RTDLinear: Leveraging Residual Time-aware Decomposition Linear for ATM Cash Balance Prediction

---

For Reproduction

1. Setup

1. Install pip install -r requirements.txt
2. Install Pytorch 2.x manually with cuda

2. Check Results

For convenience, benchmark results are displayed in a jupyter notebook: click to see results

3. Directory Structure

Assets Directories

• ./config/{xxx model}.yml: Configurations for all models
• ./data/: All the data (raw, processed, info)
• ./log/cluster{n}/{xxx model}/{experiment time}/: runtime log, metrics, model_configs, checkpoints

Source Code Directories

• ./model/: Code for models implementations
• ./trainer/: Code for Model Training and Testing
• ./util/: Code for other uses

3. To Test Trained Neural Models Individually

run python individual_test.py --model Model where parameter Model can be:

• GRU (for benchmark)
• LSTM (for benchmark)
• DLinear (for benchmark)
• TDLinear (for ablation study)
• RTDLinear

4. To Train Neural Models Individually

run python individual_train.py --model Model where parameter Model can be:

• GRU (for benchmark)
• LSTM (for benchmark)
• DLinear (for benchmark)
• TDLinear (for ablation study)
• RTDLinear

4. To Train and Test ML Models

Each ML/Stat model is trained, validated and tested in an independent notebook. Results are stored in ./notebooks/ directly.

• see ARIMA
• see MLP
• see SVR

5. To Run Clustering

• see clustering_a_s.ipynb

6. (Not Necessary) To Run Data Preprocessing

• raw data is saved in ATM cash balance.csv
• data preprocessing: see filter.ipynb

---

ATTENTION: This repository is protected by LICENSE AGPL v3.0. Any commercial use in closed source software as well as cloud service is not permitted.

---

Copyright © 2025 Liu Shu. All rights reserved.