Can similar economies improve ICT-growth nowcasts? An OECD model cut forecast error by 17%
Across 28 countries, a machine-learning model that pooled structurally similar ICT economies reduced out-of-sample error by an average 17% versus pooling every country together. The result comes from recursive backtests; the reported 2025 growth rates are estimates, not official outcomes.
The short version
- What happened
- The OECD evaluated annual ICT-value-added predictions for 28 countries using a recursive, one-year-ahead design covering 2015–2024 or the latest available year.
- Who it affects
- Public agencies and businesses may gain an earlier signal of ICT-sector direction, but the model does not identify individual job, wage or company outcomes.
- What remains unproven
- The evaluation covers 28 countries and short annual series, generally fewer than 20 observations per country.
Keep in mind
The evaluation covers 28 countries and short annual series, generally fewer than 20 observations per country.
Google Trends is sampled, platform-dependent and sensitive to changes in language, attention and search behaviour.
The 17% improvement is an average historical RMSE reduction; gains vary by country and weakened under some pandemic-era comparisons.
The reported 2025 growth rates are nowcasts with 90% intervals, not observed national-accounts outcomes.
SHAP feature contributions describe model behaviour and do not establish that AI, searches or any other topic caused economic growth.
Source analysis · 1 linked source. Primary analysed source date: 5 November 2026.
Read the full report · 10 min The evidence, methods, implications and limitations
The answer: structurally similar peers beat one global training pool in the backtest
An OECD study released on 5 November reports that a machine-learning model predicted national ICT-sector growth more accurately when it learned from economies occupying similar positions in global technology supply chains. Across 28 countries, the selective approach reduced root mean squared error, or RMSE, by an average 17% compared with a model trained on one global pool. Canada showed the largest reported improvement at 43%, followed by the Netherlands at 40% and Poland at 36%. Those percentages describe reductions in historical prediction error, not faster economic growth or a direct benefit delivered to firms.
The problem is practical. Official measures of ICT value added arrive slowly because national accounts require validation and revision. At the same time, annual country series are short: the report says there are usually fewer than 20 observations for each country. A flexible model trained separately on such a small history may fit noise. Combining countries supplies more training examples, but mixing economies whose hardware, software, trade and demand patterns differ can introduce bias. The OECD paper treats that choice of peers as part of the model rather than assuming that every country should be pooled together.
The result is consequential for finance ministries, central banks, industry analysts and businesses that need an early view of a volatile sector. It is not an urgent public event and it does not establish that official 2025 growth has occurred. The evaluation is a recursive historical backtest, and its newest country figures are model estimates with 90% intervals. The report therefore belongs in Finance & Business and Latest rather than Breaking.[1]
How the model linked search behaviour to ICT value added
The target was annual growth in the volume of ICT value added from the OECD Structural Analysis database. Most official series ran through 2023 when the model was prepared. The input features were annualised Google Trends topics, including technology subjects such as cloud computing, artificial intelligence, generative AI, DevOps and gaming computers, alongside wider signals such as inflation, energy and stock-market indices. The report describes repeated sampling, variance filtering, detrending and seasonality correction to reduce instability in the search data. Country indicators were added to account for persistent national differences.
A feed-forward multilayer perceptron then estimated the relationship between those topic signals and ICT growth. Continuous inputs were standardised. The researchers selected the network architecture, learning rate and regularisation through a grid search with nested validation. Training used early stopping on a 10% validation subset once the sample was large enough, and final predictions averaged multiple networks with different random seeds. This ensemble was intended to reduce the luck of a particular initialisation and supplied the variation used for symmetric prediction intervals.
The evaluation preserved time order. For each year, the model used only observations from earlier years and produced a one-step-ahead prediction. Errors accumulated over the evaluation window, which ran from 2015 to 2024 or the latest year available. RMSE was the primary metric. That design is more informative than a random train-test split for a time series because it imitates what an analyst could have known at the time. It is still a pseudo-real-time exercise: later database vintages and the exact historical availability of every input may differ from a live forecasting environment.[1]
The distinction that matters
How the selective nowcast was evaluated
The model combined timely search signals with lagged official data, while every historical forecast used only information from earlier years.
