ARIMA_SAX_Hybrid_Forecaster/RESEARCH_PROTOCOL.md

# RESEARCH_PROTOCOL.md

## 1. Role

Authoritative governance for the ARIMA + SAX hybrid forecasting research project.
Owned by `chiki2bum2`. This bootstraps a **new, isolated repository** and does not
reference any existing project implementation state.

## 2. Central question

> Does a hybrid ARIMA + SAX forecasting architecture provide **statistically and
> economically meaningful incremental predictive information** compared with
> appropriate baselines and with each component individually?

`D > B`, `D > C`, `D > A`, and — most importantly — whether any incremental
information provided by `D` is **stable out-of-sample** rather than merely
in-sample or restricted to selected historical periods.

## 3. Initial hypotheses (do not assume true)

* **H0** — ARIMA + SAX does not provide materially better information than the
  strongest component or baseline after OOS evaluation and explicit costs.
* **H1** — ARIMA + SAX provides statistically meaningful, stable incremental information.
* H1a — ARIMA captures structure beyond the naive baseline.
* H1b — SAX captures pattern structure beyond the baseline.
* H1c — ARIMA and SAX are partially complementary.
* H1d — Agreement has predictive value.
* H1e — Disagreement identifies unstable/uncertain regimes (rejection filter).

## 4. Design principle

**Never** average two raw price forecasts. Convert both components to a **common target**:

```
Y(t,H) = (Close[t+H] - Close[t]) / ATR[t]
```

Both models must express forecasts against the **same**:
symbol, timeframe, forecast origin, forecast horizon `H`, target definition,
normalization convention. The hybrid layer combines **evidence**, never incompatible
raw prices.

## 5. Hard rules

1. Closed-bar data only for forecast evaluation.
2. All forecast inputs strictly before the forecast origin.
3. The evaluation outcome is revealed only **after** the prediction is frozen.
4. No future observations in model fitting.
5. No arbitrary parameter sweeps to make a backtest look good.
   Any search is pre-defined, bounded, documented, confined to train/valid, and
   frozen before final OOS evaluation.
6. Never report gross results as net results. Transaction costs are explicit.

## 6. Stop conditions (hard stop)

Immediately stop and report if any of the following is detected:

* lookahead bias
* future observations entering model fitting
* training/evaluation overlap
* data integrity failure
* non-reproducible result
* major implementation discrepancy
* metric definitions changed after observing results
* retrospective parameter changes to improve OOS

Do **not** silently repair a scientific-methodology violation.

## 7. Report separation

Every report separates:
- **OBSERVED** — what was actually measured
- **INFERRED** — what the measurements suggest
- **UNKNOWN** — what has not been established
- **DECISION** — the justified next research action

```
Never turn: high correlation / R² / low price error / high historical similarity
into a claim of trading edge without economic OOS evidence.
```

## 8. Scientific conclusion vocabulary

Only: `SUPPORTED`, `INCONCLUSIVE`, `NOT TESTED`, `FAILED`.
Never: `PROFITABLE`, `EDGE CONFIRMED` (unless directly supported by an explicitly
defined and reproducible experiment).

## 9. Correct outcomes are valid

ARIMA wins / SAX wins / hybrid wins / naive wins / none has useful information /
hybrid works only in specific regimes / evidence is inconclusive — **all are valid**.
Do not bias the implementation toward a preferred conclusion.