How leonardoberti00/tlob works
TLOB is a research and training system for classifying price trends from limit-order-book sequences. The principal workflow configures a model and dataset, optionally prepares raw data, loads labeled sequence windows, trains and evaluates a selected model, then saves predictions and checkpoints. Separate preprocessing branches support BTC, FI-2010, and LOBSTER; the repository also includes several alternative model architectures and a backtesting entry point. Edges reflect shown imports, calls, or README statements; unsampled subsystem details are not inferred.
Main entry points: main.py Training entry · run.py Experiment runner · utils_model.py Model selector · engine.py Training engine
Components 20
Configuration and orchestration
- Experiment config
config/config.py - Training entry
main.py - Experiment runner
run.py
Market data pipeline
- BTC preparation
preprocessing/btc.py - FI-2010 preparation
preprocessing/fi_2010.py - LOBSTER preparation
preprocessing/lobster.py - Data transforms
utils/utils_data.py - Sequence datasets
preprocessing/dataset.py
Model architectures
- Model selector
utils/utils_model.py - TLOB
models/tlob.py - MLPLOB
models/mlplob.py - DeepLOB
models/deeplob.py - BiN-CTABL
models/binctabl.py - BiN normalization
models/bin.py - Attention utilities
utils/attention.py
Training and evaluation
- Training engine
models/engine.py - Saved checkpoints
data/checkpoints - Backtesting
run_backtesting.py
Other components
- Researcher
- Weights & Biases
Connections 25
- Researcher → Training entry: configures run
README.md - Training entry → Experiment config: loads config
main.py - Training entry → Experiment runner: starts run
main.py - Training entry → BTC preparation: prepares BTC
main.py - Training entry → LOBSTER preparation: prepares LOBSTER
main.py - Training entry → FI-2010 preparation: extracts data
main.py - Training entry → Weights & Biases: starts logging
main.py - Experiment runner → Sequence datasets: batches windows
run.py - Experiment runner → Training engine: trains model
run.py - Experiment runner → BTC preparation: loads prepared data
run.py - Experiment runner → LOBSTER preparation: loads prepared data
run.py - Experiment runner → FI-2010 preparation: loads prepared data
run.py - BTC preparation → Data transforms: transforms data
preprocessing/btc.py - LOBSTER preparation → Data transforms: labels and normalizes
preprocessing/lobster.py - Sequence datasets → Data transforms: encodes event types
preprocessing/dataset.py - Training engine → Model selector: selects architecture
models/engine.py - Model selector → TLOB: instantiates
utils/utils_model.py - Model selector → MLPLOB: instantiates
utils/utils_model.py - Model selector → DeepLOB: instantiates
utils/utils_model.py - Model selector → BiN-CTABL: instantiates
utils/utils_model.py - TLOB → BiN normalization: normalizes with
models/tlob.py - TLOB → MLPLOB: uses MLP layer
models/tlob.py - Training engine → Saved checkpoints: saves predictions
models/engine.py - Researcher → Backtesting: runs backtest
README.md - Training engine → Weights & Biases: logs metrics
README.md