As part of the annual meeting of the International Symposium on Forecasting – Virtual ISF 2021 | 27 au 30 Juin 2021, PREDICONSULT will be organizing two sessions where five papers will be presented on the topic of “How to predict with artificial intelligence”.
The meetings, which will bring together academic researchers and business practitioners, will provide opportunity for learning more about new forecasting methods based on artificial intelligence (including machine and deep learning).
Focus on two topics about the use of non-numerical information for the purposes of modeling and forecasting.
The use of Textual Data for Time Series Forecasting
The publications of various organizations (companies, institutions such as ministries and others such as public and private laboratories), information on websites, blogs and social networks are all very rich and relevant sources of information which can enhance the quality of modeling and forecasting. However, today, the information from these sources and “inputs” is barely included in the algorithms for predicting time series and any other data series (cross-section, for example). In one of the presentations, the authors of a study use textual information included in daily weather reports available in France and the United Kingdom as well as in tweets, in order to forecast electricity consumption.
Using exclusively weather reports, the authors of this study are able to predict loading time series with sufficient accuracy to be used to replace missing data. In addition, Twitter data relating to specific keywords helped adapt to new patterns of electricity consumption observed during the COVID-19 lockdown period in France, significantly reducing forecast errors in operational models.
A picture is worth a thousand data points
“A picture is worth a thousand data points”. This is the title of a presentation to be made in one of the two sessions. The authors will present an approach to forecasting time series from images based on Deep Learning. The latter has become an approach and common practice in several scientific fields, including computer vision and natural language processing, among others. Nevertheless, deep learning has been adopted at a much slower pace in the field of forecasting. In addition, existing deep learning forecasting approaches treat time series data as numerical vectors, thus not directly and to the fullest extent of recent advances in deep learning.
The authors of this study therefore devised a new approach to deep learning forecasting that transforms the traditional task of time series forecasting into a computer vision. The latter is an area of artificial intelligence that enables computers and systems to extract meaningful information from digital images (and all other visual inputs) and make recommendations accordingly. First, instead of using digital representations as input to forecasting models, they consider their visual representations as images. Second, neural networks inspired by image recognition neural network architectures (eg, ResNet) are responsible for producing point predictions from the visual representations in question. The results indicate that image-based time series forecasting methods outperform traditional methods, both statistical and machine learning.
For more information about the program and all papers that will be presented, visit the IIF’s website. To participate in this conference, you must first register here.