An algorithm for predicting job vacancies using online job postings in Australia

Accurate and timely job vacancy statistics are crucial for policymakers to respond effectively to changing economic conditions. Traditional methods, such as employer surveys (e.g., the US JOLTS, UK Vacancy Survey, Australian Job Vacancies Survey), provide highly accurate data but suffer from infrequent reporting and time lags. Online job postings offer a high-frequency alternative, but their accuracy is limited by various factors including incomplete coverage, duplicate postings, and variations in posting behavior across different platforms. This study aims to develop an algorithm using online job postings data from Adzuna Australia to accurately predict changes in job vacancies, addressing the limitations of both traditional surveys and raw online posting counts.

Previous research has explored using online job postings to measure labor market dynamics, particularly during the COVID-19 pandemic. Studies like Forsythe et al. (2020) highlighted the potential of using new job postings to track labor demand, while Krumel et al. (2021) emphasized the importance of considering the total stock of postings, not just new ones. Several reviews have examined the strengths and weaknesses of using online job postings, emphasizing their benefit in detailed analysis but acknowledging limitations like unrepresentativeness. To address the representativeness issue, some studies have calibrated online postings to official statistics (Beresewicz et al., 2019; Cammeraat & Squicciarini, 2021; Turrell et al., 2019), but these focus on improving representativeness at specific points in time rather than predicting future vacancy counts. This study differentiates itself by focusing on developing a predictive algorithm using job postings data.

The study utilizes data from Adzuna Australia, an online job postings aggregator, and the Australian Bureau of Statistics Job Vacancies Survey. The Adzuna data includes 8,460,963 new job postings from 2017 to 2022, with efforts to filter out duplicates. The official job vacancy counts are quarterly, published with a six-week lag. The algorithm treats each source's percentage change in new postings as an equally weighted signal of the percentage change in vacancies. This ensemble approach aims to reduce the influence of large sources prone to errors and leverage the accuracy of smaller, more reliable sources. A k% winsorization step is used to mitigate the impact of outliers. The study evaluates the algorithm's performance using data from May 2017 to February 2018 for algorithm design and May 2018 to August 2022 for testing. The algorithm's accuracy is assessed by comparing its predictions with the official vacancy counts, comparing it to the accuracy of using raw posting counts.

The algorithm demonstrates high accuracy in predicting changes in job vacancies. At the national level, the correlation between the algorithm's predictions and the official vacancy changes is 0.95-0.96 across the 18 quarters in the test set (compared to 0.65 for the raw posting count method). Similar results are observed for Australia's largest states, although performance decreases as the number of sources (ensemble size) diminishes. The algorithm consistently delivers significantly smaller mean absolute prediction errors than the raw posting count method (Table 1 and Table 2). Visualizations (Figures 4 and 5) demonstrate the algorithm's ability to accurately capture the trajectory of job vacancy changes over the 4.5-year period, outperforming the raw count approach. The algorithm's accuracy proves relatively robust to different levels of winsorization (Figure 6), with optimal results around 5-10% winsorization.

The findings demonstrate the efficacy of the proposed signal averaging algorithm in creating an accurate and timely indicator of job vacancy changes using online job postings. The superior performance compared to using raw posting counts highlights the importance of equally weighting signals from diverse sources. This approach effectively mitigates the influence of unreliable sources, improving the accuracy of the overall prediction. The algorithm's ability to provide estimates six weeks ahead of official statistics provides significant value for policymakers, enabling faster responses to labor market shifts. The method's applicability extends beyond Australia, and the methodology itself provides a framework adaptable to other big time-series data from multiple sources.

This study presents a robust algorithm that accurately predicts changes in job vacancies using online job postings data. The ensemble approach, combined with winsorization, provides a significant improvement over using raw posting counts. This method offers timely and high-frequency insights into labor market conditions, aiding policymakers in making informed decisions. Future research could explore adaptive weighting schemes based on source accuracy and extend the method to finer levels of regional and industrial aggregation.

The algorithm's reliance on consistent recruitment agency posting behavior and platform scraping procedures is a potential limitation. Significant changes in these factors could negatively impact the algorithm's accuracy. The study's focus on larger regions limits its applicability to smaller regions with fewer data sources. Future work should explore alternative methods suitable for smaller regions and industries.