Data-Oriented Businesses Part 1

In the eyes of the consumer, online markets can be thought of as having a big supply of products, which in reality are bound to physical stores and production lines. Though, this perception of big supply has given the advantage to those consumers to easily look around within the web, to find exactly what they are looking for, in the right color, at an adequate shipping rate and above all, at the right price. This liberty that the customer has at its fingertips makes it somewhat impossible for certain retailers to keep up due to rapid changes in trends within the market. Therefore, organizations are forced to constantly look for best practices in order to optimize their operations and remain competitive (Šustrová, 2016).

To alleviate this issue, companies must deal with the challenging task of managing inventory in an efficient matter (Sachs, 2015). To a large extent, inventory can represent one of the most crucial functions within manufacturers and retailers since it tends to have great impact on their performance as a whole (Šustrová, 2016) Hence, the optimal level of stock needs to be determined. Though, one of the main dilemmas concerning inventory management relates to stochastic demand (Raa & Aghezzaf, 2005). Without any knowledge or indication of what is going to be bought, it is difficult for merchants to plan ahead and replenish stock in an efficient, accurate and timely matter, which understandably calls for a need of higher inventory levels and therefore generates more costs. Thus, it is essential to find the right predictive method to achieve efficiency (Dhini, Surjandari, Riefqi, & Puspasari, 2015).

A popular method such as Linear Regression has been recommended and used for decades in the industry to make numeric forecasts (Witten, Frank, & Hall, 2011). Unfortunately, one of the core assumptions of this method requires linearity between dependent and independent variables (Francis, 2003), which is often not the case in reality. On the other hand, machine learning techniques have allowed organizations to solve some of the demand uncertainty issues (Bertsimas, Kallus, & Hussain, 2016). Specifically, Artificial Neural Networks (ANNs) or simply Neural Networks (NNs), have gained popularity throughout the years due to their ability to solve complex problems where there is a high degree of uncertainty (Šustrová, 2016). Although providing fast and accurate results in some cases (Radzi, Haron, & Jahori, 2006), they tend to be difficult to interpret (Tso & Yau, 2007) and hard to setup (Forte, 2015) Most importantly, they can be very computational and time expensive compared to more traditional models (Han, Pool, Tran, & Dally, 2015).

 It has been argued that the method of choice to make predictions depends entirely on the data at hand. A comparable method to ANNs is Multivariate Adaptive Regression Splines (MARS). In Abraham & Steinberg (2001) these two methods were compared and results showed that MARS outperformed the NN model, whereas in Francis (2003), though while using a small dataset, the opposite was the case. Hence, as much information as possible should be gathered in order to find an appropriate method as based on this, the method might differ. Bertsimas et al. (2016) highlights the use of information generated through a company´s systems, as it can be remarkably useful for retail planning and thus useful for predictions. Furthermore, external factors which are not always considered such as weather conditions, could also be quite beneficial to decide on order quantities (Sachs, 2015).