OECD Compendium of Productivity Indicators 2025

9. Insights on productivity developments in 2024

Context and key findings

Productivity is central to economic policy. It underpins decisions on growth strategies, competitiveness and living standards. Yet official labour productivity data are published with a one- to two-year delay. Quarterly measures of GDP per hour worked can offer earlier insights, but they are not uniformly available across OECD and accession countries. When they exist, they are often subject to substantial revisions and sometimes provide only limited information on final annual estimates of productivity growth.

This chapter presents experimental nowcasts of annual labour productivity growth for 2024 across 40 OECD and accession countries. It builds on the approach of Dorville et al. (2025[1]), which offers improvements over standard alternatives in three respects. It is set in a panel setting, thus allowing to overcome the lack of data over a long time span, it considers a range of models, including dynamic factors models and machine-learning methods, and optimally exploit high frequency data through mixed-data sampling (Ghysels, Santa-Clara and Valkanov, 2004[2]; Borup, Rapach and Schütte, 2023[3]). The resulting estimates can provide policymakers with more timely insights into productivity dynamics, while highlighting areas of heightened uncertainty.

Key findings

• Labour productivity growth is estimated to have remained weak across the OECD in 2024, despite a modest rebound in Europe. The average estimated growth in labour productivity for OECD countries (excluding Türkiye) is 0.4%. The results suggest labour productivity gains in 2024 were modest in most regions: Europe (0.4%), Asia (0.2%), and North America (0.2%).

• Labour productivity growth nowcasts varied widely across countries, with few clear leaders. Only a handful of countries exhibited growth beyond 2.5% in 2024, including Croatia, Finland, Luxembourg and Chile. However, these figures come with sizeable uncertainty, especially Croatia and Chile, and should be interpreted cautiously.

• The average labour productivity of G7 countries is estimated to have declined slightly (- 0.3%). The nowcast points to a clear divergence between the United States – with estimated labour productivity growth of 1.5% – and the rest of the G7, which are expected to have experienced negative or near-zero productivity growth.

• Nowcasts for 2024 reinforce the view that the post-pandemic productivity recovery remains fragile and uneven. While estimates continue to converge towards historical averages, country-level nowcasts remain heterogenous and are surrounded by large uncertainties.

What happened in 2024?

Labour-productivity growth is expected to have remained weak in 2024. For the 40 OECD and accession countries included in the analysis, the nowcast points to an average increase of roughly 0.2 percentage points, a small improvement on the mild contraction recorded in 2023, but still far below pre-pandemic levels (Figure 9.1). For the G7 as a whole, the picture is mildly negative (-0.3%).

Taken together, the 2024 nowcasts reinforce the view that the post-pandemic productivity recovery remains fragile and uneven. Among regions, Europe appears closest to showing signs of a rebound (Figure 9.2). While the 26 European OECD countries are projected to have seen labour productivity growth by about 0.4%, this masks a pronounced split: strong gains in Finland, Luxembourg, and Croatia contrast with notable declines in large economies including Germany and Italy.

OECD countries outside Europe are estimated to have experienced only modest gains in labour productivity growth in 2024. Korea and Israel would have registered modest gains, whereas Japan would have slipped slightly into negative territory and Australia showed little change, leaving the regional average at only 0.2%. In 2024, North America is estimated to have exhibited a two-speed pattern, with solid productivity gains of about 1.5% in the United States contrasting with a small fall in Canada and a larger drop in Mexico, resulting in a regional mean of approximately 0.2%. South America would have experienced the strongest labour productivity growth at roughly 1.2%, primarily due to Chile. Yet, the region also exhibits the broadest confidence intervals, reflecting considerable uncertainty.

At the country level, dispersion is high (Figure 9.3). Twenty-three of the 40 OECD economies are projected to have seen positive labour productivity growth in 2024, but most of them by less than one percentage point. Only Croatia, Luxembourg, Finland and Chile break the 2.5% threshold, and all four come with sizeable uncertainties that call for caution when interpreting the numbers. Sixteen economies are expected to have experienced declines in labour productivity, led by Austria and Germany, while Ireland is projected to have experienced no change. Most point estimates fall within a fairly wide one percentage point confidence band.

Figure 9.1. Labour productivity growth nowcast for 2024

OECD average, Per cent

Note: The blue shaded area displays the 95% confidence intervals associated to the point estimate for the OECD aggregate. They are computed as $\left[\widehat{\mu }\pm p_{97.5}\widehat{\sigma }/\sqrt{N}\right]$ where $\widehat{\mu }$ corresponds to the unweighted average prediction and $\widehat{\sigma }$ the average Root Mean Squared Error. $p_{97.5}\approx 1.95$ is the 97.5^th percentile of the normal distribution, and $N=40$ the number of countries in the respective region. The red dotted line depicts the consensus nowcast, i.e., the average prediction of the five best models. Grey shaded areas represent global crisis episodes. Türkiye is excluded from the sample due to a large share of missing input data.

