LSTM Time-Series Forecasting

A hands-on project for forecasting time-series with PyTorch LSTMs. It creates realistic daily data (trend, seasonality, events, noise), prepares it with sliding windows, and trains an LSTM to make multi-step predictions. The project tracks errors with RMSE, MAE, MAPE and shows clear plots of training progress and forecast results.

---

Features

• Synthetic daily series generation (configurable length & seed)
• Sliding-window dataset preparation for supervised learning
• LSTM model with dropout, Adam optimizer, early stopping
• Multi-step forecasting (direct & recursive)
• Metrics: RMSE, MAE, MAPE
• Visualizations: training curves & forecast plots
• Saved artifacts: best_lstm.pt, scaler.pkl, metrics.json

Results

• RMSE: 22.25
• MAE: 16.09
• MAPE: 7.64%

---

Forecast vs Actual

---

Training & Validation Loss

---

Project Structure

lstm-time-series-forecasting/
├─ README.md
├─ LICENSE
├─ requirements.txt
├─ data/
│  └─ generate_series.py
├─ src/
│  ├─ train_lstm.py
│  ├─ evaluate.py
│  └─ utils.py
└─ outputs/
   ├─ metrics.json
   ├─ forecast_plot.png
   ├─ training_curves.png
   ├─ best_lstm.pt
   ├─ scaler.pkl
   └─ (auto-created figures & reports)

Setup

python -m venv .venv
# Windows:
.venv\Scripts\activate
# macOS/Linux:
source .venv/bin/activate
pip install -r requirements.txt

Generate Data

python data/generate_series.py --start 2020-01-01 --end 2025-12-31 --seed 42 --out data/daily_series.csv

Train Model

python src/train_lstm.py --input data/daily_series.csv --horizon 30 --lookback 60 --epochs 30 --batch-size 64 --outdir outputs --seed 42

Evaluate

python src/evaluate.py --input data/daily_series.csv --model outputs/best_lstm.pt --lookback 60 --horizon 30 --outdir outputs

Outputs

• outputs/metrics.json – RMSE, MAE, MAPE
• outputs/training_curves.png – training & validation curves
• outputs/forecast_plot.png – forecast visualization
• outputs/best_lstm.pt – trained PyTorch model
• outputs/scaler.pkl – fitted scaler