Strategic sector, generic funding

This analysis uses the IEP-COMPET dataset from Bocconi University, which links Horizon 2020 and Horizon Europe grant records (CORDIS) to firm-level financial and ownership data (Orbis) and patent records (Orbis Intellectual Property). The sample covers Horizon projects with at least one private company, 2015 to 2025, with company information taken at the grant year.

Deep tech taxonomy

Firms are sorted into three mutually exclusive deep tech groups, Hard Tech, Digital Tech, and Biotech, using a technology-based classification rather than sector R&D-intensity measures, which fail to capture cross-cutting technologies such as biotechnology whose applications span several industries. The taxonomy is anchored in the Strategic Technologies for Europe Platform (STEP), operationalised through the EIC deep tech domain structure, and cross-mapped to the legacy Key Enabling Technologies framework for backward compatibility. Twelve domains are defined, covering advanced materials, advanced manufacturing, micro- and nanoelectronics, nanotechnology, photonics, quantum technologies, space and aerospace, clean energy and climate, medical devices and diagnostics, AI and digital, robotics and autonomous systems, and biotechnology.

For Hard Tech and Digital Tech, each project-participant row is scored independently against all domains using four evidence signals: high-specificity keywords in the CORDIS objective text; the programme name, call identifier, or topic label; the firm’s four-digit NACE industry code from Orbis; and, where available, the International Patent Classification codes on the firm’s linked EPO patents. Confirmation requires the keyword signal plus at least one of the other three. Each row is assigned to its highest-scoring domain, with ties broken by a fixed priority order (patents, text, programme, NACE). Because assignment is exclusive, subsector counts sum to group totals without double-counting.

Biotech classifier

Biotech is classified following the definition in the EU Biotech Act (COM(2025) 1022): under Article 2(1)(1), the application of science and technology to living organisms, or parts, products, or models thereof, to produce knowledge, goods, or services. To do this we use a machine-learning classifier trained on the language of the Act together with expert-labelled Horizon project descriptions (84 for Stage A, 53 for Stage B). The classifier works in two stages: it identifies biotechnology projects (Stage A), then splits them into Health and Non-Health biotech (Stage B). Article 2(1)(2) defines health biotechnology as that directed at the promotion, protection, or restoration of human health, including veterinary public health and food safety where these serve human health protection. Non-health biotechnology is the statutory residual: all confirmed biotechnology not directed at human health, covering agriculture, environmental remediation, industrial bioprocessing, bio-based materials, and clean energy.

Both stages share the same architecture. Each project description is converted into a 768-dimensional dense vector using the all-mpnet-base-v2 sentence embedding model, which captures semantic meaning rather than surface vocabulary. A logistic regression classifier then assigns a probability score between zero and one.

The decision rule uses symmetric probability thresholds. At Stage A, projects scoring between 0.35 and 0.65 are flagged as borderline; 0.65 or above confirms biotechnology, and 0.35 or below confirms non-biotechnology. The same threshold governs the health versus non-health distinction at Stage B, with additional rule-based flags triggered by edge cases the Act explicitly acknowledges: food safety, veterinary applications, bioremediation, probiotics and microbiome research, and One Health framing. In total, 112 projects (0.4 percent of all Horizon projects in the dataset) were flagged at Stage A, and 244 confirmed biotech projects (5.3 percent of the Biotech projects) at Stage B. All 356 were submitted for structured secondary review using Claude (claude-sonnet-4-6), with the prompt framed against the statutory language of the Act and the edge-case decision rules in its Explanatory Memorandum. Borderline cases cluster around genuinely ambiguous categories: bioremediation may serve ecological restoration, drinking-water safety, or both; probiotic research spans food manufacturing and clinical microbiome therapy; veterinary research is health biotechnology when it addresses zoonotic disease risk but non-health when it concerns livestock productivity. Cases that remained indeterminate after review are excluded from the confirmed-sector analyses. Health biotechnology, as defined by this procedure, comprises 2,131 unique private firms; non-health biotechnology comprises 1,026.

Firm stage

Firms are placed in early-stage, growth, and mature tiers from three equally weighted signals, headcount, age, and revenue, with cut-offs aligned to the EU SME definition and the biotech sample distribution. Stage assignment covers 2,309 of the 3,270 biotech firms (70.7 percent); the remaining 961 could not be matched to Orbis or had no recorded firm-level EU contribution in CORDIS.

Revenue effect

The revenue analysis estimates the effect of grant receipt on firm revenue. Because firms received their first grant in different years, we line them up by the moment of receipt rather than by calendar year: each firm’s own grant year becomes year zero, and we compare revenue before and after that point across firms treated at different times. The effect is estimated in two forms: a before-and-after comparison, and a year-by-year path around receipt.

Before and after grant receipt. The binary specification is a two-way fixed effects panel regression, ln(Revenue)ᵢₜ = β·1{t≥t₀ᵢ} + αᵢ + δₜ + εᵢₜ, where t₀ is the firm’s first year of positive EU contribution and the indicator switches from zero to one in that year and stays on. Firm fixed effects absorb time-invariant company characteristics (size, country, founding year) and year fixed effects absorb common shocks (recessions, the pandemic, interest-rate moves), so β captures the within-firm revenue change pooled across all post-grant years. The outcome is the natural logarithm of annual revenue, so β is the effect on log revenue; for small values this approximates a proportional change, and the exact proportional change is exp(β) − 1, the form used when effect sizes are reported in the text. Standard errors are clustered at the firm level. This specification is also what the subgroup forest plots use: the same regression run separately on different groups of firms, for example independent SMEs or early-stage biotech, with each group’s estimated effect and confidence interval on one row so the groups can be compared side by side. The event study is not used for the subgroups because the per-year coefficients become too imprecise once the sample is split into small cells.

