Hybrid.
• No model-shopping is permitted. No model may be chosen based on observed results. Secondary models cannot establish the primary boundary claim.
• Drift definition (existing, unchanged): linear slope over the last 40 closes projected H bars ahead, normalized by ATR(20) (src/baselines.py, DriftBaseline). Naive baseline = zero-return forecast (NaiveBaseline), used as the error reference.

J. Cost model (frozen, pre-registered)

• Headline round-trip cost: 7.57 bps
  • median tick spread ≈ 1.57 bps (R1-R.4V tick validation, 103.3M ticks)
  • commission assumption = 6.00 bps
  • total frozen cost = 7.57 bps
• The 7.57 bps value is a PRE-REGISTERED GLOBAL CONSTANT. It is applied identically to every origin and every primary horizon.
• Conversion to entry-ATR units:

cost_atr(o) = (7.57 / 10000) × close(o) / ATR20(o)

• Prohibitions: no future-looking or per-origin future-window spreads; no spread derived from ticks after the origin; no cost tuning after results.
• Sensitivity: 0.5× and 2× cost multiples may be reported descriptively only; they cannot alter the primary decision.

K. Primary metrics

Primary directional metrics (computed on frozen origin lists, Drift model):

1. Direction accuracy (fraction of origins where the Drift direction sign matches the realized sign of tau_H).
2. Mean gross return in ATR units (mean of sign-matched realized tau_H).
3. Mean net return in ATR units (mean gross − cost_atr).
4. Time-normalized / holding-period economic return (gross and net per unit of elapsed holding time).

Secondary metrics (diagnostic only): MASE (vs naive); gross/cost ratio; cost/range ratio; trade coverage; number of origins; effective sample size; turnover / trade frequency; annualized or time-normalized economics; chronological EARLY / MIDDLE / LATE descriptive results.

Reporting rules:

• Gross and net MUST always be reported separately.
• Longer horizons must NOT appear superior merely because they trade less frequently — time-normalized economics are reported alongside per-trade economics.

L. Statistical inference (frozen)

• Method: Moving Block Bootstrap (MBB).
• Block length: B = 60 M1 observations.
• Bootstrap repetitions: 10,000.
• Seed: 20260826.
• Confidence level: 95% (percentile CI).
• Effective sample size: n_eff = floor(n_usable / B); the actual n and n_eff are published for every horizon.
• Dependence-aware CIs are used for direction accuracy, gross return/trade, net return/trade, and the error comparison vs naive.
• These parameters are fixed before execution. B, seed, and replication count may not be modified after observing experimental results.
• A horizon that is statistically underpowered for the frozen decision rule is classified INCONCLUSIVE. No significance is manufactured from a low-n horizon.

Expected scale (audited, acknowledged, not a guarantee):

• H15: high usable-origin count (13,193 origins; n_eff ≈ 219).
• H60: moderate effective sample (3,169 origins; n_eff ≈ 52).
• H240: low effective sample (863 origins; n_eff ≈ 14).

H240 must not be claimed to have the same statistical power as H15; an underpowered H240 result cannot establish a positive boundary. Power limitations are reported as scientific results, not hidden caveats.

M. FWER procedure (multiple-horizon inference)

• Primary FWER family: {60, 240}.
• H0 (15 min) is the in-study R1-R control and is not treated as independent evidence for the new horizon-boundary FWER claim.
• Method: joint cross-horizon block bootstrap that preserves chronological dependence and cross-horizon dependence (origins aligned on a common M1/elapsed-time axis; blocks of 60 M1 observations resampled jointly across horizons).
• Statistic: max-statistic under a re-centered global null (each horizon's statistic re-centered to its null mean; the maximum over the family is the test statistic). A simple independent per-horizon z-max approximation is not used.
• Family-wise α = 0.05.
• Per-horizon interpretation is permitted only after the global gate is rejected.
• No horizon shopping; no post-hoc multiplicity changes.

N. No-lookahead / forensic gates (frozen)

Gates F1–F9 (each must distinguish OBSERVED / INFERRED / UNKNOWN / DECISION):

gate	item
F1	feature construction
F2	target construction
F3	ATR construction
F4	SAX window construction
F5	training data boundary
F6	normalization
F7	horizon endpoint determination
F8	cost information
F9	tick-derived information

• Mutation testing is extended to every primary horizon (15, 60, 240). Future-bar perturbation must not alter predictions at tested origins.
• Minimum spot checks: ≥ 3 origins per primary horizon.
• Any no-lookahead failure = INVALID / HARD STOP.
• DEFECT-001 SAX z-normalization invariant is preserved (candidate windows z-normalized before SAX encoding; regression test tests/test_sax_analog_zscore.py). Any regression = HARD STOP.

