Smart cargo search and market price discovery: Trans.eu’s algorithms

- Author: Agnieszka Szubert

With the help of artificial intelligence, digital logistics platforms can recognize the patterns behind completed freight orders. If the volume of data is large enough, it can be used to calculate market-compliant price proposals that shippers, forwarders or carriers can use to significantly improve their negotiating position. As one of the largest digital logistics platforms in Europe, Trans.eu is one of the pioneers in this field.

Artificial intelligence can make people happy: dating sites discovered machine learning a long time ago to improve the search results of lonely hearts. With massive investments in advertising and technology, the leading providers in this segment have gained high market shares and user numbers. Parship alone is said to have more than 4.5 million members, giving the algorithms a solid database.

But it is not only private happiness that benefits from the achievements of digitization. The search for suitable transport partners and the determination of fair freight prices can also be optimized by smart algorithms. Here, too, the quality of the results depends on the mass of data available. The greater the number of freight orders completed each day, the more information the algorithm can use for its work – and the more accurate its results will be.

Learning from past data

The logistics platform Trans.eu, which operates throughout Europe, draws here on the data and daily inquiries of 40,000 users. This data pool is analyzed through machine learning. Trans.eu has developed a powerful algorithm that “learns” the behavior of market participants using data from the past. On this basis, a model is created that can be used to determine price proposals for future tours based on the learned rules.

In practice, this means that intelligent algorithms compare the stored offers and agreements and, on this basis, create proposals for current freight prices. These take into account factors such as the respective routes, time aspects, spot and standard offers, and other factors. This provides all parties involved with an ongoing insight into constantly changing market prices – an ideal basis for price negotiations. The other advantages of pricing with the help of machine learning are also obvious: the use of the algorithm saves the client (shipper or freight forwarder) a significant amount of time, helps avoid errors, and increases effectiveness. Another plus point is that the knowledge is no longer lost when employees change. Since Trans.eu’s algorithm has learned behavior based on data from the past, a new employee will also benefit from this knowledge.

The term “artificial intelligence”

Artificial intelligence (AI) is one of the megatrends of digitalization, but the exact meaning of AI is unclear to many stakeholders: in the human brain, there are about 100 billion neurons that are networked with each other. The transmission of information between neurons takes place via electrical impulses. This enables humans to learn, draw conclusions and think abstractly. In so-called “artificial intelligence”, neurons are replaced by artificial neurons and trained by means of algorithms. However, human intelligence is not replicated – machine learning is used to learn pattern recognition based on a variety of data.

Machine Learning can, for example, automatically learn a set of rules based on training data. This saves companies from having to manually create a model and the associated effort, such as defining rules, checks and interpretations. The quality of the training data is crucial for success.

Learning without memorizing

When developing a machine learning model, two tasks are particularly challenging. These include so-called feature selection, by which is meant the selection of a subset of relevant features of a data set from the numerous features of past transport jobs. This involves, for example, the selection of the destination, weight or transport type. The second challenging task is the so-called “overfitting/underfitting”. The model must be mathematically complex enough to learn human behavior. However, it should not learn by rote. The desired solution is called a generalizing model by machine learning engineers.