Source: Dorville et al. (2025[1]).

Statlink

Figure 9.2. Labour productivity growth by region, nowcast for 2024

Region-specific unweighted average of labour productivity growth, Per cent

Note: The grey area refers to the 95% confidence bands associated with each predicted point. They are computed as $\left[\widehat{\mu }\pm p_{97.5}\widehat{\sigma }/\sqrt{N}\right]$ where $\widehat{\mu }$ corresponds to the unweighted average prediction and $\widehat{\sigma }$ the average RMSE, $p_{97.5}\approx 1.95$ is the 97.5^th percentile of the standard normal distribution, and $N$ the number of countries in the respective region. Red dotted lines correspond to the consensus prediction. Asia: Korea, Japan, Australia and New Zealand. Euro area: Austria, Belgium, Croatia, Estonia, Finland, France, Germany, Greece, Italy, Ireland, Latvia, Lithuania, Luxembourg, the Netherlands, Portugal, Slovak Republic, Slovenia and Spain. Other Europe: Denmark, Bulgaria, Romania, Switzerland, Hungary, Iceland, Poland, Sweden and Norway. North America: United States, Canada and Mexico, South America: Chile, Colombia and Costa Rica.

Source: Dorville et al. (2025[1]).

Statlink

Figure 9.3. Country-level labour productivity growth, nowcast for 2024

Predicted labour productivity growth per country, Per cent

Note: Vertical lines refer to the 90% confidence intervals associated with each predicted point.

Source: Dorville et al. (2025[1]).

Statlink

The underlying forces driving the predictions

Feature importance statistics reflect how much the model depends on each predictor to generate the nowcasts. Specifically, they quantify how much each predictor improves the accuracy of the nowcast on average.

First, activity-related indicators, such as those tracking industrial production and capacity utilisation, emerge as the most influential in driving labour productivity growth nowcasts across most countries, especially in advanced economies (Figure 9.4). The external environment also plays a role: external indicators (e.g. trade and foreign orders) are important in smaller, open economies like Czechia and Lithuania, while financial and uncertainty variables contribute more substantially in emerging or volatile contexts (e.g. Romania) or financial hubs like Luxembourg. The nowcast models for some countries (e.g. France, Canada, Greece) show a heavy reliance on global variables, indicating strong sensitivity to worldwide economic trends, while others (e.g. Switzerland, Slovak Republic, Iceland) rely more on country-specific indicators (Figure 9.5). In contrast, labour market indicators remain broadly secondary but can play a role, particularly in transition economies where changes in hours worked and employment exert greater influence.

Figure 9.4. Feature importance by domains

Average feature importance by domains

Note: Feature importance is measured as the average decrease in prediction error allowed by each feature. The resulting values are then normalised. Values are obtained from country-specific model of the imposed-weights MIDAS method and calculated over the period 2015-23.

Source: Dorville et al. (2025[1]).

Statlink

Figure 9.5. Feature importance by variable scope

Average feature importance by country-specific and global features

Note: Feature importance is measured as the average decrease in prediction error allowed by each feature. The resulting values are then normalised. Values are obtained from country-specific model of the imposed-weights MIDAS method and calculated over the period 2015-23.

Source: Dorville et al. (2025[1]).

Statlink

Data sources

OECD (2024), Labour productivity forecast (indicator), https://doi.org/10.1787/cb12b189-en.

References

[3] Borup, D., D. Rapach and E. Schütte (2023), “Mixed-frequency machine learning: Nowcasting and backcasting weekly initial claims with daily internet search volume data”, International Journal of Forecasting, Vol. 39/3, pp. 1122-1144, https://doi.org/10.1016/j.ijforecast.2022.05.005.

[1] Dorville, Y. et al. (2025), “Towards more timely measures of labour productivity growth”, OECD Statistics Working Papers, No. 2025/01, OECD Publishing, Paris, https://doi.org/10.1787/436ecbb5-en.

[5] Fosten, J. and S. Nandi (2023), “Nowcasting from cross‐sectionally dependent panels”, Journal of Applied Econometrics, Vol. 38/6, pp. 898-919, https://doi.org/10.1002/jae.2980.

[7] Ghysels, E. and H. Qian (2019), “Estimating MIDAS regressions via OLS with polynomial parameter profiling”, Econometrics and Statistics, Vol. 9, pp. 1-16, https://doi.org/10.1016/j.ecosta.2018.02.001.

