This thesis investigates various time-series methods to forecast the evolution of different technology areas in the field of artificial intelligence (AI), based on historical patent application data from the United States. Understanding trends in AI, using various time-series methods can help inform R&D planning and innovation policy.
This study used open-source patent application data from the PatentsView database by the United States Patent and Trademark Office (USPTO) and a taxonomy for AI-based domains to categorise patent classification codes into technology areas of interest. Patent classification codes based on the Cooperative Patent Classification (CPC) scheme. Moreover, several exogenous variables relating the public and private sector spending in AI were also used for the predictions.
Forecasting was performed with statistical methods like SARIMAX, as well as additive methods such as the Prophet model. Supervised learning methods such as XGBoost were also used by augmenting the training data with temporal features. The predictive performance of all models was compared over monthly and quarterly frequencies. Evaluation metrics such as the Root Mean Squared Error (RMSE) and the Mean Absolute Percentage Error (MAPE) were employed to validate the model, along with a novel metric particularly suited for time series analysis, the symmetric Mean Absolute Percentage Error (sMAPE).
The findings demonstrated that the XGBoost model trained on augmented data outperformed the other models. The results also analysed the differences in prediction accuracy between monthly and quarterly forecasts, concluding that quarterly predictions yielded superior performance. Finally, the study validated the forecasts using various error metrics, enabling comparative evaluations across models and technological areas.
