ICAIF v9 handoff — November 1, 2026

The maintained default branch is master. The ICAIF adapter/trainer and the completed FOMC research branch have been merged here. The architecture remains Jordi Corbilla’s stock_prediction_lstm.py, v9 shared 128/64 LSTM encoder with the expected-return head. Its SHA-256 is 2579877f487ed6e7b311804c56c47a66adf61955c6223e055d43bbada50aabd7.

Code and evidence map

stock_prediction_lstm.py: original architecture/factory.
quant_forecast_lab/icaif.py: canonical causal features and portfolio targets; the frozen arithmetic order is the default.
quant_forecast_lab/icaif_training.py: three-seed training through that factory; run from a source checkout, using the documented purged samples.
quant_forecast_lab/icaif_fomc.py: separately documented research extension, not part of the submitted model. Its later-year promotion gate failed.
Adapter methodology and historical experiments.
FOMC experiment and rejected promotion.
Fresh offline reproduction result.

The competition API client, portfolio accounting, sealed decisions and private credentials belong to the sibling JordiCorbilla/acm-icaif-2026-trading-agent-competition repository. They are not required to replay the public demonstration. Models and datasets are supplied in the frozen review package, rather than committed as new Git artifacts here.

Frozen reproducibility package

Final submission 2131 was uploaded once on October 31. Its owned receipt was VALID, PENDING_REVIEW on November 1; final ranking/review is not confirmed. Official trading decision 1545 is the earlier decision reproduced by this package.

Shared materials folder contains source.zip, the video, README, disclosures, dependencies and checksums. The canonical source.zip SHA-256 is aeca6ce28c755486e53c14b2713ac09c0b33bbe7a75093dc583a45e3db5183d6. Do not replace it with a new archive of today’s default branch.

Download source.zip, check that hash, and extract it to a dedicated folder. From the extracted root, follow its pinned Python 3.11/Windows setup:

py -3.11 -m venv ..\icaif-review-env
& ..\icaif-review-env\Scripts\python.exe -m pip install -r requirements.txt
& ..\icaif-review-env\Scripts\python.exe -B reproduce.py --verify-only
& ..\icaif-review-env\Scripts\python.exe -B reproduce.py --output ..\icaif-review-output

The output directory must be new and outside the extracted package. Replay needs no account credentials, data downloads or trading submissions. Package installation needs package-index access. Frozen checkpoints are canonical; new training is unnecessary and is not claimed to yield identical weights.

November 1 checks

The full unit/runtime suite passed in the existing LSTM environment, together with lint, compilation, dependency consistency and training CLI help checks. The competition suite passed all 138 tests. Lineage checks verified 34 archived models and the frozen live checkpoint hashes. Current canonical features match the stored 2023 train/validation/test tensors and the Official 30×20×6 input bit for bit. An independent replay using the pinned clean environment verified 188 packaged files and reproduced the three forecasts, ensemble and all weights. Legacy v7 compatibility checks emit existing scaler-version/deprecation warnings; those checks do not establish v7 compatibility across arbitrary runtimes.

An upstream running-sum optimization changed feature rounding by up to roughly 1e-6 on the checked historical tensors. It remains available explicitly as enhanced_features_fast for research; the default enhanced_features retains the exact frozen calculation. A small historical tensor fixture guards this. Neither frozen models nor cloud review materials were modified.

All remaining ICAIF automations were deleted at the user’s request. No scheduled monitoring or upload is needed. Keep shared materials accessible and unchanged through organizer review; respond manually if organizers request clarification.