O. Economic analysis

• Q2 (economic survival) passes only if the dependence-aware 95% CI of net return under the frozen 7.57 bps cost has lower bound > 0.
• A positive gross result that fails Q2 is PARTIAL / ECONOMICALLY NOT VIABLE, not a successful boundary.
• Cost is applied at every origin identically; time-normalized economics are reported to prevent horizon-frequency artifacts.

P. Secondary diagnostics

• MASE (Drift vs naive) per horizon; gross/cost ratio; cost/range ratio; trade coverage; n and n_eff per horizon; turnover/trade frequency; annualized or time-normalized economics; EARLY / MIDDLE / LATE descriptive results.
• Secondary models (ARIMA, SAX, Hybrid) may be reported as robustness diagnostics only; they cannot establish the primary boundary claim.
• 960-minute diagnostic results (if any) are reported diagnostic-only and never enter the FWER family or the primary decision.

Q. Stop conditions

Before committing the frozen protocol, STOP if any of the following occurs:

• authoritative HEAD mismatch;
• working tree not clean before freeze;
• origin/main mismatch;
• R2 namespace conflict;
• source dataset hash cannot be verified;
• protocol contradiction;
• unresolved statistical specification;
• unresolved cost specification;
• unresolved model specification;
• unresolved gap rule;
• unresolved multiplicity procedure;
• accidental R1-R modification;
• any attempt to execute the experiment;
• any attempt to generate forecasts/backtests;
• any attempt to tune parameters from results.

An unresolved scientific ambiguity is not "fixed" by improvisation — it is returned to the owner.

R. Outcome matrix

outcome	definition
NULL	Global gate not rejected; no horizon in {60, 240} shows gross signal; consistent with R1-R null persisting.
BOUNDARY	Global gate rejected; ≥1 family horizon passes Q1 (A, B, C) AND Q2 (net CI lower bound > 0).
PARTIAL / ECONOMICALLY NOT VIABLE	Global gate rejected; ≥1 horizon passes Q1 but fails Q2.
INCONCLUSIVE	Underpowered horizon(s) prevent the frozen decision rule from resolving; or no-lookahead/specification ambiguity prevents classification.
INVALID	No-lookahead failure, DEFECT-001 regression, data provenance failure, or protocol violation → HARD STOP.

Q1 conditions (Drift only, all required):

• A. Direction accuracy dependence-aware 95% CI lower bound > 0.50.
• B. Gross return/trade dependence-aware 95% CI lower bound > 0.
• C. Error vs naive satisfies the pre-registered error criterion: the dependence-aware 95% CI of the paired error difference (|e_naive| − |e_drift|) has lower bound > 0 (equivalently, Drift MASE < 1 with significance).

Q2 (economic survival): dependence-aware 95% CI of net return under frozen 7.57 bps has lower bound > 0.

S. Reproducibility requirements

• A new R2-H manifest (results/R2-H/R2_H_MANIFEST.json) records: protocol; configuration; source dataset; source hash; horizon definitions; origin lists; origin hashes; code/config identity; statistical parameters; cost parameters; seed; generated design artifacts.
• Origin lists are generated twice and verified byte-identical (see §H).
• All future execution artifacts (if later authorized) must be hash-linked to this frozen protocol.
• R1-R evidence hashes are never reused as R2-H evidence.

T. Execution authorization boundary

• This document freezes design only.
• EXECUTION IS NOT AUTHORIZED by this freeze. No model fitting, forecasting, backtesting, parameter optimization, simulation, prediction, metric computation from new experimental runs, or economic experiment may begin until a separate EXECUTION AUTHORIZATION gate is granted by the owner.
• The required next governance gate is EXECUTION AUTHORIZATION.

U. Version / freeze metadata

field	value
protocol id	r2-h-v0.1.0-frozen
status	DESIGN FROZEN — EXECUTION NOT AUTHORIZED
freeze date	2026-08-26
branch	main
R1-R anchor	67be79aa75ae98eace7c67b74c93797f4327d51a
dataset	results/R1_real_data/XAUUSDc_M1_raw.json
dataset sha256	5472357e67f70f7d1bfef8f9de67fcce91ee6f26208072d7a6766b06c158ee6e
generator	scripts/run_r2h_origins.py
origin lists	results/R2-H/origins_H015.csv, origins_H060.csv, origins_H240.csv
cost	7.57 bps global constant
bootstrap	MBB B=60, reps=10,000, seed=20260826, 95% CI
FWER family	{60, 240} minutes, α = 0.05, joint block bootstrap, max-statistic

---

DESIGN FROZEN — EXECUTION NOT AUTHORIZED.

End of R2-H frozen protocol.