Data Science

You build a data science project from scratch to forecast daily sales volume using real-world retail data. You perform exploratory data analysis to uncover underlying trends and anomalies in the data, select clustering strategies to group articles and stores together, and train machine learning models to predict future demand. MLFlow systematically tracks your experiments and performance metrics along the way. An interactive Streamlit dashboard visualizes your forecasts against actual sales and baselines. Finally, you test different backtesting strategies to simulate a live production environment.

Week 1: Exploratory Data Analysis & Business Context

Preprocess and join raw data tables into a unified dataset. Conduct exploratory data analysis to deepen your data understanding and uncover underlying sales trends, seasonality, and demand anomalies across tables covering sales volumes, base prices, active promotions, store lifecycles, and operating hours.

Week 2: Store & Article Clustering

Apply unsupervised machine learning techniques to analyze and cluster high-dimensional retail data. Use algorithms such as K-Means or Agglomerative Hierarchical Clustering to segment articles and stores based on selected features, and evaluate overall cluster quality. As a bonus challenge, implement K-Means from scratch.

Week 3: Time-Series Foundations & Baselines

Build univariate time-series forecasting pipelines to predict daily sales across different horizons while preventing temporal data leakage. Establish naive baseline benchmarks and train machine learning models from Linear Regression to advanced algorithms across multiple articles. Track all experiments systematically using MLFlow, evaluating performance with retail metrics like WMAPE and BIAS. As a bonus challenge, implement hyperparameter tuning to optimize your model.

Week 4: Feature Engineering & Global Models

Perform feature engineering to predict daily sales one week into the future. Train machine learning models and systematically evaluate feature contributions using Gain-based importance or SHAP values and translate model insights for business stakeholders. As a bonus challenge, leverage your Week 2 clusters to build a global forecasting model and compare its performance against local models.

Week 5: UI Dashboard, Backtesting & Model Drift

Build an interactive Streamlit dashboard to generate sales predictions based on your pretrained models. Display predictions, baselines and metrics on various aggregation levels. Simulate production deployment through historical backtesting to evaluate different retraining strategies, optimizing update frequencies while detecting data drift. As a bonus challenge, evaluate model resilience on a shifted dataset from a new environment and implement techniques to mitigate concept drift.

Entry requirements:

• Solid Python fundamentals

• Comfortable using Jupyter Notebooks

• Knowledge of core data structures

• Basics in statistics and linear algebra (lightweight: vectors, matrices, dot product)