Demand forecasting and machine learning

The traditional statistical models are not designed to learn continuously from data and make decisions. Hence, they become obsolete when new data comes in and forecasting is undertaken. The solution to this critical problem is machine learning, that can help a supply chain to forecast accurately and manage it efficiently.

A fundamental concern in forecasting is the measure of forecasting error for a given data set and a given forecasting method. Accuracy can be defined as “goodness of fit” or how well the forecasting model is able to reproduce data that is already known (Makridakis and Wheelwright, 1989).

Decentralization is getting increasingly easy. Machine learning and advanced statistical algorithms capable of preparing the data, while automated data visualizations provide the user with insights on a silver platter. To handle these challenges of Big Data, new statistical thinking and computational methods are needed. For example, so many old methods that perform well for medium sample size do not scale to large amount of data. In the same way, the statistical methods that can perform well for less volume data are facing challenges in analyzing large volume data. To design effective statistical procedures for exploring and predicting Big Data, we need to find Big Data problems such as heterogeneity, noise accumulation, spurious correlations, and incidental endogeneity, in addition to balancing the statistical accuracy and computational efficiency.

The most basic neural network (a single neuron network) can outrun the traditional forecasting methods of moving averages, exponential smoothing and regression. The first order functional link network yields better results for one - week ahead forecast. The single layer, multiple input, feedforward network provides better performance for three – week ahead forecasting, according to a comparison study.

It is difficult to forecast demand for a product which has no sales history, to overcome this Machine Learning models can consider technical indicators such as product attributes, web analytics etc. A global leader eyewear company, Luxottica adds new styles to its collection annually. It is using machine learning to cluster the behaviors of past launches, and selecting the suitable performance for the new product, then “ common demand behaviors in the first launch period. Therefore, WMAPE (Weighted Mean Absolute Percent Error) improved by 10% and reduced the forecast baseline on new launches by about 30%.

According to Mr. John Wang; what if, instead of asking for an accurate forecast, you asked, "what is the optimum policy?" This is the realm of prescriptive analytics, and it can provide better results than planning based on a forecast (inevitably with errors). By integrating with a decision support system, prescriptive analytics can give decision makers timely recommendations of suitable course of action, help to conduct "what-if" analyses, and evaluate options under constraints. Ultimately, as more data enters the machine learning system’s storage, it will become more intelligent and the data will become efficient and easier to interpret.