Merqato Price Forecasting

A time-series forecasting project designed to predict strawberry prices with a 2-week horizon using a simple Transformer model. The dataset includes weekly data between 2013 and 2023 with weather and pricing features.

📁 Project Structure

├── data/
│   ├── dataset.py         # PyTorch Dataset and DataLoader split logic
│   └── preprocess.py      # Preprocessing, time-aware imputation logic
│
├── local_data/
│   ├── models/            # Trained model weights (e.g. model.pt)
│   ├── plots/             # Saved performance plots
│   ├── processed/         # Preprocessed parquet files
│   └── raw/               # Raw CSV data
│
├── model/
│   ├── model.py           # Transformer model definition
│   ├── simulate.py        # Evaluation script with metrics + matplotlib plots
│   └── train.py           # Training loop with wandb integration

🧠 Core Functionality

1. Preprocessing

• Safe handling of missing values
• Weekly date reconstruction
• Time-aware imputation of prices with Random Forest regressors if desired

python -m data.preprocess --data local_data/raw/senior_ds_test.csv --impute

2. Dataset

• Custom PriceForecastDataset for rolling-window normalization
• Splits data by year into train/val/test
• Efficient DataLoader setup

3. Model

• Minimal Transformer encoder
• Sequence-to-one prediction
• Continuous price output

4. Training

• Logs to Weights & Biases (wandb)
• Early saving of best model

python -m model.train --data local_data/processed/data_2013_2023.parquet

5. Evaluation

• MAE, RMSE, MAPE, R² metrics
• Matplotlib plots for predictions vs actuals

python -m model.simulate --model_path local_data/models/original.pt

⚙️ Requirements

Install everything with:

conda env create -f environment.yml

Then install preferred PyTorch in that environment, e.g.:

conda activate dylan-merqato

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

📌 Notes

• All preprocessing is done without data leakage (only past data is used).
• The model is very lightweight and intended as a prototype — you can easily extend it with more features, attention masks, or richer architectures.
• Data is stored locally to prevent large files in version control.

👋 Author

Dylan Prins @ Neurality for Merqato · 2025