protocol.md

# Forecasting v7: Single-Split End-to-End Fit and Development-Test Export

## Objective

Train three independently seeded copies of the selected shared AAPL/ABBV/MCD
architecture end to end, evaluate them once on a frozen 2022-2023 development
test, and preserve complete prediction outputs for later strategy research.
The three copies have identical architecture and data but use matched
seed-0/1/2 early-add initial checkpoints and independent deterministic training
orders. The sealed interval
`[2023-08-08, 2026-08-08)` remains inaccessible.

## Time contract

All ranges are half-open XNYS session-date intervals.

| Role | Interval | Use |
|---|---|---|
| Train | `[2016-08-08, 2021-08-08)` | Gradient fitting and train-only calibration |
| Validation | `[2021-08-08, 2022-08-08)` | Epoch selection for the frozen recipe |
| Refit | `[2016-08-08, 2022-08-08)` | Final development-test model fit |
| Strategy-development test | `[2022-08-08, 2023-08-08)` | One-shot prediction export and model scoring |
| Sealed reserve | `[2023-08-08, 2026-08-08)` | Never read, derived, summarized, or scored |

The 2022-2023 interval had feature exposure in an obsolete pipeline, although
no model metric was computed. It is therefore a strategy-development test, not
an untouched holdout. Model architecture, inputs, optimizer family, class
count, queried horizons, and output schema freeze before its labels or scores
are opened. Model development does not resume after opening it.

## Frozen model

- One shared checkpoint for AAPL, ABBV, and MCD.
- Context: 128 completed one-minute close tokens.
- Sequence inputs: close return, observed flag, elapsed wall-clock time, and
  regular-session linear/sine/cosine position.
- Early additive ticker embedding.
- Continuous RoPE-style horizon query with base 16.
- Supervision at 2, 4, 8, 16, and 32 minutes.
- Query support and evaluation at every integer horizon from 2 through 32.
- One shared 21-class output projection; no ticker- or horizon-specific heads.
- Auxiliary input: 5- and 20-completed-session net return, RMS return, range,
  terminal drawdown, and availability, plus session-to-date return from open,
  distance to maximum and minimum close so far, range position, and
  availability.
- All temporal, query, ticker, auxiliary-fusion, and output parameters are
  trainable in the end-to-end stage.

Every anchor with at least one exact same-session target is retained. Target
masks preserve late-session 2-, 4-, 8-, or 16-minute examples when longer
horizons are unavailable.

## Fit protocol

1. For seeds `0, 1, 2`, initialize from the corresponding locked shared
   early-add checkpoint whose training interval is exactly
   `[2016-08-08, 2021-08-08)`. Architecture and checkpoint recipe must match;
   only seed changes.
2. Add the identity-initialized typed auxiliary lane.
3. Train every parameter for the complete fixed epoch budget on the train
   interval. Score validation after every epoch and select its lowest
   equal-asset canonical NLL checkpoint.
4. Reinitialize from the same locked checkpoint and auxiliary identity,
   combine train and validation, and refit for the selected epoch count.
5. Freeze weights and calibration. Generate each seed's development-test
   prediction artifact exactly once.

The validation-selected epoch is a fitting decision. Development-test scores
or strategy results cannot alter it.

## Prediction artifact

Write exactly one immutable Parquet prediction sheet per seed. Each sheet
contains all three assets so stable row IDs can be joined across seeds without
rewriting an artifact. Checkpoints, calibration arrays, manifests, and hashes
are separate supporting artifacts.

Prediction arrays:

- stable row ID;
- asset and anchor row index;
- anchor timestamp and prediction-availability timestamp;
- session date and minute of session;
- queried horizons `2..32`;
- full float32 probability tensor `[row, horizon, 21]`;
- expected return implied by train/refit-only class means;
- expected positive-return probability;
- entropy and maximum-class probability;
- target-availability mask.

Outcome arrays are separate:

- stable row ID;
- exact target timestamp;
- realized log return;
- class label under frozen refit boundaries;
- anchor and target raw close;
- validity mask.

Calibration contains per-asset class edges, return means, positive rates,
priors, token normalization, and auxiliary normalization. The manifest records
all date boundaries, row counts, code/data/checkpoint hashes, probability
normalization checks, and maximum source/anchor/target timestamps.

## Ensemble uncertainty and strategy-development protocol

This protocol was frozen before any v7 development-test score was computed.
The probability average is the primary ensemble. Member entropy, ensemble
entropy, mutual information (ensemble entropy minus mean member entropy),
pairwise probability disagreement, expected-return dispersion, variation
ratio, and horizon consistency are candidate uncertainty variables. Individual
seed scores remain controls.

The first mining hypotheses are deliberately broad:

1. larger absolute ensemble expected return has monotone realized-return
   separation;
2. low seed disagreement strengthens a fixed expected-return signal;
3. agreement in sign across the `2..32` minute curve strengthens the signal;
4. entropy, adaptive prediction-set size, or conformalized scalar uncertainty
   identifies lower-risk subsets;
5. probability averaging improves NLL, Brier, ranked probability score, and
   calibration over the median member.

Full-year plots and tables are descriptive. Threshold or functional-form
discovery uses `[2022-08-08, 2023-02-08)`; a chosen rule is then frozen and
retested on `[2023-02-08, 2023-08-08)`. Every result is also sliced by asset,
month/quarter, session third, horizon, direction, volatility, and confidence,
with leave-one-asset and contiguous time-block checks. Search tables retain all
attempted slices and use simultaneous or false-discovery-aware uncertainty
rather than reporting only the maximum.

Conformal methods follow Angelopoulos and Bates (2021), but overlapping minute
targets are not exchangeable iid examples. Marginal prediction-set coverage is
therefore reported as a diagnostic unless calibration is grouped by complete
sessions. Any development-test conformal threshold is strictly past-only after
a frozen burn-in and is evaluated on later sessions; it never uses a future
outcome to score an earlier prediction. Group-conditional coverage by asset and
horizon is reported where sample size supports it.

Later strategy research joins predictions and outcomes by stable row ID. Entry
cannot precede the prediction-availability timestamp. Initial confidence
research ignores execution costs as requested, but every table is labeled
gross and does not claim tradability. Subsequent strategy evaluation uses
tradable prices and reports turnover, holding time, overlap, exposure, costs,
average return, cumulative return, Sharpe, drawdown, and breadth by asset and
period. Development findings cannot change the architecture, training recipe,
or seed ensemble. The sealed three-year reserve is not opened during strategy
development.