[2] Ghysels, E., P. Santa-Clara and R. Valkanov (2004), “The MIDAS touch: Mixed data sampling regression models”, CIRANO Working Papers.

[9] Jaax, A., A. Mourougane and F. Gonzales (2023), “Nowcasting services trade for the G7 economies”, The World Economy, Vol. 47/4, pp. 1336-1386, https://doi.org/10.1111/twec.13477.

[8] Kotchoni, R., M. Leroux and D. Stevanovic (2019), “Macroeconomic forecast accuracy in a data‐rich environment”, Journal of Applied Econometrics, Vol. 34/7, pp. 1050-1072, https://doi.org/10.1002/jae.2725.

[6] Mourougane, A. et al. (2023), “Nowcasting trade in value added indicators”, OECD Statistics Working Papers, No. 2023/03, OECD Publishing, Paris, https://doi.org/10.1787/00f8aff7-en.

[4] Woloszko, N. (2020), “Tracking activity in real time with Google Trends”, OECD Economics Department Working Papers, No. 1634, OECD Publishing, Paris, https://doi.org/10.1787/6b9c7518-en.

Annex 9.A. Methodology

This annex provides a short overview of the approach to nowcast labour productivity growth, measured in terms of total hours worked, for 40 OECD and accession countries. A key specificity of the approach is to run models in a panel setting. As the target variable is annual and only available for the years 1995-2022, a panel approach increases the robustness of results (Woloszko, 2020[4]; Fosten and Nandi, 2023[5]; Mourougane et al., 2023[6]). In addition, models are estimated in a quasi real time accounting for publication lags, identified using lags observed in 2023.

The main steps are summarised in Annex Figure 9.A.1.

Annex Figure 9.A.1. Main steps of the empirical strategy

Source: OECD illustration.

Data

The dataset spans national-accounts aggregates, labour-market statistics, trade and business-cycle proxies, as well as a range of financial, geopolitical and policy-uncertainty measures for forty OECD countries. Because many of these series are not published in real time, the most recent observations are “now-filled”: monthly indicators are prolonged with ARIMA projections, while higher-level series are extended with dedicated machine-learning models. The result is a balanced, country-by-frequency panel that replicates the information set that would have been available to forecasters at each historical point in time.

Model selection

Labour-productivity growth, the target variable, is annual and observed only for 1995-2022. Estimating models in a country panel therefore increases the effective sample size and exploits cross-sectional variation, an approach that has proved advantageous in earlier work by Woloszko (2020[4]), Fosten and Nandi (2023[5]) and Mourougane et al. (2023[6]). Publication delays are handled in quasi real time by imposing the calendar lags that prevailed in 2023 on every historical vintage. High-frequency information is folded into the annual horizon through Mixed-Data-Sampling regressions (Ghysels, Santa-Clara and Valkanov, 2004[2]). Three MIDAS variants are considered: an unrestricted version in which high frequency variables are included with no transformation; a restricted version in which high frequency variables follow a parsimonious lag polynomial; and an imposed-weight specification that aggregate high frequency variables with an equal weight, the later can be interpreted as a special case of the restricted model.

On top of the MIDAS bridge, six forecasting algorithms are explored: a dynamic-factor model that summarises co-movements across the data panel, penalised linear regressions (LASSO, Ridge and Elastic Net) that offer automatic variable selection; and tree-based machine-learning techniques, random forests and gradient-boosted trees, that capture non-linearities and interactions. Combining each algorithm with the three MIDAS lag-structures yields eighteen candidate specifications; an AR(1) in annual growth serves as a traditional benchmark, bringing the total to nineteen. All machine-learning hyper-parameters are tuned by grid search.

Performance is evaluated on one-year-ahead forecasts with the Root Mean Squared Error (RMSE). Rather than imposing the same specification on every economy, a single best-performing model is selected for each country. To guard against in-sample over-fitting, the evaluation follows an expanding-window cross-validation that starts in 2015 and moves forward one year at a time. For restricted MIDAS regressions, the polynomial lag structure and the regression coefficients are jointly estimated with the profile-likelihood technique of Ghysels and Qian (2019[7]), which retains estimation efficiency despite the added non-linearity.

Nowcasting

The selected country-specific models are applied to generate nowcasts for 2024. In parallel, a “consensus” estimate, the simple average of the five best individual models, is computed following the evidence that such ensembles reduce idiosyncratic model noise (Kotchoni, Leroux and Stevanovic, 2019[8]; Jaax, Mourougane and Gonzales, 2023[9]). Uncertainty is quantified with symmetric confidence bands whose half-width equals one model-specific RMSE, interpreted as the forecast-error standard deviation. Finally, for tree-based models, feature importance, measured as the decrease in squared error averaged over nodes, provides insight on the features underlying the nowcasts.