)
Focused decision-making and data-driven tooling, his main focus is on pricing and recommendation while leading a team of machine learning engineers.
In this blog post, he explains how and why we combine machine learning and human intuitiveness to suggest the optimal price of trucking lanes.
One of the core challenges of a logistics company is getting the prices right. Thankfully, getting the prices right is very hard, so this is an interesting problem which will never become boring.
If you look at the information that goes into pricing from the point of a freight broker, you can divide it up into three types:
- Knowns: Route distance, time of pick-up, time of drop-off, vehicle type, …
- Known unknowns: General interest of a given carrier in one specific load, the effect of upcoming legislative changes, the effect of weather or pandemic conditions, the desirability of a given pick-up or drop-off point, sickness of drivers, operational hick-ups on the carrier’s or shipper’s side, road accidents, …
- Unknown unknowns: No idea. In hindsight, the shutdowns during COVID-19 qualify here. However, the main quality of the unknown unknowns is that at the point where you would care about it, you have no idea that you should.
As a modern digital freight forwarder, we have two types of agents that do pricing estimations: We have a human agent and we have a machine/pricing algorithm.
These agents compliment each other as each has distinct strengths and weaknesses.
Humans are good at high uncertainties
Human agents are great at navigating a space of low information, where they can make use of their experience and gut feeling to find a comparatively good solution for a problem, where there is no clear way to check the quality of a solution other than to try it out. Humans are also great at building relationships with other humans, leading to longer-lasting business relationships, which tend to be more profitable for both parties. However, two humans rarely agree.
If you look at the following chart, where one dot symbolizes one human price estimate and the mean human price estimate would be 1.0, you can see that humans wildly disagree. Ask two of them, get three opinions.
Machines are good at everyday stuff
Machine agents are great at doing repetitive tasks fast and reliably. They have no moods and gut feelings that cloud their judgment. They also have perfect memory and hindsight. They know everything that ever happened in the company and if they look at a problem, then their suggestion incorporates every bit of codified knowledge of the organization.
Think about how a human agent makes their pricing decision. They usually look up historical prices of a specific lane and then, trained by their experience, add/subtract a bit based on a gut feeling reflecting current market conditions. Both steps here are driven by limited information. An algorithm would more or less do the same, but it would do so with the experience of hundreds of human traders who never forgot anything and it would compare not only historical prices on the specific lane, but on all similar lanes, and some dissimilar ones if they are useful.
However, machines might also make insensible mistakes, if not trained properly and they are not able to work with non-codified, context-related, highly non-linearly complex data. E.g. a machine will not know how to price the impact of an upcoming rule-change for palette exchange.
Letting machines run unsupervised by humans is what caused a lot of extreme stock market destabilization, of which the 440-million-Knight-Capital-disaster might be the most famous example.
Our integrated present and future
At sennder, we want to combine the benefits and have the best of both worlds.
We want to leave the boring, repetitive everyday stuff to the machines, and let them have their fun with automated workflows, and have humans supervise at best.
And we want our human agents to specialize in the tricky stuff that machines can’t do. E.g. build relationships, care for more complicated set-ups, and come up with ingenious solutions to business needs.
So to close the circle in terms of the initially given classification of information, our philosophy of who should handle the different information types at sennder becomes:
- Knowns: Route distance, time of pick-up, time of drop-off, vehicle type, …
Who? Let the machine handle suggesting costs to human agents during negotiation or pricing loads in our fully automated order flows. - Known unknowns: General interest of a given carrier in one specific load, the effect of upcoming legislative changes, …
Who? Let the human handle finding solutions to hard-to-quantify problems. - Unknown unknowns: No idea.
Who? Have fast-adapting processes and structures to be able to quickly act should unknown unknowns become known.
While in practice, real processes are not as tidy and clearly separable. E.g you might also want to have machine suggestions for the tricky cases and support them with analytics and risk estimations. Or you might want the carrier to still be able to talk to a human agent for a given standard load.
However, the overview model above is the guiding line which we successfully have begun to implement in the past year. So far, the results have shown that we are on the right track and have confirmed our strategic direction.
Most of our business that goes to our open market platform is supplied with automated pricing. Theoretically, this means that there is no need for costly human interaction anymore. This automated pricing is amended by human agents monitoring the outputs and results and tweaking certain meta-screws. And tricky regions where our machine needs a helping hand are monitored more closely by human agents. So we are moving towards a setup where we fuse the best of both worlds: Rapid repetitive task replication by machines combined with the agile adaptation and complex thinking that humans excel in.
)
)
)