Lead source: research scan, 2026-10-16 — an article in npj Complexity and Michele Coscia’s blog
Sometimes science begins not in a laboratory, but in the middle of someone else’s presentation. Michele Coscia — a network scientist who usually measures economic complexity through 20th-century trade flows — was sitting through a talk by archaeologist Petra Heřmánková when he suddenly realized that his two worlds were colliding like protons in the LHC: she had half a million Latin inscriptions with coordinates in her hands; he had the Economic Complexity Index toolkit. Out of that collision came a paper in which the “résumé of an empire” was assembled from epitaphs.
What hooked me was the second layer: having calculated Rome’s economy from stone, the authors found that 1500 years later it was still “there” — the correlation between Roman and modern economic-complexity rankings is 0.439 at p < 0.05. Either that is a deep regularity, or an elegant hallucination of the method — and I wanted to find out which one was right.
The LIST dataset (Latin Inscriptions in Space and Time) contains 525,870 Latin inscriptions from the 1st–4th centuries, merged from two major epigraphic databases, EDH and EDCS (77,091 records overlap between them — deduplication trims the list). 511,973 inscriptions have geocoordinates, 206,570 have numerical dates. Then comes what I call the “archaeological funnel”: the entire corpus is put through computational reading, yielding 8,475 inscriptions that mention an occupation. That is 1.6%. For every hundred stones that survive to this day, a little more than one tells us what a person did in life.
Why so few? Because a tombstone is not a census but a shop window. A Roman epitaph is a résumé drafted by the heirs: “To the spirits of the dead. Here lies Gaius, a fuller, a worthy man.” In such a display, an occupation gets mentioned if people were proud of it, and omitted if you were a slave porter or a peasant with neither money for a stone nor a word in the will. Of these 8,475 inscriptions, 37% (3,116) are funerary, 11.4% (952) honorary, with cursus honorum, the magistrates’ career ladder. Occupation mentions can be georeferenced in 10,353 cases, and all of them are distributed across 512 distinct occupations — from agricola (farmer) and faber tignarius (carpenter) to curator (administrator), along with dozens of craft specializations whose existence we know only because someone paid to have them carved in stone.
At the input is a “province × occupation” matrix, most of whose cells are empty. The team does exactly what economists of the complexity school do with global exports: RCA (revealed comparative advantage), binarization, the second eigenvector. The logic of ECI mirrors the modern one: a province is “complex” if it has occupations that are rarer than average across the empire; a province with only what everyone else has is simple. Classic Hidalgo–Hausmann ECI, made to work on stone.
But first come four corrections, each one the crutch of an honest person. Status: the elite (1–5% of the population) are overrepresented, so the counts of “lower” occupations are inflated to demographic proportions. Gender: there are twice as many men as women in the inscriptions, so female counts are doubled. The primary sector: agriculture employed ~85% of the population, but no one bought peasants monuments — so the agricola count is inflated. Research bias: inscriptions have been digitized unevenly; some places were excavated more intensively and their campaigns better funded. Each crutch is not a scalar but a probability distribution: thousands of plausible matrices are generated, ECI is calculated for each one, and the output is a distribution rather than a point. Methodologically, it is beautiful: not “here is the exact value,” but “here are the bounds within which we can trust it.”
The foundation holding it all up is nestedness. After correction, the “provinces × occupations” matrix is 5.7 times denser to the left of the isocline, p<0.01 against a thousand randomizations: the structure is not random. Complexity here too is arranged like a set of Russian nesting dolls, as it is in the modern world: complex provinces have everything simple ones have, plus the rare stuff. Rome with Latium is the outlier at the top, followed by commercial gateways: Venetia, Carthage, Baetica. The Danubian provinces, meanwhile, clump into a single cluster — a military belt, not a market: the same set of occupations around the legions.
Now for the main thing. The authors calculate “ancient ECI” within the borders of modern countries located on the territory of the Western Roman Empire — there are 23 of them — and correlate it with today’s ECI from UN Comtrade (SITC Rev 2, 1962–2022, the same algorithm built on the same principles). Spearman rank correlation: ρ = 0.439, p < 0.05 — across 23 observations and a span of a millennium and a half.
The top 10 then and now (Table 1 from the paper):
| No. | Ancient ECI (from inscriptions) | Modern ECI (from trade) |
|---|---|---|
| 1 | Italy | Germany |
| 2 | France | Switzerland |
| 3 | Germany | Austria |
| 4 | Spain | United Kingdom |
| 5 | Belgium | Slovenia |
| 6 | Algeria | Hungary |
| 7 | Slovenia | Italy |
| 8 | Hungary | France |
| 9 | Serbia | Belgium |
| 10 | Switzerland | Spain |
The key thing in the table is that 8 of the 10 countries appear in both top lists. Algeria and Serbia have “dropped out”; Britain and Austria have “entered.” Notice: not a single rank matches exactly — it is the membership of the top tier that matches, not the places within it. Spearman was not chosen by accident here: Italy is such a sharp outlier in ancient ECI that Pearson would show whatever it liked.
And let us hear how the authors themselves interpret the result:
We don't yet have an explanation for this finding. We cannot establish whether there is a direct causal relationship, or whether some common third factor could explain both. The observation is astounding nonetheless. It is at least consistent with the idea that economic complexity is sticky: our work cannot prove this idea, but we have definitely failed to disprove it.
