The year 1999 marked the start of the European higher education reform called the Bologna Process. One of its goals is a competency-based system. After graduation, each student has a series of competences in solving specific problems.
This approach is based on the assumption that problems will forecasted and the problem solvers will be there. The spectacular advancements in machine learning, leading to the term of AI revolution, give us confidence that indeed, we have the technology to forecast future problems. We have deep learning approaches towards forecast which are derived from the very powerful algorithms used in natural language generation (generating future values of a sequence of numbers is much like generating next word in a sentence), but there is a hidden trap in these advanced techniques.
One natural question is do these powerful algorithms ever fail to provide good results and what to do if this would happen. In fact, such events happened many times in history, powerful armies were defeated on the battlefield, but I would refer to a moment in the history of football.
At the 1958 FIFA Word Cup, all the media was focused on the The Soviet Union’s team, because they were coming with a novel, scientific approach towards football. The acted like a deep leaning approach towards forecast and they were prepared for many game conditions, except the unexpected. In the game with Brasil, The Soviet Union had no solution against players like Pele, Garrincha and Vava and they were dominated from the very beginning.
Our results on restaurant occupancy forecast show that under an unexpected event like the COVID 19 pandemic, an advanced deep learning approach considering features derived from weather data and calendar fall harder than a time series approach.
This is why we recommend having a deep learning approach together with one based on timeseries, It is just having a regular guy to double Superman for the case when kryptonite in present.
The is a second hidden trap to forecast: evaluating. We cannot measure how good the forecast is, but only how good the forecast was, after the event happened.
In the 1985 film Back to the future, when Marty travels back in time, he meets his parents and things turn bad: his mother falls in love with him and his father is constantly haressed by the school bullies.
There is the iconic ballroom scene, when things got fixed. Marty replaces the injured guitar player, because his parents first kissed at that ball, and with no music they cannot dance, hence they cannot kiss and fall in love, so he would disappear.
In fact, Marty took a management decision based on a forecast. He had a picture with him and his siblings and characters started to vanish from the picture forecasting his disappearance. Based on this forecast, he took the decision to replace the injured musician and play himself the guitar.
However, if we would like to evaluate the accuracy of the forecast, it turns out to be bad, because Marty survived in spite of the fact that the vanishing characters from the picture suggested otherwise.
From this perspective, the ideal fate of any forecast is to be bad. If the forecast of a restaurant occupancy for the next days is bad, the manager is able to take actions and come with a promotion or advertising campaign. The occupancy increases, but afterwards, when measured, the forecast turns to be less accurate since it predicted a low occupancy.
There is only one way to assess how accurate a forecast is: by comparing it to what a human expert would predict, since exactly this difference is what gives a return of investment.
Here at Next It Project we offer forecast solutions tailor made for any needs a customer might have.