OECD Compendium of Productivity Indicators 2024

2. Insights on productivity developments in 2023

Context and key findings

Developments in productivity have become more and more uncertain, as several shocks, including the COVID shock, the energy crisis, and more recently heightened geo-political tensions hit economies, with potential long-term scarring effects for some of them (OECD/APO, 2022[1]). This has added to long-term trends such as population ageing, declining competition and stalling globalisation, which can also hamper productivity developments. At the same time, digitalisation, Artificial Intelligence and the transition to a green economy offer opportunities to revive productivity growth (OECD, 2023[2]). Getting preliminary insights on most recent developments in productivity growth is thus useful to inform policymaking, identify policy needs or monitor the effects of policies.

Labour productivity statistics are usually released with a lag of one or two years, posing challenges for timely analysis and policy design. This chapter provides information on labour productivity (as measured by GDP per hours worked) developments in 2023 relying on experimental estimates for 38 OECD countries. Estimates are derived using a range of machine learning models, of varying accuracy across countries and should be interpreted with caution, especially for countries where confidence bands around estimates are large.

Key findings

• Overall, experimental estimates point to labour productivity growth of about 1.4% in 2023 on average across OECD countries (excluding Türkiye), close to the average over the long period (2001-2019). These estimates are surrounded by large uncertainties.

• Labour productivity growth is estimated to have been modest in most OECD European and Asian countries in 2023 (1.5% in Europe and 1.8% in Asia on average across countries).

• A sizeable increase in productivity growth from -1.6% in 2022 to 1.5% in 2023 is estimated in the United States. Volatility in labour productivity growth during the COVID period blurs signals in Canada.

What happened in 2023?

Average labour productivity growth, measured as GDP per hour worked, across OECD countries (excluding Türkiye) is estimated at 1.4% in 2023, close to pre-pandemic average of 2001-2019. The increase in labour productivity growth as compared with 2022 is estimated to be at best modest in all regions covered in the analysis, North America, Asia and Europe (Figure 2.1 and Figure 2.2). The absence of a significant improvement in productivity growth in 2023 is consistent with the expectation of a moderate real GDP growth for 2023, coupled with a decrease in hours worked relative to that of 2022 (OECD, 2023[3]).

However, labour productivity growth is estimated to vary widely in 2023 across countries (Figure 2.3). Within North America, the United States and Mexico are expected to experience a significant rebound in productivity growth while productivity growth is estimated to be negative in Canada in 2023. Within Asia, a mild increase in productivity growth is estimated in both Japan and Korea. Within Europe, Ireland stands out as the best performer, although large volatility in the data lowers the accuracy of the nowcasts for this country. Labour productivity growth is estimated to be much more modest in the rest of European countries in 2023.

Experimental estimates derived from models estimated before the COVID crisis, would point to very similar outcomes in most countries. Labour productivity growth in the OECD is estimated to 1.7% using a model that does not include the information from during the COVID crisis as opposed to 1.4% when this information is included. There is no systematic under or over estimation across countries.

There are some notable differences, though. In Canada, where models fail to capture productivity developments during the COVID crisis, estimates range from -3.9% when the information from the COVID period is included to 1.3% when it is not. Other significant differences (above 2 percentage points) are visible for Greece and Slovak Republic.

Figure 2.1. Labour productivity growth by region

Per cent

Note: Data refers to the unweighted average labour productivity growth. North America includes the United States, Canada and Mexico. Asia contains Korea, Japan, Australia and New Zealand. Europe includes 27 EU member countries. The blue shaded area depicts nowcast for 2023 together with the associated 95% confidence bands.

Source: Luu et al. (forthcoming).

Figure 2.2. Labour productivity growth in selected OECD countries

Per cent

Note: The blue shaded area depicts nowcast for 2023 together with the associated 95% confidence bands.

Source: Luu et al. (forthcoming).

Figure 2.3. Labour productivity growth in 2023

Per cent

Note: 95% confidence bands are reported.

Source: Luu et al. (forthcoming).

How to read the indicators

Experimental estimates of labour productivity growth have been derived using statistical models (Dynamic Factor Models and machine learning techniques) applied to 38 countries (Figure 2.4). A similar methodology was previously employed for nowcasting trade in value added (Mourougane, 2023[4]). A key specificity of the approach is to run models in a panel setting to mitigate small sample bias, as the target variable (labour productivity growth) is annual and only available for the years 1995-2022. Evidence from the literature suggests that such an approach increases the robustness of results in case of small samples (Woloszko, 2020[5]) (Fosten and Nandi, 2023[6]). In addition, models are estimated in a quasi-real-time – i.e. using only information available before the period that is predicted.

Figure 2.4. Main steps of the empirical strategy

The first step is to collect and process the input data that are used to estimate labour productivity growth. Predictors include national accounts data, labour market indicators, trade and business statistics, and measures of geopolitical risks and uncertainty. Non-stationary data are differenced. Indicators whose predictive accuracy is expected to be high, but which are not sufficiently timely, have been extended using the same methods and model selection criteria as for nowcasting labour productivity growth.

