Backtesting the screen: two decades, out of sample

1 · How the test works

Each vintage screen is a deliberately simple reconstruction: an equal-weight blend of three EU-wide percentile ranks knowable at the time — prosperity (real GDP per head, constant 2020 prices), market scale (total real GDP) and momentum (real GDP growth over the prior five years). It is not today's 13-pillar screener — most pillars have no deep history — but these three signals sit at its economic core, and equal weighting is a plain screen, not an optimized model.

All series are JRC ARDECO regional accounts (real GDP at constant 2020 prices, with GDP per head derived over population; workplace employment; census-break-harmonized annual-average population), backfilled to 1995 across 1,151+ NUTS-3 regions. Universe: 1,151 of 1,211 regions (60 recoded regions and microstates lack full history; see caveats).

Everything below is rank correlation between a ranking knowable at the vintage date and subsequent outcomes — it shows what the ranking anticipated, not what it caused.

2 · What the 2015 screen anticipated: people, then jobs

Sort regions into ten deciles by their 2015 screen score and the realized population-growth ladder runs monotone from bottom to top: the screened-out decile lost population at -0.74%/yr while the top decile gained +0.68%/yr — a 13-percentage-point divergence in demand base over nine years. Rank correlation: population ρ = +0.63; workplace employment ρ = +0.19 (positive but weaker — top decile +1.12%/yr vs +0.46%/yr, with jobs lumpier than residents).

3 · The replication: a 2005 screen, judged through 2014

A backtest over one window is a single historical draw. So we ran the identical experiment a decade earlier: the same three signals as they stood in 2005 — before the global financial crisis — judged on 2005→2014 outcomes. The structure repeats almost exactly: population ρ = +0.64 (vs +0.63 in 2015→2024), a monotone ladder from -0.59%/yr to +0.75%/yr, within-country ρ = +0.50, positive in all 21 of 21 countries (median +0.54), ex-CEE ρ = +0.61, employment ρ = +0.19. Two different decades, two different crises — the same answer.

4 · What neither screen anticipated: %-growth of GDP

In both windows the screen has no meaningful correlation with realized real-GDP growth rates: ρ = -0.03 (2015→2024) and -0.02 (2005→2014). One signal is negative in both windows: real GDP per head at the vintage date against realized real per-capita growth is ρ = -0.52 (2015→2024) and -0.17 (2005→2014) — beta-convergence, far more forceful in the recent decade when catch-up in Central and Eastern Europe ran at full speed, and muted in the crisis decade that hit Southern Europe hardest. Poorer EU regions grew faster in percentage terms overall.

The honest generalization: a screen built on prosperity level and recent real momentum alone could not rank future %-GDP growth in either decade — %-growth league tables are structurally dominated by low-base catch-up effects. Treat any tool that promises to find “the fastest-growing regions” by ranking on today's prosperity with suspicion; this test, run twice, is the evidence. (Signals of a different kind — innovation intensity, demographic structure, industry mix — are a separate question this test does not answer.)

For siting and real-asset demand the null matters less than it sounds: %-GDP growth is not the demand variable. People and jobs are — and those followed the screen, twice.

5 · Which signal did the work

Prosperity level pulled hardest. Recent momentum was the weakest of the three — and its correlation with future %-real-GDP growth was essentially zero. Recent growth alone does not persist; where prosperity already concentrates, people keep arriving.

6 · Against naive baselines — and how slowly the signal decays

After an adversarial external review (four independent models critiquing this page), we added the two tests a skeptical quant asks first: does the screen beat a naive baseline? and how fast does the signal decay?

Baseline honesty first: if your only question is “where will people be”, demographic momentum itself is the strongest single predictor — population is enormously persistent (prior-5-year population growth vs the next nine years: ρ = +0.80). The economic screen never sees a demographic input, still reaches ρ = +0.63, and retains a positive partial correlation of +0.20 after controlling for demographic momentum entirely. Its role is not to out-forecast demography — it ranks economic mass and prosperity, which carry independent signal a demand-side analyst does not get from population trends alone.

And unlike a timing signal, it barely decays across a fund's deployment window — near-identical rank correlation at one, three, five and nine years (the 2005 window repeats the pattern: +0.54 / +0.55 / +0.56 / +0.64). This is a structural screen, not a trade.

7 · Winners and losers, named

8 · Caveats — read these

• Two nine-year windows are two draws, not a panel of independent trials — and regions are spatially and nationally correlated, so naive significance thresholds overstate certainty. The within-country decompositions and the cross-decade replication are the load-bearing evidence, not an n=1,151 t-statistic.
• ARDECO series are model-harmonized official accounts: census breaks are smoothed and recent years are partly provisional (the final outcome year includes nowcast components; re-running window B with a 2023 endpoint changes no conclusion). Population here is an annual average, so levels differ slightly from Eurostat 1-January counts.
• Employment is workplace-based (jobs located in the region, not resident workers) — commuter hubs score differently than residence-based measures. The residence-based labour-force outcome gives a stronger correlation (ρ = +0.56 in window B); we show the weaker workplace number.
• 60 of 1,211 regions are excluded, and not at random: NUTS-recoded regions and microstates lack full back-history on current boundaries.
• Percentiles and boundaries are today’s — a reconstruction, not a literal vintage snapshot. Regions never exit the panel, so there is no survivorship bias, but boundary harmonization can shift country composition slightly.
• This validates ranking signals against demographic and employment outcomes — not investment returns. No pan-EU rent, price or yield outcome exists in public data.
• Equal weighting is a plain screen, not an optimized model — in window B the population correlation holds between roughly +0.57 and +0.65 across 50/25/25 weight permutations, and strengthens in the pre-COVID subperiod.