In the end, forecasting is nothing more than converting information we have into useful information we don’t have, as Agrawal, Gans and Goldfarb put it in their book Power and Prediction. We convert information we have: historical weather conditions and current measurements, through the application of complex models, into information we don’t have: the probability of rain tomorrow afternoon.
The data and the model
As we see in the example, forecasting basically requires two elements: data and a model. We live in a time when both elements are abundant. Never in the history of mankind have we had so much data and the speed at which the amount of information is growing is staggering. This trend will only accelerate with ever increasing storage capacity and more connected devices generating lots of data. On the modelling side, advances in AI give us access to increasingly complex, robust, fast and powerful models that can handle and learn from unthinkable amounts of information and data and can therefore deliver increasingly accurate predictions. In short, the ability to predict today is vastly, radically greater than it was just a couple of years ago.
Examples of prediction
We know many examples: predicting whether the patient in an X-ray has cancer, predicting whether the object moving in a camera is a person, predicting whether an image corresponds to a person, predicting whether a sound corresponds to a word. The models, which have learned using a lot of historical data, capture current data and with that give a probability to some event. The probability that this sound corresponds to this word is greater than X percentage, so the model can claim that it ‘heard’ that word. We know that today the accuracy with which speech recognition models identify words is higher than what we humans have.
Now, imagine what our processes could be like if we had the ability to predict. Today we have many of our processes always the same: vanilla for everyone. Or worse, more expensive and cumbersome than they could be: we always carry the umbrella because we cannot predict whether it is going to rain or not.
We make them more complicated than they could be, because we don’t differentiate what is going to happen to the cases going through the process. We could be much faster and more efficient if we could predict what kind of case the process is dealing with at any given time. It is as if we always carry a suitcase with clothes and accessories for all possible climates, instead of predicting and just carrying the right attire.
There are predictions that have a lot of value. For example, predicting whether a sales lead will turn into a real sale, whether the customer will not be at home when we arrive for service, how long maintenance will take at a customer’s home, or what materials I will need to repair the service on a mobile phone tower.
According to the prediction I can have different processes, not the same for all cases. This would generate a lot of speed and efficiency, as long as these processes are mostly automated, so that the variation in processes does not imply complexity in decision making. Nor does it take more time. For example, we should not take longer to ensure that we only carry what we need for the maintenance we have to do. In order to achieve this, the process of enlisting and defining materials needs to be automated. Another example, the way we treat a sales lead should be different according to the prediction of probability of purchase, we should have different processes and maybe different channels according to the prediction.
We are in an era where predictions are going to be increasingly varied, common and accurate. Incorporating predictions into our processes is the next step we need to take in order to move towards having increasingly autonomous, fast and productive processes. In this way we will offer a more agile and personalised experience to our customers while becoming a smarter and more efficient company.