Over time. A single before-and-after figure conceals whether the effect builds, holds, or fades. The event-study specification replaces the one post-grant indicator with a separate indicator for each year relative to receipt, ln(Revenue)ᵢₜ = Σ_{k≠−1} β_k·1{k=t−t₀ᵢ} + αᵢ + δₜ + εᵢₜ, binned to three years before receipt and to five years after. The year immediately before receipt (k=−1) is the baseline year each other year is measured against, so each β_k gives log revenue in that event-year relative to it. This traces the full path of the effect across the post-grant years, which the single before-and-after coefficient cannot show.

Sample and controls. Both specifications use the same sample. It is restricted to firms holding a single Horizon grant, so the estimated change ties to one identifiable grant rather than an accumulating portfolio. The outcome enters as a strict natural logarithm of revenue, which is undefined for firm-years with zero or missing revenue; these observations are dropped rather than transformed or imputed. Of 92,721 firm-years in the single-project panel, 63,672 enter the regression. Almost all of the loss is missing revenue (28,037 firm-years), reflecting the delay with which small firms file accounts in Orbis; a smaller share (1,012 firm-years) is dropped because revenue is recorded as zero and its logarithm is undefined. These zero-revenue years fall disproportionately on early-stage firms yet to record sales. Ownership comparisons restrict both groups to Orbis-classified SMEs so that independent and corporate-group firms are compared at a similar scale. Ownership type follows the Orbis classification: independent firms (entity types IndepCo and SingleLoc) are standalone companies with no identified parent, while corporate-group firms (entity types GUO, ContrSubs, and Branch) belong to a wider group as ultimate owners, controlled subsidiaries, or branches.

Parallel trends. The event-study coefficients for the pre-receipt years test the assumption behind both specifications. Treated firms show significantly higher revenue two years before receipt (k=−2) than in the reference year immediately before treatment (β₋₂ = +0.069, p < 0.001), while three years before is indistinguishable from the reference year (β₋₃ = +0.017, p = 0.297). A joint Wald test rejects the hypothesis that the pre-receipt coefficients are together zero (χ²(2) = 23.4, p < 0.001), so the pre-trends fail jointly and not only at a single year. The pattern is consistent with an Ashenfelter’s dip (Ashenfelter 1978; Heckman and Smith 1999): firms tend to apply for public funding when performance has recently softened, so the year just before receipt looks unusually weak compared with earlier years. Because every event-time coefficient is measured against this reference year, the magnitudes should be read as conditional on that baseline rather than as clean deviations from an undisturbed trend. As a robustness check, we re-estimated the event study using alternative baselines: the earliest pre-grant bin (k=−3), the average of all pre-grant years, and a version dropping the suspect year. The post-grant effects survive all three, so the direction and persistence of the finding are not an artefact of the depressed baseline, even if the precise size of each coefficient is more sensitive to the choice of reference year.

Estimator caveat. Both specifications are conventional two-way fixed effects estimators, and both are exposed to a known problem under staggered treatment timing with effects that change over time. In the before-and-after specification, the estimator identifies the effect partly by comparing later-treated firms against firms treated earlier, and those earlier firms are still undergoing their own evolving grant effect, so they act as imperfect controls and can bias the pooled coefficient. The event-study specification is subject to a related contamination when cohort-specific effects differ. The estimators designed to address this underlying bias (Sun and Abraham, 2021; Callaway and Sant’Anna, 2021) work by estimating each treatment cohort separately, drawing controls from firms not yet treated, and then reading off either the corrected per-year coefficients or a corrected average effect. Applying them reliably requires each treatment cohort to be well populated across the event window, and that condition is not met here. Once the sample is narrowed to biotech and split further by ownership, stage, and health status, the individual cells contain too few firms to support cohort-by-cohort estimation. For this reason the effects are reported as plain two-way fixed effects estimates, carrying the staggered-adoption limitation described above rather than as heterogeneity-robust ones.

Disclaimers and limitations

• This analysis does not claim to be an exhaustive review of Horizon-funded deep tech firms, though it covers a substantial share of private companies across biotech, hard tech, and digital tech, 2015 to 2025.
• Because the biotech classifier is more compute-intensive than the scoring approach used for the other two groups, biotech’s share may be slightly overstated relative to hard tech and digital tech. Firms often span multiple domains, so classification is inherently subjective; groups were assigned for clarity and comparability rather than granularity.
• Revenue data thins sharply in the most recent years. Orbis compiles revenue from filed company accounts, which arrive with a delay, so revenue is available for more than 36,000 firm-years in a typical year (2015 to 2024) but only 2,412 in 2025, while EU grant data is populated through 2025.
• The biotech sub-sample becomes small once split by sub-sector and stage, so the revenue findings should be read as directional rather than definitive. The two-way fixed effects estimate is a within-firm before-and-after comparison identified in the standard way; the thin never-treated pool rules out heterogeneity-robust estimators such as Sun-Abraham (2021) and Callaway-Sant’Anna (2021), which require sufficient never- or not-yet-treated firms within each cohort.
• We restrict ourselves to firms with exactly one grant for the regression analysis to more accurately estimate the effect of a grant, but single-grant firms differ systematically from repeat recipients – smaller, less networked, less experienced with EU processes. So the revenue estimates describe single-grant firms, not Horizon recipients generally.
• Financial and ownership metrics are drawn from Orbis and may not reflect standardised accounting across all firms. SME status and ownership type follow Orbis classifications, not official EU SME designation.

This analysis draws on the IEP-COMPET dataset, developed by the Institute for European Policymaking at Bocconi University (IEP@BU). We are grateful to Giorgio Presidente (IEP@BU) for providing access to the dataset and for his input on its application in this analysis.