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

Job vacancies are a key indicator of labour market tightness and are closely monitored by policymakers for monetary and fiscal decisions. Official vacancy measures, derived from employer surveys (e.g., JOLTS in the US, ONS Vacancy Survey in the UK, and ABS Job Vacancies Survey in Australia), are accurate but reported with time lags and at monthly/quarterly frequencies, limiting responsiveness during rapid economic changes such as the COVID-19 shock. Online job postings offer higher-frequency signals but are unrepresentative due to incomplete coverage, cross-platform scraping, duplicates, postings representing multiple positions, persistence of filled postings, and differential representation across occupations and regions. Consequently, postings are only moderately correlated with true vacancies (e.g., correlation of 0.65 in Australia from 2018–2022 using a six-week window of new postings as a proxy for stock). The study’s objective is to develop an algorithm that uses the distribution of postings across many sources to generate an accurate, timely predictor of changes in official vacancies, improving on raw counts by down-weighting platforms with large, potentially noisier volumes.

Following the COVID-19 onset, researchers used online postings to track labour demand dynamics (OECD, 2021; ONS, 2021). Forsythe et al. (2020) documented a sharp contraction in new postings early in the pandemic, while Krumel et al. (2021) argued that focusing on the stock of postings showed a smaller contraction and highlighted the role of longer vacancy durations in raising total posting stock. Reviews (Carnevale et al., 2014; Duenser & Mason, 2020; Kureková et al., 2015; ONS, 2021) emphasize strengths of postings for granular analysis and shortcomings due to representativeness issues. Prior work has calibrated postings to official vacancy distributions to improve representativeness at a point in time (e.g., sectoral reweighting in Turrell et al., 2019; calibration in Beręsewicz et al., 2019). In contrast, this paper aims to predict official vacancy changes ahead of publication by aggregating signals from multiple sources, emphasizing equal weighting and robustness to outliers rather than ex-post calibration to published statistics.

Data: The study uses Adzuna Australia job postings (2017–2022; 8,460,963 new postings) comprising postings submitted directly to Adzuna and postings scraped from thousands of sources (employer sites, job boards, recruiters). To mitigate duplication, Adzuna performs deduplication, and the authors apply an additional filter removing suspected duplicates based on identical titles, locations, similar posting dates, and near-identical descriptions. Weekly counts of new postings are computed, and each posting’s source (website or recruitment agency) is recorded. Official vacancy counts come from the ABS Job Vacancies Survey conducted on the middle Friday of Feb, May, Aug, and Nov, with high response rates (>95%) and precise estimates. To relate flow (postings) to stock (vacancies), the total number of new postings in the six weeks preceding each survey date is used as a proxy indicator for the vacancy stock on that date.

Algorithm (ensemble of sources): For each pair of consecutive survey dates t-1 and t, identify all sources j with at least 100 postings in both periods. Compute each eligible source’s percentage change in postings Δx_jt = (x_jt − x_j,t−1)/x_j,t−1. Apply k% winsorization to Δx_jt across sources (replace values below the k-th percentile with the k-th percentile and values above the (1−k)-th percentile with that cutoff). Estimate the percentage change in vacancies as the mean of the winsorized Δx_jt across sources. The algorithm assigns equal weight to each source’s signal (subject to winsorization), based on evidence that prediction accuracy is not related to source volume beyond very small sources (<100 postings). Extreme values are winsorized to reduce undue influence without discarding data, improving robustness.

Parameter selection: Using training data (May 2017–Feb 2018), the authors examined absolute prediction errors across winsorization levels and found large errors without winsorization, which were reduced substantially with k between 5–15%. They therefore test k = 5%, 10%, and 15%. The minimum source count threshold (100) and k are data-dependent and should be tuned similarly for other datasets.

Study design and evaluation: Training period: May 2017–Feb 2018. Test period: May 2018–Aug 2022 (18 quarters). Predictions are formed 6 weeks prior to the publication of official ABS estimates by applying the algorithm to the six-week posting window before each survey date. Performance is compared against the benchmark method using raw total counts of postings (across all sources) to estimate percentage changes in vacancies. Accuracy is assessed via correlation with ABS changes and mean absolute prediction error across 18 quarters at national and state levels (excluding small states/territories with too few sources for ensemble averaging). Robustness to winsorization levels is also evaluated by varying k from 0% to 50% (k = 50% corresponds to using the median across sources).

• The ensemble algorithm’s predicted quarterly percentage changes in vacancies track official ABS changes closely over 18 quarters (May 2018–Aug 2022). At the national level, correlation is 0.95–0.96 (vs 0.65 using raw postings counts).
• State-level performance is also strong, though it declines with smaller ensemble sizes: New South Wales 0.85–0.86 (raw 0.69), Victoria 0.92 (raw 0.77), Queensland 0.84–0.89 (raw 0.54), Western Australia 0.84–0.88 (raw 0.71), South Australia 0.74–0.77 (raw 0.52). Application to the smallest regions was infeasible due to insufficient sources.
• Mean absolute prediction errors (percentage points) are substantially lower than the raw counts approach. At the national level: 4.7–5.6 vs 14.1. NSW: 10.6–11.2 vs 15.3; Victoria: 6.8–7.0 vs 13.1; Queensland: 8.1–9.2 vs 14.3; Western Australia: 10.1–11.6 vs 14.9; South Australia: 16.4–17.8 vs 20.5.
• The algorithm captures major quarterly movements more accurately than raw counts and provides a close approximation to the vacancy index trajectory over 2018–2022.
• Results are robust across winsorization levels; equal-weight averaging across sources improves accuracy relative to raw counts even with no winsorization (k=0%). Optimal k is around 5–10% for this dataset.
• The method delivers predictions 6 weeks earlier than official publication and can be updated at higher frequencies (daily/weekly).

The findings show that treating each data source’s percentage change in postings as an equal-weight signal and robustly averaging them yields substantially better predictions of official vacancy changes than aggregating raw posting counts. This addresses the central challenge of unrepresentativeness and noise in big postings data, especially from large platforms prone to duplication and scraping idiosyncrasies. By dampening outliers and avoiding dominance by high-volume sources, the ensemble improves alignment with true vacancy dynamics. The approach provides earlier and higher-frequency insights (6 weeks ahead of official releases), supporting more timely policy responses to shifting labour demand. The concept generalizes to other multi-source time series where reliability varies, offering a simple and robust transformation of heterogeneous big data into accurate indicators of underlying economic quantities.

The paper presents an ensemble, winsorized equal-weighting algorithm that transforms multi-source online job postings into a timely, accurate predictor of official vacancy changes. Applied to Australia (2018–2022), it strongly outperforms the raw counts benchmark, closely tracking ABS vacancy movements and reducing mean absolute errors. For broader adoption, postings providers should expose source-level metadata. With these data, policymakers can monitor labour market conditions in near real-time. Future work should test performance-based weighting schemes, assess risks of overfitting, and develop methods suited to finer spatial/sectoral granularity where ensemble sizes are small.

The algorithm’s accuracy depends on relative stability in recruiters’ posting behavior and platforms’ scraping/deduplication procedures. Systematic changes (e.g., increased duplicate posting activity or shifts in scraping access/policies) can introduce noise not related to true vacancies. Equal weighting across many sources mitigates the impact when only a few sources are affected, but accuracy declines as more sources are impacted. Limited ensemble sizes in small regions constrain applicability and reliability at fine geographic levels.