“We do not yet have an explanation. We cannot establish either direct causality or some common third factor. But the observation itself is so strange that it is at least consistent with the idea that complexity is sticky — our work does not prove it, but it definitely failed to disprove it.”
“We failed to disprove it” is a rare formulation of such integrity in a world where every other preprint promises a revolution. And that honesty makes you want to ask the next question: why might complexity be “sticky” at all?
There are three sharpest questions here.
1. Wealth, seen twice. A wealthy region in Roman times left behind more inscriptions — more stones, more literate heirs. A wealthy region today has been excavated more thoroughly and digitized more completely. If wealth changes slowly over time, then research bias is measuring the very variable it is trying to control for: modern complexity correlates with the visibility of stone not because Rome passed along a baton, but because the region was rich then and is rich now. This is the sharpest angle of attack, and it was named outright in the Hacker News thread: the authors correct for status, gender, and peasants, but Roman stone itself determines where digging happens in the 21st century. The paper has sensitivity estimates, but the outline of this correction remains an open wound.
2. The Greek blind spot. The eastern half of the empire wrote in Greek: the GIST dataset contains 217,863 inscriptions. But NLP extraction of occupations from Greek texts currently produces more than 50% false positives — the detector “finds” an occupation where none exists (this was checked by a human epigrapher’s reading). Which is why “Western Roman Empire” in the title is not an aesthetic choice but a technical necessity; the East appears only in a robustness test, where Greek counts are approximated with Latin coefficients — rather like measuring China’s economy with French dictionaries.
3. Epitaph ≠ economy. Even with corrections, what we see is stone that a person of status could afford. 512 occupations across 1.6% of artifacts — that is a shop window, not a map of the labor market. The fact that the correlation remains significant after all the adjustments does not mean the signal is clean — only that the biases are measurable. The question of “the same wealth counted twice” remains open.
Even if we accept the result, three mechanisms that are not always distinguishable may lie behind it.
The authors honestly refuse to choose among them. But the three mechanisms make different predictions, and that opens the way for a future test. Each hypothesis explains the table in its own way. Endogenous stickiness is happy with 8 out of 10 matches: the baton kept passing wherever continuity did not break. Geography explains both the matches (the Rhine, Rhône, Po — riverine axes of trade in both eras) and the asymmetry: Britain, having lost Roman administration for a couple of centuries, returned to the top — you cannot hide an island with ports; the Danube, Rome’s military belt, leads nowhere today. Algeria, however — ancient No. 6, absent from the modern top tier — is awkward for stickiness: Africa was among the leaders then; the Maghreb is not there now.
What happened to Roman Africa — the empire’s granary and one of its centers of complexity? After the Vandals, Byzantines, and Arab conquest, urban continuity was broken in places. This is my reading, not the authors’ conclusion, but it is a testable hypothesis: stickiness is not a sentence; it works where the urban grid, infrastructure, and transmission of skills survived the ruptures. A couple of breaks like that are enough for the 0.439 line to sag to zero.
For me, the main story here is not “Rome was cool,” but three things.
First: the epitaph as Big Data. LIST and its Greek sister GIST are what happens when the classical humanities digitize themselves: 525,870 records, an ML classifier for inscription type across 22 categories, coordinates, dates, all stitched together from more than two centuries of manual epigraphic labor. A discipline in which people spent their entire lives deciphering one inscription at a time has suddenly acquired a weapon of mass investigation.
Second: scientific honesty as a feature. The authors got an intriguing result and, instead of saying “we discovered the stickiness of complexity,” said “we failed to disprove it.” In an age when every preprint sells a revolution, this is the kind of hygiene you want to see in every study you read.
Third: if stickiness is real, developing countries have a problem. The uncomfortable conclusion of the paper is that relative rankings of economic complexity are conservative. The first commandment of economic policy then becomes not “catch up within a generation,” but “do not lose what you have.” The authors leave this conclusion in cautious form — and that caution is no accident: the correlation is weak (0.439), the sample is small (23), the mechanism unknown. But if this is even partly true, the gap between rich and poor countries is not an anomaly but an entrenched structure, one that cannot be shaken without a deliberate global effort.
Coscia’s blog ends with a joke: “As a third and final action point, I am of the opinion that Carthage must be destroyed.” “Furthermore, I consider that Carthage must be destroyed” — Cato, alive 22 centuries later. The HN thread, naturally, decoded the joke in the very first comments. The difference is that the real Cato ended every speech with it, while Coscia just plays too much Crusader Kings III. And there, the entire scientific method is more visible than in any textbook: first the eureka moment at someone else’s presentation, then a thousand randomizations, then “we failed to disprove it” — and at the very end, inevitably, “Carthage must be destroyed.”
Sources: Mazzamurro, Heřmánková, Coscia, Brughmans — “The Economic Complexity of the Roman Empire” (arXiv:2508.19892; npj Complexity, 2026); michelecoscia.com/?p=2613 blog; LIST dataset (Zenodo, v1.1); Hacker News discussion (item 50084790).
📝 2026-10-16 · Curiosity · Rome, economic complexity, epigraphy