Demand Planning and Forecasting for Tito’s Vodka using both classic and Machine Learning (ML) models.
For most modern sales professionals, the job requires an arsenal of hard and soft skills in order to have a competitive advantage. Demand planning and forecasting are essential tools for sales leaders to stay competitive in a rapidly changing business environment and optimize their operations.
Sharper forecasting allows teams to develop a better long range strategy, set short to long term goals, improve brand spends and design more efficient incentives.
Project Goals
This project aimed to forecast Tito Vodka sales using a live and publicly available dataset to narrow the best performing demand planning and sales forecasting tools.
To achieve this, the project tested 12 different forecasting tools, narrowed them down to four, and created live and updating dashboards in Tableau to measure and compare the accuracy of the chosen models for this upcoming year (2023).
The project aimed to identify the least expensive and easy-to-deploy forecasting models that can quickly and easily improve forecasting.
The models were tested using MAPE and Absolute Error, with uni-variate data sets and kept all models tuned to default parameters.
The project tested three types of models, including straight line, moving averages and simple linear regression.
Key Takeaways
Models will be evaluated at years end, January 2024, against the predictive modeling in February 2023. And compared against their tested accuracy.
The classic models tended to outperform the more complex (though the ML were trained on a small and uni-variate set), with the exception of XG Boost.
With increasingly optimized machine learning models (such as XGBoost), abstracted and easy to use libraries (such as Darts), its easier and less computationally expensive to use a variety of models and tools.
All models were trained on Tito’s Vodka 750mL (9L) cases from 2012 through 2021 and tested against 2022 (12 months). MAPE and APE were used on 3, 6 and 12 month forecasts.
Top tested performing models were Seasonal Naive, XGBoost and ARIMA models.
Links to Products and Resources
Please visit the below links for products, code and full technical write-up.
Tableau Public Profile: Live Updates (monthly) against actuals
Technical Write Up: Comparing forecasting models for use in demand planning and sales forecasting using Tito’s Vodka
Press enter or click to view image in full size
Screenshot of Tableau dashboard created around chosen forecasting models
The Approach
Dataset: What are we capturing?
Off Premise Licenses ONLY
Liquor Sales (excludes beer, wine and non-alc)
Data Limitations
Dataset does not include On-Premise (bars, restaurants or any sale by drink “on” the premise)
Does not include Beer and wine for broader pulse on the Beverage Alcohol market.
Training the model
The goal was to simplify the training and testing of the models. For that, I used a uni-variate data set, one independent variable, and kept all models tuned to the default parameters. For any ML models, supervised only.
For the training set, Tito’s 750 mL (9L) case sales from 2012–2021 (10 years) was added to the model. For the testing set, the additional sales history for 2022 (1 year). The model performance was measured using the MAPE accuracy for 3 separate time periods.
In a real-world scenario, sales leaders need a minimum of 3 and 6 months accuracy to set sales goals, manage supply chain, and align incentives. For sales planners and executives, the need is for 12 months or greater accuracy, to set strategy, develop OKR’s and build out Future State models.
Testing Models
The goal was to simplify the training and testing of the models. For that, I used a uni-variate data set, one independent variable, and kept all models tuned to the default parameters.
Types of Forecasting Models
Straight Line: method assumes a company’s historical growth rate will remain constant. Forecasting future sales involves multiplying a company’s previous year’s revenue by its growth rate.
Moving Averages: the forecasts of all future values are equal to the average (or “mean”) of the historical data. In this example, we will look at a 4 month moving average periods to adjust for Warehouse and Larger retailer front-load purchases (Example: Kroger may order 90 days worth of goods vs monthly to ensure receipt, warehousing, distribution and merchandising).
Simple Linear Regression: regression model that estimates the relationship between one independent variable (sales) and one dependent variable (time) using a straight line.
Multivariable Regression: as the name would state, uses multiple variables. For this project, multivariable
Measuring accuracy
Moving Average Percent Error (MAPE): one of the most common forecast accuracy measurements. Sum of the individual absolute errors divided by the demand over the course of each test month. Average of the percentage errors.
Absolute Percent Error (APE): This was chosen to test the reliability of the model to accurately predict the cases for each time period. All models were tested out to 3, 6 and 12 months to determine how close the forecast came to the actual case amounts in 9L. Need an accurate estimate for demand planning, goal setting and revenue estimate.
Model Performance
Comparing across time periods and % errors, 4 models stood out, with the classical models taking an early first round.
Press enter or click to view image in full size
Top performing models were…
Seasonal Naive (classical): as described, this method is useful for highly seasonal data. In this case, we set each forecast to be equal to the last observed value from the same season. Essentially, applying trend to the value from the last period (previous year).
XGBoost (ML): an open-source Python library that provides a gradient boosting framework. It produces a highly efficient, flexible, and portable model implementation of the gradient boosted trees algorithm. Gradient boosting is a supervised learning algorithm, which attempts to accurately predict a target variable by combining the estimates of a set of simpler, weaker models.
Seasonal + Drift = Combined Forecast (classical): This is equivalent to drawing a line between the first and last observations, and extrapolating it into the future. Though it seems to have slightly displaced the accuracy as the brand has seen meteoric growth and has a much higher slope than a more established brand.
Auto-regressive Integrated Moving Average (ARIMA) (classic): auto regressive techniques use past behavior to predict future behavior, and that model is integrated with a moving average forecast, a statistical method used for forecasting long-term trends. The technique represents taking an average of a set of numbers in a given range while moving the range. As expected, the model makes major improvements out to 12 months.
Project Next Steps
Live and interactive dashboard has been loaded onto Tableau Public to be revisited 30 days following the close of 2023 (Feb 2024) to evaluate the forecasting models against the foretasted predictions.
Tableau Public has dashboard updating YTD MAPE and APE for each model
About the author:
Harry McKaig is the CEO of Double Cross Vodka and a board advisor to emerging CPG brands. With over two decades of experience in beverage alcohol and an MS in Finance from Georgetown University, he brings deep expertise in revenue operations, go-to-market strategy, and brand development. Follow him for insights on spirits, sales, and scaling smart.
