← Back to Projects

Time Series Forecasting 📈

Project Details

This project enhances TV ad attribution by introducing a forecasting-based baseline model for website traffic. The system trains a model to predict the expected website traffic without TV effects, capturing normal user behavior influenced only by seasonality and trends. This predicted traffic — the baseline — is then compared against actual website visits. Any significant uplift beyond the baseline is attributed to TV campaigns, enabling a more accurate measurement of spot performance.

We used Prophet, an open-source forecasting model, to capture complex seasonality and trend patterns while supporting additional features. The end-to-end pipeline was developed and deployed using Metaflow, Netflix’s open-source framework for orchestrating machine learning workflows.

Time Series Forecasting - MLOps

The system retrains models daily with the last 6 weeks of traffic data, ensuring forecasts incorporate the latest behavioral trends.

Key Features

My Contribution

I contributed to the design and development of the abstraction layer that dynamically builds machine learning pipelines from configuration files, enabling flexible experimentation and deployment. As part of this, I developed three core pipelines: a training pipeline to train new models and register them in MLflow, an inference pipeline to load trained models from MLflow and generate website traffic predictions, and an evaluation pipeline to run multiple experiments in parallel with different configurations while tracking results in MLflow. I also played a key role in selecting the best-performing model, tuning its hyperparameters, and engineering features to improve its performance. Finally, I productionalized the training and inference pipelines in AWS, ensuring the system could run automatically and scale to meet the demands of daily operations. Beyond development, I shared our approach publicly: I presented how we leveraged scikit-learn Pipelines with Metaflow at the DataGeeks Conference P7S1 x Munich Datageeks.

Challenges and Solutions

Outcome

Technologies

Python Time Series Forecasting Website Traffic Prophet Metaflow Pandas Scikit-Learn Evaluation MLflow Preprocessing AWS EventBridge AWS Step Functions