Technology Intelligence Platform

Forecasting patent filings

In the first three articles in our series on the Technology Intelligence Platform (TIP), we explored technological fields and their evolution, focusing on emerging technologies and Europe's role as a driver of cutting-edge innovation. By leveraging PATSTAT data and combining it with TIP’s data processing and visualisation capabilities, the accompanying notebooks provided powerful insights into the evolution of technical fields. These analyses highlighted emerging technologies through visualisations designed to support decision-makers and stakeholders.

Fig.1: 5 years forecast (2023-2027) for applications filed to EPO members states. GDP values of the members states is included in the training phase of the algorithm.

In this edition, we focus on time series forecasting for patents. By incorporating external libraries and implementing open-source models, we use historical filing and economic data to generate mathematical predictions of filing activity. This article is accompanied by two companion notebooks, designed to be read sequentially. Please refer to the “Further information” section at the bottom of this page.

DataLoading

The first notebook, DataLoading, focuses on data preparation and preliminary analysis. It loads three key datasets: PATSTAT, EPA, and the Geographic Dataset, which categorises applicants by region (United States, Republic of Korea, EPO member states, People’s Republic of China and Japan). Patent and geographic data are combined with economic indicators as exogenous factors for modelling. TIP’s data visualisation tools are used to highlight trends, seasonality and structural patterns, including the correlation between GDP and filing volumes.

N-BEATS

The second notebook, N-BEATS, implements a state of the art, deep learning model for time series forecasting. It demonstrates model setup, training and the integration of exogenous variables such as GDP for varying forecast horizons. Performance metrics and visualisations assess the accuracy and relevance of the model’s predictions.

These two notebooks do not aim to provide definitive conclusions about future filing trends but rather serve as an introduction to advanced time series applications within TIP. Readers are encouraged to clone the repository, explore the notebooks and conduct their own analyses.

Keywords: data processing, visualisation, patent data analysis