📈 Time Series Forecasting with Azure Machine Learning

This project is part of the DP-100 Certification and focuses on developing and deploying a robust time series forecasting solution using Azure Machine Learning.

---

🧠 Project Summary

• Developed a machine learning pipeline for forecasting sales data using the OJ Sales Simulated dataset from Azure Open Datasets.
• The original dataset contained over 12,000 CSV files representing 3 stores for each date.
• Initial exploration and aggregation were done to unify the dataset and prepare it for training.
• Preprocessed data to include lag features, date components (day, month, year), and handled multiple time series identifiers.
• Managed compute clusters and resources in Azure ML for efficient model training and testing.

---

⚙️ Modeling Process

• Attempted automated ML in Azure with time series forecasting mode. Faced issues due to multi-store data.
• Filtered the data to a single store (Dominicks 4128) for more accurate modeling.
• Top-performing models from AutoML included ARIMA, Seasonality, and Moving Average.
• Custom model built using SARIMAX (from statsmodels) with extensive hyperparameter tuning.

---

🧪 Evaluation Metrics

Metric	Value
MAPE	28.78
MAE	3299.61
RMSE	4617.57
Spearman Correlation	1.0
R2 Score	0.971

---

🧩 Modular Coding & MLFlow Integration

• Implemented function-based modular coding using main.py in Spyder IDE.
• Integrated MLflow for tracking hyperparameters, metrics, and artifacts.
• Included logging, exception handling, and visualizations (residuals, subplots, etc.).

---

☁️ Azure Deployment: Batch Endpoint

• Registered the model and deployed it via an offline batch endpoint on Azure.
• Used azureml and azureaiml libraries for workspace authentication and job submission.
• Created compute instance and batch endpoint using BatchEndpoint and BatchDeployment.
• Deployed the MLflow model with configurations like:
  • instance_count, mini_batch_size, output_action, retry_settings, etc.
• Monitored endpoint progress in Azure ML Studio and downloaded predictions from predictions.csv.

---

📤 Output & Results

• All inference results were logged and stored in Azure's default datastore.
• Final predictions visualized using Pandas and Matplotlib.

---

🔗 Key Technologies

• Azure ML SDK
• Python (Pandas, Numpy, Statsmodels)
• MLflow
• AutoML
• SARIMAX
• Batch Endpoint
• Logging & Modular Code

---

🙌 Acknowledgments

This project was built under the SkillUpOnline DP-100 Certification Program. Special thanks to the mentors and community that supported this journey.

---

🔗 Let’s Connect

If you're interested in collaborating or have feedback:

📎 Connect on LinkedIn