The second step is to select the best model to nowcast labour productivity growth. A range of models have been tested, including dynamic factor models, penalised regressions (Lasso, Ridge as well as Elastic Net), tree-based approaches such as random forest and gradient boosted trees (GBM) as well as a neural network. In addition, a “consensus” model which is the average of all machine learning models and the dynamic factor model is tested. Models are compared to a first-order autoregressive model (AR1), a standard benchmark model in the nowcasting literature. A cross-validation process is implemented to prevent overfitting – i.e. a situation when the model performs well in-sample, but fails in generalising out of sample (Hastie, 2009[7]). The best models are selected based on the root mean squared errors (RMSEs) for one-year ahead predictions.

The last step is to use the best models to nowcast productivity growth in 2023. Note that the model that performs best is selected for each country, (i.e. one best model per country), rather than the model that would perform best on average across all the countries.

Nowcasting models are found to outperform an AR benchmark

The “best” nowcasting models tend to perform better than the AR1 benchmark when performance is measured in terms of one-year ahead RMSE (Table 2.1). Overall, the benchmark model is outperformed for 37 out of the 38 countries. The GBM is selected most often as the best model (for 17 countries), indicating that it has a higher predictive accuracy than the other models. The penalised regressions also display a relatively good performance, while the neural network was chosen in only two instances. The relative performance gain compared to the benchmark AR1 ranged from 9 to 87%, as measured by (1 - relative RMSE) x 100. An additional performance metric, the Forecast Directional Accuracy (FDA) also suggests that nowcasting models predict the direction of annual developments in labour productivity growth in most cases. Indeed, for 37 countries, one-year ahead models predict the accurately the direction of change of productivity growth in at least 63% of the cases.

Table 2.1. Best model selection based on one-year ahead RMSE

	Benchmark AR1	Lasso	Ridge	ElasticNet	Random Forest	Gradient Boosted Trees	Neural network	Consensus	Total number of countries
Number of instances selected as best model	1	1	4	8	1	17	2	5	38

Note: One single model is selected as best for each country. The number of indicator-country instances are shown in the table where selected best models correspond to one of the following statistical models: benchmark AR1, lasso, ridge, elastic net (EN), random forest (RF), gradient boosting tree (GBM), neural network (NN) and consensus. Türkiye is excluded from the sample due to a large share of missing input data.

Source: Luu et al. (forthcoming).

Performance is stable at aggregate and country levels

Overall, the models demonstrate a satisfactory in-sample performance, as assessed by RMSEs, for labour productivity growth (Figure 2.5). Most models do not fully account for the COVID-19 shock, that increased the volatility of the series and led to an artificial increase in productivity growth (see (OECD, 2023[8]) and (OECD/APO, 2022[1])).

Figure 2.5. Model performance across OECD countries

Productivity growth in %

Note: Türkiye is excluded from the sample due to a large share of missing input data.

Source: Luu et al. (forthcoming).

Nowcast performance varies across countries (Figure 2.6). Despite deploying different machine learning algorithms in a panel setting, not all countries performed equally well. Larger economies demonstrated relatively good performance, in particular the United States, Japan, Germany and France, where absolute RMSEs are lower than the average of countries (1 percentage point). Other large economies, such as Italy, the United Kingdom and Korea also display relatively lower RMSEs.

Figure 2.6. In-sample errors are lower in large economies

Average RMSE, percentage points, 2015-2022

Note: Country-average RMSEs were estimated for across all years of the in-sample training period. The horizontal dashed line corresponds to the overall average absolute RMSE (1.6 percentage points). Türkiye is excluded from the sample due to a large share of missing input data.

Source: Luu et al. (forthcoming).

By contrast, a group of economies is harder to nowcast. Ireland stands out particularly, with deviations of around 5 percentage points from the average country RMSE (at 1.6 percentage points). Chile, Canada, Croatia, Estonia, Greece, and Lithuania also present above-average RMSE values.

There are several reasons behind these differences in nowcasting performance. Some countries display highly volatile patterns in labour productivity growth at the start of the COVID-19 crisis, which the models fail to capture accurately, compared to those countries with better prediction performance. In Canada, for instance, labour productivity growth experienced a notable increase in 2020 before falling markedly in the subsequent years. In contrast, in the United States, Australia and many European countries, labour productivity growth was less volatile between 2020 and 2022. Relatively disappointing nowcasting performance also stems from large volatility in productivity growth over the whole period (e.g. Ireland and Croatia). In addition, missing data persist even after pre-processing and could potentially explain poor prediction performance. For Chile, Canada, Croatia, Estonia and Ireland around 10% of observations are imputations.

References and further reading

[6] 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] Hastie, T. (2009), The Elements of Statistical Learning, Data Mining, Inference, and Prediction, Second Edition, Springer Series in Statistics.

[4] Mourougane, A. (2023), “Nowcasting trade in value added indicators”, OECD Statistics Working Papers.

[2] OECD (2023), Job Creation and Local Economic Development 2023: Bridging the Great Green Divide, OECD Publishing, Paris, https://doi.org/10.1787/21db61c1-en.

[8] OECD (2023), OECD Compendium of Productivity Indicators 2023, OECD Publishing, Paris, https://doi.org/10.1787/74623e5b-en.

[3] OECD (2023), “OECD Economic Outlook, Volume 2023 Issue 1”, OECD Publishing, Paris.

[1] OECD/APO (2022), Identifying the Main Drivers of Productivity Growth: A Literature Review, OECD Publishing, Paris, https://doi.org/10.1787/00435b80-en.

[5] Woloszko, N. (2020), “Tracking activity in real time with Google Trends”, OECD Economics Department Working Papers, OECD Publishing, Paris,.