Recent research points to a measurement-first, hybrid forecasting strategy rather than a simple substitution of traditional econometrics with machine learning. Digitalization changes what is produced, how it is sold, and where transactions occur, while official statistics often measure these changes imperfectly. The unresolved problems include quality-adjusted prices, platform and “free” services financed by advertising, cross-border digital trade, intangible assets, sector classification, and the boundary between market output and consumer welfare. For some digital activities, price, quantity, quality, and even revenue are difficult to observe; this can create a wedge between measured GDP and changes in economic welfare. (Coyle, 2025) Digital-economy statistics also face ambiguous definitions, inconsistent statistical standards, and limited access to basic data. (Yan, 2025)

For forecasting, the first practical implication is to use real-time data vintages, not only the latest revised dataset. Macroeconomic observations have different degrees of maturity, and treating all observations as final can reduce the accuracy of both point forecasts and uncertainty estimates. Research recommends either modelling the revision process, using revision-robust reduced-form methods, or restricting some applications to relatively unrevised indicators, depending on whether the target is the first release or the eventual revised value. (Clements, 2019) Models that explicitly incorporate real-time vintages or a model of revisions improve density forecasts for US and UK output growth and inflation. (Clements, 2022) This is not a minor technical detail: a study of Polish macroeconomic data found that revisions changed the forecasting process by more than 10% for many variables, with extreme cases exceeding 100%. (Ziembińska, 2021)

The second implication is to combine high-frequency and unconventional data with an economic model. Internet searches, social-media sentiment, financial transactions, administrative records, firm accounts, and other digital traces can reduce the lag of official indicators and improve nowcasting. But evidence is mixed: a critical review concludes that big-data forecasts have had varying success, and that the strongest applications embed these data in a coherent economic framework rather than using a purely black-box model. (Richardson, 2019) A recent real-time nowcasting study reports gains from combining high-frequency and non-traditional data with econometric and machine-learning models. (Malysheva, 2023) At a longer horizon, firm-level accounting data can preserve heterogeneity, nonlinearities, interactions, and structural breaks lost through aggregation; a 2026 study reports more accurate GDP forecasts from machine-learning models applied to large panels of firm earnings. (Cui, 2026)

The strongest modelling toolkit therefore appears to be ensemble-based and adaptive: dynamic-factor or targeted-predictor models for dimensionality reduction, mixed-frequency methods for asynchronous releases, machine-learning models for nonlinearities, and forecast combinations for model uncertainty. Evidence from a 143-series US dataset suggests that automatically selecting factors is not necessarily useful, whereas emphasizing low-frequency factor information improved some forecasts and produced the best GDP results when combined with a small preselected predictor set. (Chudý, 2019) Other work finds that forecast combinations and machine-learning ensembles can reduce errors relative to traditional models; one application to Turkey reported at least a 30% reduction in forecast errors, although that result should not be generalized without country-specific validation. (Bolhuis, 2020) Forecast evaluation should use rolling or expanding pseudo-real-time tests, vintage-specific data, regime and structural-break checks, and density as well as point-forecast metrics. Inference itself must be adjusted for data revisions; recent work develops a block-bootstrap procedure for predictive-ability tests with real-time revised data. (Gonçalves, 2024)

For your project, the central research design is to define the forecasting target precisely—first release, revised estimate, or latent economic activity—then build a measurement layer before the prediction layer. The measurement layer should document definitions, quality adjustment, missing digital activity, geographic and sector coverage, revisions, and provenance for every predictor. The prediction layer should compare a transparent benchmark, a revision-aware econometric model, factor or mixed-frequency models, and several machine-learning models under identical rolling real-time evaluation. Report performance separately for stable periods and shocks, and include prediction intervals. Digital variables should be retained only when they add out-of-sample information after accounting for release timing, revisions, data drift, and changing platform coverage. This approach treats the digital economy as both a source of predictive information and a source of measurement error—the distinction that recent research identifies as decisive for credible macroeconomic forecasting.