Datasets:

mythezone
/

LOBench-A-share-processed

This snapshot contains the experiment implementations and configurations for the accepted IEEE Transactions on Big Data manuscript. Public configurations use asset_NNN identifiers and the existing NPZ files. The classification code maps labels -1, 0, 1 to class indices 0, 1, 2 internally.

Environment and checks

Python 3.12 is recommended. Install dependencies in a virtual environment and select a PyTorch build appropriate for your GPU if needed:

python -m pip install -r requirements.txt
python -m unittest discover -s experiments/tbd_revision/tests -v
python smoke_test.py

The validation environment is recorded in environment-tested.json. Results can vary with software versions and hardware. The smoke test checks execution on a small subset, not agreement with reported scores.

Main 50-stock comparison

python experiments/tbd_revision/scripts/train_representation_baselines.py \
  --config experiments/tbd_revision/configs/final_revision_ashare50_adaptation_3seed.json

This runs the three main encoders (timae, lobgrid_timae, simlob2026), three seeds, and direct training, frozen probing, and full fine-tuning. Pooled pretraining uses the five-asset training set specified in the configuration. Validation selects checkpoints; test labels are not used for fitting.

For one formal case, preserve the hyperparameters and restrict the case list:

python experiments/tbd_revision/scripts/train_representation_baselines.py \
  --config experiments/tbd_revision/configs/final_revision_ashare50_adaptation_3seed.json \
  --stocks asset_001 --methods timae --seeds 2026

Results and checkpoints are written under experiments/tbd_revision/outputs/.

Reconstruction and imputation

The reconstruction structural-weight experiment uses the original two configurations:

python experiments/tbd_revision/scripts/run_representation_task_suite.py \
  --config experiments/tbd_revision/configs/final_revision_ashare5_structural_tick_3seed.json
python experiments/tbd_revision/scripts/run_representation_task_suite.py \
  --config experiments/tbd_revision/configs/final_revision_lobgrid_structural_tick_3seed.json
python experiments/tbd_revision/scripts/run_representation_task_suite.py \
  --config experiments/tbd_revision/configs/final_revision_imputation_structural_3seed.json

Run the nine supplementary architecture adapters with:

python experiments/tbd_revision/scripts/run_architecture_task_coverage.py \
  --config experiments/tbd_revision/configs/final_revision_architecture_task_coverage.json

These are benchmark implementations under a shared task and decoder, not a claim to reproduce each original method's published score.

Additional prediction and sensitivity experiments

Use train_representation_baselines.py --config with these configurations:

Configuration	Experiment
`final_revision_ashare5_pooled_3seed.json`	Five-stock pooled pretraining
`final_revision_ashare5_local_3seed.json`	Five-stock local pretraining
`final_revision_ashare50_tlob_supervised_3seed.json`	Supervised TLOB-based reference

All runner paths below start with experiments/tbd_revision/scripts/ and all configurations are under experiments/tbd_revision/configs/.

python experiments/tbd_revision/scripts/run_absolute_prediction_baselines.py \
  --config experiments/tbd_revision/configs/final_revision_ashare50_adaptation_3seed.json \
  --results-dir experiments/tbd_revision/outputs/absolute_baselines

python experiments/tbd_revision/scripts/run_hyperparameter_sensitivity.py \
  --base-config experiments/tbd_revision/configs/final_revision_ashare5_hyperparameter_base.json \
  --output-root experiments/tbd_revision/outputs/hyperparameters \
  --variants baseline latent64 latent256 width64 width128 heads2 heads8 mask20 mask60 lr100 lr1000 batch64 batch256 wd0 wd1e4

python experiments/tbd_revision/scripts/run_label_threshold_retraining.py \
  --base-config experiments/tbd_revision/configs/final_revision_label_threshold_retraining.json \
  --output-root experiments/tbd_revision/outputs/label_threshold \
  --variants threshold0001 threshold0005 threshold0010

profile_final_revision_complexity.py measures parameter counts and timing. The summarize_*.py scripts aggregate generated case outputs. Consult each runner's --help for input and output paths, and compare runtime measurements only on equivalent hardware.

Experiments requiring additional data

FI-2010: obtain the official files and set the source paths in final_revision_fi2010_decpre_train_only_data.json. Run experiments/tbd_revision/src/prepare_fi2010_official.py --config with that file, then the common training runner with final_revision_fi2010_decpre_train_only_local_3seed.json. Preserve the normalization variant and chronological boundaries.

LOBSTER: obtain source files under the provider's terms and set raw_root in final_revision_lobster_corrected_natural.json. Run experiments/tbd_revision/src/prepare_nasdaq_lobster.py --config with that file, then the common training runner with final_revision_lobster_local_3seed.json.

Five-stock training history: the article compares seven versus 237 training dates with the same year-end holdout. The public January panel cannot substitute for those inputs. prepare_training_duration_pilot.py with training_duration_pilot_data.json prepares the two conditions when the authorized 2019 source data are available. The four training_duration_{short10,full_year}{,_pooled}.json configurations use the common trainer. The preparation example uses alias-based CSV filenames; the data custodian must map the authorized source files locally.

Raw-feed reconstruction: data_pipeline.py, the external adapters, and the A-share preprocessing configuration describe the processing pipeline. The private source-to-alias mapping is not distributed. Reproduction from this repository begins with the processed tensors unless the relevant source data are supplied separately.

Packaging changes

Replaced A-share identifiers with the existing public aliases and removed machine-specific input paths.
Pointed public A-share configurations to ../deidentified_50stock_panel/.
Added loading of natural_train_idx when a public NPZ lacks legacy train_idx; legacy indices retain precedence.
Reused the manifest reader for architecture-coverage normalization, including scale validation.
Preserved relative trading-day groups when absolute timestamps are absent.
Added tests and a bounded CPU smoke check. Model equations, losses, and formal experiment hyperparameters were not changed.

RELEASE.json and checksums.sha256 identify this code snapshot. Publication details remain pending; use the repository's CITATION.bib.