Mid-Price Forecasting

Limit order books are noisy, high-frequency market microstructure data. This notebook studies whether short-horizon mid-price movement can be predicted from order-book features, then compares simple baselines with deeper sequence models and interpretability probes.

Why This Matters

Mid-price forecasting is a useful applied ML problem because it forces the model to deal with class imbalance, temporal dependence, and weak signal-to-noise ratios. The project is also a good testbed for interpretability methods: a model that predicts price movement should expose which levels, sides, or engineered features drive its decisions.

Method Overview

flowchart LR
    A[Kaggle LOB text files] --> B[Parse order-book rows]
    B --> C[Clean and normalize features]
    C --> D[Create train/validation/test splits]
    D --> E[Logistic baseline]
    D --> F[MLP baseline]
    D --> G[DeepLOB-style sequence model]
    E --> H[Accuracy and class report]
    F --> H
    G --> H
    G --> I[Feature ablation, saliency, integrated gradients]

Loading

Current Results

The checked-in notebook reports these representative metrics on the current split:

Model	Validation accuracy	Test accuracy	Notes
Logistic regression baseline	0.6087	0.7091	Strong baseline for simple normalized features
MLP baseline	0.5494	0.5943	Did not outperform the simpler baseline on this split
DeepLOB-style sequence model	0.8040 validation during training	Notebook includes sequence evaluation workflow	Best direction for further work

These numbers should be treated as exploratory notebook results, not production trading claims.

Dataset

This project uses the benchmark dataset by Ntakaris et al., available through Kaggle:

• Dataset: https://www.kaggle.com/datasets/praanj/limit-orderbook-data?resource=download
• Paper: Ntakaris, Adamantios, et al. "Benchmark dataset for mid-price forecasting of limit order book data with machine learning methods." Journal of Forecasting 37.8 (2018): 852-866.

After downloading, create a Data/ folder in the repository root and place the dataset text files there.

Reproduce

1. Install a Python environment with Jupyter, pandas, scikit-learn, PyTorch, matplotlib, and Captum.
2. Download the Kaggle dataset into Data/.
3. Open baseline.ipynb.
4. Run the notebook from top to bottom.

The notebook contains the full data preparation, model training, evaluation, and interpretability workflow.

Limitations

• The dataset is external and not committed.
• The results depend on the current split and notebook state.
• The project is research exploration, not a deployed financial system.
• Further work should add a scriptable training entry point, fixed environment file, and saved result artifacts.

Interested in this area? Email me at praneeth.suresh.s@gmail.com.