Monthly Demand Forecasting for a multi-billion dollar multinational in the advanced materials sector
In this project, I was tasked with forecasting unit sales for the next six months using historical monthly data from a large multinational in the industrial coatings sector (Mexican subsidiary, anonymized for confidentiality).
🧠 Project Scope
The goal was to provide accurate volume forecasts for various brand-region combinations. I worked with monthly time series data that included:
- Seasonality and trend components
- External regressors (1 categorical, 1 continuous)
- Autoregressive structure
🛠️ Modeling Strategy
Rather than building a single global model, I opted for a grouped modeling approach, training one model per (brand, region) segment. This allowed for finer control and more accurate modeling of local patterns.
I evaluated several forecasting models per group:
- SARIMAX (with and without differencing)
- Seasonal Autoregressive State Space (SAS)
- Holt-Winters
- Prophet
Each model was tuned using Optuna for hyperparameter optimization with a Bayesian search strategy, guided by a 5-fold expanding window cross-validation process. The primary evaluation metric was Root Mean Square Error (RMSE) on out-of-fold predictions.
📊 Cross-Validation Results
Below is a summary of cross-validation performance for the top 10 brand-region groups:
📈 Forecasts
For each group, I generated:
- Out-of-fold predictions (to visualize model accuracy)
- Future forecasts (next 6 months)
- 95% prediction intervals
The following plot is representative of the final outputs delivered, for one of the top groups that was fit best by a Prophet model:
Consider that this is an out-of-fold prediction plot, not a training set fit, which would likely look very close to the actual data for all months.
🧹 Data Handling
Significant effort went into preprocessing and cleaning:
- Outlier handling
- Missing value imputation
- Encoding categorical features
- Normalization of external regressors
The data was aligned to ensure all time series were consistent and complete before modeling.
🧪 Deliverables and Deployment
The final deliverables included:
- Forecast plots and prediction files for each group
- Saved model files per group for potential reuse
- Clear documentation on retraining and inference strategy
While the project didn’t include deployment, the structure supports batch retraining and monthly prediction, making it easy to integrate into a production pipeline.
💼 Value
This project demonstrates my ability to:
- Work independently on large forecasting initiatives
- Identify and correct input data errors and inconsitencies
- Select and validate models in a robust, unbiased way
- Communicate results clearly and professionally
If your organization needs time series forecasting or custom AI solutions, feel free to reach out — I’d be happy to collaborate.