jed: just enough delivery
Increase sales and reduce sell-outs and returns of newspapers
and magazines through a smarter distribution.
One of the problems in logistics is to have the right amount of products at every point of sale. The distribution of single-copy newspapers is a typical example. Newspapers have an extremely short lifetime: with the arrival of a new edition, the previous one is worthless.
The obvious goal is then to match the delivery with the demand for each point of sale, in other words, to have no sellouts and no returns (or maybe one to make sure that there is no more demand).
This demand is different for different points of sale, for different days of the week, for different times of the year, and may also depend on the weather conditions, news content and so on. How do you know how many newspapers to send to each point of sale?
One of the (not so good) options is to work with a fixed set of rules. For example, "if among the last 3 deliveries there were at least 2 sellouts, the number of deliveries is increased by 10%". Fine, but how do you know that this is indeed a good strategy? Machine learning can provide a solution if historical sales figures are available. Suppose that there is a huge database with a few years of delivery and sales figures for all points of sale (many newspapers companies do have such a database). Somewhere in this database, there must be relevant information that can be used to determine the next delivery.
The solution is to train a model on this database: to find the model that best explains the observed sales figures. The same model can then be used to predict the upcoming demand and to determine the next delivery based on that demand and the strategy of the company. SMART Research BV has developed a software system based on this principle, specifically for the distribution of newspapers and magazines.
NEURAL NETWORK MODEL
The model is based on a neural network, with all kinds of special (Bayesian) features for optimal use of the huge amount of available data. Points of sales learn from their own data, but if this is insufficient, they extrapolate from the data of other points-of-sales. Furthermore, the model can learn and predict on the basis of data that is incomplete, i.e., data that has sell-outs or unknown sales. Besides learning on the basis of past sales data, other information such as holidays, events, promotion actions etc. can be incorporated to improve learning and prediction.
JED: just enough delivery
Jed is a tool to optimize deliveries. To do so, JED predicts not only the sales for each outlet, but also the uncertainty. As a simplified example, consider an outlet with predicted sale of 10 +/- 1 and another outlet with a predicted sale of 10 +/- 5. In a conservative sales approach, one may choose to deliver 9 and 5 copies respectively. In a more expansive approach, the delivery of 11 and 16 copies to these outlets could make more sense. It is also possible to compute optimal deliveries for a given total number of deliveries.
Within JED, the optimal deliveries for all (or a subset of all) points of sales can be computed based on your sale strategy.
JED has been implemented for De Telegraaf in The Netherlands and Público in Portugal. In tests at De Telegraaf, JED scored significantly better than their previous software systems that was based on more traditional techniques for time-series analysis: With the same amount of deliveries, JED realized 1.7% more sales.
JED IN YOUR ORGANIZATION?
JED relies on the availability of historical sales figures at the level of individual points of sale. This makes JED useful for newspaper companies and magazine distributors that:
Each newspaper or magazine has its own characteristics: a different reader group, different distribution channels and so on. Therefore, JED is custom made for each newspaper or magazine and installed and maintained by SMART Research BV at your offices.
- work with a right of return;
- have sales figures for at least one year;
- collect sales figures at the level of individual outlets.
Two articles on JED have been published in the IFRA Magazine one in the journal 'Neural Computing & Applications'. Furthermore there is a short brochure. Download the pdf files: