Role overview
Own the development of forecasting and marketing analytics models, helping move an existing regression-based approach toward more advanced time-series methods. The work spans exploring model options, validating them against future data, and turning results into recommendations that can support marketing decisions.
Responsibilities
- Design, train, and evaluate time-series forecasting models, including transformer-based approaches and alternatives.
- Run experiments with Python and its scientific computing and machine-learning libraries, and use managed cloud tooling for model training.
- Build robust evaluation plans using backtesting, temporal splits, rolling or expanding windows, and safeguards against data leakage.
- Compare model performance and explain the reasons behind differences between approaches.
- Contribute to marketing mix modeling, attribution, and budget optimization by analyzing channel impact, saturation, and investment scenarios.
- Communicate assumptions, trade-offs, and findings to technical and business stakeholders.
Requirements
- Strong machine-learning and data-science experience focused on time series and forecasting.
- Practical experience building and validating forecasting models with future or unseen data.
- Proficiency in Python, with experience using PyTorch, NumPy, Pandas, and Scikit-learn.
- Experience training models with AWS SageMaker.
- Solid grounding in statistics, inference, optimization, and model interpretability.
- Ability to independently frame open-ended problems, choose and defend a modeling approach, and connect results to business outcomes.
Nice to have
- Experience with Chronos-2 or other time-series transformer architectures such as TFT, PatchTST, or Informer.
- Familiarity with Bayesian optimization or scikit-optimize.
- Experience with marketing mix modeling, attribution, or causal modeling, including concepts such as adstock, saturation, incremental contribution, and budget allocation.
- Familiarity with tools such as Robyn, PyMC-Marketing, or LightweightMMM.