Piero Ferrarese

This is a presentation based on a cloud systems administration company project. Since the data was private, we replaced the customer data with public data, i.e., Wikipedia page requests.

The task is to forecast the requests of a page to trigger an alert when the demand may become too high for servers to be handled. I employed Google Trends data as alternative data to enhance the accuracy of the predictions. The model used is an LSTM with encoder-decoder representation.

The results have been presented at a major German conference for machine learning applications.

The link shows the full list of publications in peer-reviewed journals produced during my time in academia. The Ph.D. thesis is publicly available, too, as it is published through the Göttingen University library.

I developed a time-series forecasting engine that blends state-of-the-art algorithms, from traditional statistics autoregressive models to decision trees and neural networks, to produce the best forecast for the end-user. The application can be deployed on the cloud and scale easily with increasing data. Third-party data can be suitably integrated to improve the forecasting power of the algorithms.