Zero-Shot Time Series Forecasting of the Online Gig Economy Using the CHRONOS Pretrained Language Model

Lead Researcher(s): Dharyll Prince M. Abellana
Status: Published

Abstract/summary: This study introduces a zero-shot forecasting method for the online gig economy applying the CHRONOS pre-trained language model to the online labor index (OLI) data. The paper concentrates on forecasting demand for six freelance job categories using historical weekly data from January 2017 to August 2024, without retraining the model specifically for these tasks. Although the one-step-ahead (weekly) forecasts demonstrate highly satisfactory accuracy, with a mean absolute scaled error below 1 for most freelance occupations, there is a decline in performance for multi-step forecasts (at least one month ahead), shown by increased errors. This result underscores the importance of pre-trained models in learning semantic information that can be easily transferred to other applications, especially in zero-shot setups. This paper makes a substantial contribution to the field by using OLI to enhance our understanding of predicting the demand for freelance jobs based on structural patterns in economic indicators. This is crucial given the volatility and distinct behaviors of the online gig economy. Furthermore, it illustrates the potential of using pre-trained transformer-based language models in time series forecasting, even with limited historical data, proving that advanced pre-trained models can still produce valuable predictions.

Keywords:

• Zero-shot time series forecasting
• pre-trained language model
• online gig economy
• online labor index