28 countries
Rank peers by ICT value-chain similarity
Country-specific pools from 2 to 23 peers
2015–2024
Train on prior years and predict the next
Recursive out-of-sample RMSE
Selective pool
Choose the prior best-performing pool size
17% lower error on average than global pooling
2025 estimate
Apply the model to current signals
4.9% OECD point nowcast, not an observed result
What 'similar' meant—and how each country's pool was chosen
The study did not group countries simply by geography or income. It used the OECD Trade in Value Added database to describe two sides of each economy's ICT production network. Forward linkages captured where domestically produced ICT value added was ultimately consumed. Backward linkages captured where the foreign value added embedded in a country's ICT exports originated. Similar forward and backward profiles were treated as evidence that two countries might encounter related demand and supply shocks.
For every forecast year, the proximity calculation used only trade data available before that year. It combined cosine similarity in the distribution of value-added flows with overlap among the ten largest partners. The report weighted the continuous cosine measure at 70% within each linkage profile, then weighted forward linkages at 80% when combining forward and backward proximity. Sensitivity tests examined other parameter choices. Those weights are modelling decisions grounded in earlier trade-network literature; they are not natural constants, and another sector or period may require different settings.
For a target country, the algorithm ranked all 28 countries by structural proximity and tested pool sizes from one—effectively no international pooling—to 28, the global pool. In each evaluation year it selected the size with the lowest cumulative out-of-sample RMSE from prior years, subject to a minimum history. The preferred pool varied sharply: the Netherlands and Slovenia used two-country pools, Spain seven, Italy 16, France 19 and Germany 20. This variation is central to the paper's claim. More observations helped only while the added economies remained relevant enough to offset the bias introduced by their differences.[1]
What the comparisons show—and where the improvement breaks down
The selective model generally beat both a first-order autoregressive benchmark and the 28-country global pool. In the full evaluation, its country RMSE values ranged from 1.5 for Canada to 9.2 for Iceland and Sweden, reflecting very different underlying volatility. For the United Kingdom, the report gives RMSE of 3.1 and error ratios of 0.9 against the autoregressive benchmark and 0.8 against global pooling. A ratio below one indicates that the selective model made the smaller error; it does not mean the forecast was correct within a fixed percentage on every year.
The COVID-19 period was a harder test. From 2020 to 2022, errors rose and the selective model was not universally best. It remained ahead of both comparators in many countries, but Denmark's error ratio was 1.1 against the autoregressive benchmark, Estonia's was 1.1 against global pooling, and the United States reached 1.7 against the autoregressive benchmark. The authors conclude that structural similarity is most useful in stable conditions, while a broad global pool can become more competitive when a common shock hits many economies at once.
A ridge-regression robustness comparison in the annex generally supported the neural-network result, and country-level charts suggest that the average improvement was not created by one outlier. Even so, the same historical period informed both model development and the cumulative choice of pool size. The paper uses careful recursion, but it is not an independent prospective test by a separate team. Volatile countries also gained less from selective pooling, showing that value-chain similarity cannot remove large domestic shocks or measurement error.[1]
The 2025 numbers are useful signals, not observed economic results
Applying the selected pools, the model estimates average OECD ICT-value-added growth of 4.9% for 2025. The highest point estimates were 9.1% for Lithuania, 9.0% for Hungary and 8.0% for Germany. At the lower end, it estimated 0.6% for the Slovak Republic, 1.1% for Italy, 1.7% for Canada and 1.9% for Switzerland. The distribution was narrower than the model's 2024 estimates, which the report interprets as possible convergence after several unusually strong country results.
These figures should not be reported as official growth. The underlying national-accounts target is lagged, the search indicators can shift as platforms and user behaviour change, and the model's 90% bands are wide for several countries. Ensemble variation is also not the same as a full accounting of data revisions, model choice and structural breaks. A government or investor could use the estimates as one early signal, but decisions should be checked against labour, business, trade, tax and production data rather than relying on search behaviour alone.
