Artificial intelligence is already a reality. The miracle of the machines that think occurs daily and the amount of researchers and investment behind it suggests that it will have a great future. Machine learning, or automatic learning, is one of its branches and it is going to give a lot of talk in the world of logistics due to its ability to refine demand predictions, fight against the bullwhip effect and to help in planning supply chains.
What is machine learning
Machine learning is the analysis process that allows computers to extract knowledge from huge amounts of data, without needing to be specifically programmed where and how to search. Machine learning manages to generalize behaviors from data, information and examples. And, as a result of this reasoning and the new data it receives, it achieves “learning”.
How machine learning can help logistics
It is easy to imagine why the ability to draw insights through large amounts of data is very attractive for logistics.
Demand forecasting is one of the most delicate and complex tasks when it comes to taking care of a supply chain. With traditional models, forecasting requires intensive work, often manual or not fully automated, and often does not give the expected results.
What variables should we introduce? How do we weight its importance in our demand algorithm? What conclusions can we draw from our prediction errors? How much time, effort and money does it cost us to continue doing these calculations and modifications constantly?
These questions are what machine learning answers, by being able to continuously adapt without outside intervention. Machine learning gradually learns which variables affect our demand the most, adapting for future calculations, without the need for a human to reanalyze the entire process.
Data does not mean knowledge unless companies are able to find patterns, causes or connections that explain what has happened and/or what is going to happen. Machine learning looks for patterns, decision trees, similarities… and makes the extracted knowledge available to us, as well as being able to apply it on its own.
By contrast, traditional models are not built to automatically learn from new data. Machine learning allows computers to learn without having to be explicitly and continuously reprogrammed.
Fighting the bullwhip effect
The bullwhip effect consists of the mismatches created along the supply chain between demand predictions and actual demand. This effect can be very detrimental and generates traffic jams, supply problems, stock outs, etc. One of its main causes is the inability to share information completely and efficiently between the different actors in the supply chain.
There are several reasons for these problems with information: technological difficulty, reluctance of participating companies to make certain data public, the lack of alignment of interests between the different businesses, the lack of trust to establish long-term relationships, human errors when entering data and even obstacles to giving up power within the organizations themselves. And it is that, despite the fact that we talk about the supply chain, many times the logistical relations are produced simply from one link to another, without really integrating everyone.
This means that demand forecasting, in a scenario where information is perfectly shared, is subject to incomplete data. If it is already difficult to make accurate predictions with all the data, it is even more difficult to do it with scattered data. And in this situation, machine learning is able to work better and more successfully integrate into its mathematical models the lack of certainty present in all demand predictions.
Why is it difficult to predict demand?
Some may be reluctant to continue entrusting their business demand to artificial intelligence, but there are good reasons to do so. These are some of the obstacles that traditional methods have:
Large number of promotions in a large number of references or SKUs: promotional campaigns are especially difficult to predict: how much material is going to be sold, how much product needs to be sold to cover campaign expenses and/or promotional material that is not sold , etc. Which means that sometimes the errors in the prediction eat up all the profit obtained.
In addition, there are sectors such as retail in which promotions are constant and in lots of products, making it very difficult to manually analyze their results in each new promotional cycle. It is very difficult to anticipate the success of a campaign, a price drop, a new product, a new clothing line, etc. But machine learning is giving hopeful results.
Growing number of products: at the beginning of 2017, Amazon sold 398,040,250 different references. It is impossible and financially unsustainable to create human teams that have to manually analyze demand forecasts or even continually revise them in light of new data.
Long tail products: this is the name given to the products that sell the least in our portfolio. Even companies with much smaller volumes will suffer if they have to study each item in their catalog on a case-by-case basis. In addition, there are more and more of these products, increasing the difficulty of their analysis. A fact that is repeated with personalized products: less sales of each model and more number of models.
Tendency of companies to become more complex: companies, as time goes by, tend to diversify their offer. New products, new lines, new geographical areas -which are also a new variable to take into account-, etc.
Too much manual work: obtaining forecasts continues to require many hours of dedication and on many occasions it does not provide the expected results, resulting in the frustration of those who carry it out and of those who work attending to it. This wear and tear can cause failures to worsen and lead to a negative spiral that endangers the position of the personnel dedicated to these tasks. Furthermore, human subjectivity is also a source of error in the models.
Excess of data: the mere agglomeration of data means nothing. Businesses today generate an enormous amount of data and it can be hard to know what to pay attention to. And, most importantly, have the resources to do it. Machine learning helps us manage these volumes and automate their processing.
New data sources: Data sources are increasingly heterogeneous. A good example is social networks, whose temperature has already begun to be taken into account to anticipate the reception of the products that are going to reach the market according to the reactions of the users.
Taking into account tools like Twitter for our supply chain is something that is going to sound less and less strange, especially in some sectors. Again, machine learning makes it easier to take this data into consideration and assess what the significance is.
Other uses of machine learning
Until now we have focused on demand forecasting, but the applications of machine learning do not end there.
Another activity that generates a lot of data on a daily basis and is very likely to be optimized. The ‘machine learning’ evaluates for you the changes that are gradually taking place in your routes and of which you may not be able to realize.
The most basic model of these suggestions when we buy in an ecommerce is to offer us products that are very similar to the ones we are visiting. However, machine learning has more ability to establish relationships that are not obvious. Knowing how to find these relationships will also allow us to better calculate our demand.
If machine learning allows us to realize trends earlier, it can also give us valuable information on when a product’s life cycle is ending, if buyers have already started the change of season when buying clothes, etc. Having this information will make it easier for us to execute better pricing policies, discounts, replacements…
Knowing how to adjust inventories, looking for the ideal point between the safety stock and the reduction of stocks, has a direct impact on our income statement. Machine learning avoids excessive fear of stockouts or, at the other extreme, allowing ourselves to be seduced by their dangerous reduction, reducing subjectivity in our decisions.
Supplier evaluation and choice
How often do you evaluate your suppliers? Are you able to detect upward or downward trends in your service? Weak or strong points? Geographical areas where it is delivered better or worse? The machines can also learn to discern the best suppliers, the estimated delivery times -beyond the conditions promised by your supplier-, etc. Theoretically, machine learning can warn of the threat of a stock-out based on past experience and by analyzing current conditions.
If we accept that these technologies will know more reliably what service we can expect from the various providers. Why not let them make the decisions? Allow them, without the need for supervision, to choose the right suppliers for the different shipments.
Finding unexpected relationships can be a very important source of income and improvements. For example, Walmart discovered that there was a relationship between the weather and the type of meat that was sold in its stores. During warm, cloudy or windy days, more steaks were dispatched, while on days of hot and dry air the winners were the hamburgers.
This discovery led to an 18% increase in hamburger sales. This is another way to take advantage of the greater ease of machine learning to find buying patterns that are initially unrelated or between products that are not similar.
When will machine learning
The Gartner Hype Cycle is a classification that evaluates the point of evolution in which new technologies are: emergence of innovation, high point of expectations, a subsequent decline (performance below expectations) and, finally, the point at which which these new technologies begin to be really productive.
machine learning still had five to 10 years to go before it reached that widespread adoption that would start to pay off on a large scale. The most competitive supply chains in the coming years will be those that are able to extract valuable meaning from diverse sources. In a world with more and more data and with increasingly comprehensive and cheaper sensors and measurement devices, those who manage to differentiate useful information from noise will have the advantage.