A leading Australian automobile manufacturer was in need of higher quality, more accurate demand forecast for their products. This included knowing the optimum volume and type of stock each dealer should hold in order to maximise sales while minimising excess working capital tied up in inventory. Additionally, they wanted to know the effect of competitors and their own sales promotions on the sales of their products.
Our solution provided stable and accurate forecasts at the colour, model and dealer level by integrating the effects of promotional forecasting, marketing expenditure, delivery lead time and supply constraints in addition to yearly, monthly, weekly, and cyclical seasonalities. Our implementation also accounted for the effects of exogenous factors such as holidays, weather, and local demographics. Moreover, promotional forecasting identified the effect of promotions on substitute and complementary goods such that all affected stocks were optimised to meet demands and reduce excess inventory.
Our approach balanced statistical and machine learning disciplines to construct a hybrid system that provided both the flexibility of nonparametric models and the stability of statistical methods. The initial stage of preprocessing augments existing data by imputing outliers and restructuring the dataset for various predictive models. Additional features were also engineered through the underlying time series in reflection from our past experiences and expertise. Subsequently, predictions from statistical models, in conjunction with nonparametric and deep learning based recurrent networks, were stacked together to mitigate overfitting of any singular model.
Our demand forecasting implementation provided a dynamic, large-scale and systematic way of optimising retail inventory management and increased the accuracy of budgeting.
As a result, the client was able to better focus resources and expertise in their core objective of providing a better experience and service for their customers.