Nor does the paper measure how much AI caused ICT growth. AI-related search topics influenced some predictions, but SHAP values allocate a model's output; they do not identify causal effects. A rise in searches may reflect investment, curiosity, news coverage or concern. The model also includes macroeconomic topics because ICT activity responds to wider demand, energy and financial conditions. Treating the result as proof that generative AI produced national growth would go beyond the design.[1]
What this means for people, and what evidence should come next
For workers and businesses, a better early estimate could help identify where demand for digital services, equipment and skills is accelerating or slowing. Public agencies could use it to time surveys, training support or infrastructure analysis. But nowcasts do not tell an individual whether a job will be created, whether wages will rise or which firm will succeed. A model optimised for aggregate value added can be accurate while missing distributional effects across occupations, regions and company sizes.
The most useful next test would be prospective: publish frozen forecasts before official releases, then score them against subsequent national-account vintages. Independent teams should reproduce the pipeline using other high-frequency sources and test whether the selected peer groups survive revisions to STAN and TiVA. Quarterly targets would provide more observations, as the authors suggest, but would also introduce seasonal and revision problems. Reports should continue to publish country-level intervals and failures rather than only an OECD average.
The defensible conclusion is narrower than the most tempting headline. Within a 28-country annual backtest, selecting peers from similar ICT value chains reduced average error relative to pooling all countries, with large differences by country and period. That is a useful advance in economic measurement. It does not turn Google searches into official statistics, make the 2025 point estimates observed facts or show that AI caused the growth being predicted.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring machine learning: compare the finding and its evidence limits.
Can machine learning show income inequality before surveys arrive?
An OECD working paper combines machine learning with tax-benefit simulation to estimate EU income inequality sooner. It improves on ordinary regression, but simple no-change and non-ML benchmarks remain competitive; estimates are preliminary and cannot determine an individual household's benefit entitlement.
Read the next reportExplore the full source and evidence record
What this means for people
- Public agencies and businesses may gain an earlier signal of ICT-sector direction, but the model does not identify individual job, wage or company outcomes.
- Workers should not read a national growth nowcast as a prediction for their occupation or region; aggregate value added can hide uneven effects.
- Decision-makers should combine the estimates with official releases and other timely indicators, especially where the country interval is wide or shocks are unusual.
Global context
The sample spans Europe, North America, Costa Rica and Mexico, but it is limited to countries with sufficiently comparable OECD data. The structural-pooling idea may travel beyond that set, yet value-chain data quality, search-platform coverage, language and informal economic activity differ substantially. Economies outside the sample need their own validation rather than an assumption that the OECD error reduction transfers unchanged.
What the evidence does not yet show
- The evaluation covers 28 countries and short annual series, generally fewer than 20 observations per country.
- Google Trends is sampled, platform-dependent and sensitive to changes in language, attention and search behaviour.
- The 17% improvement is an average historical RMSE reduction; gains vary by country and weakened under some pandemic-era comparisons.
- The reported 2025 growth rates are nowcasts with 90% intervals, not observed national-accounts outcomes.
- SHAP feature contributions describe model behaviour and do not establish that AI, searches or any other topic caused economic growth.
What to watch next
- Frozen prospective forecasts scored against later official data vintages.
- Independent replications with alternative high-frequency indicators and other country groups.
- Whether quarterly models improve timeliness without becoming unstable or revision-sensitive.
- Country-level failures during common global shocks and large domestic disruptions.
- Clear separation between AI-related search signals, ICT-sector growth and causal claims about productivity or employment.
Living evidence record
Impact record IAI-1EJ4R6A
Evidence stage
Observed
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
5 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what OECD published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 5 November 2026
Research question, topics and regions
Machine-learning nowcasts of national ICT-sector growth
Topics: Economic nowcasting · ICT sector · Machine learning · Google Trends · Global value chains
Regions: OECD · Europe · North America · Latin America · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.