iiinova — where your digital self keeps evolving

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

That's a good question. And thanks for your interest! Looking at places where we also observe apprentices, 52% of them also had a non-classical school. 63% either had a grammar or a non-classical school. So, we conclude that the spread was quite significant.

010

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

Link to paper (and thanks for reading): arxiv.org/pdf/2606.28063

arxiv.org

000

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

Alongside this, the paper surveys how the field already uses these tools and gives concrete guidance on model choice, digitization, and reproducibility. We think this is really useful advice. I'll do a couple of posts on these in the coming weeks.

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

We give advice on when it can be used in economic history. The good news is, in many cases. Basically, whenever ML is used to replace human annotators, debiasing is applicable.

Flowchart sorting ML applications in economic history into three task types and the recommended strategy for handling bias in each

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

All you need is small N random sample of the original data that you'd need to annotate by hand. Note that this is where good historical scholarship comes in!

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

It delivers what neither the naïve ML estimate nor the labeled subsample can offer alone: An estimate that is at once unbiased and efficient.

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

Returning to Stockholm, debiasing recovers the true gender wage gap almost exactly, with a confidence interval far tighter than using the small hand-coded subsample on its own.

The Stockholm coefficient plot with the debiased (DSL) estimate added. It coincides with the true "oracle" gap and has a confidence interval much narrower than the labeled-only estimate.*

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

The corrected estimates are consistent and retain most of the statistical power.

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

The good news is that there's already a remedy. Debiasing methods (Angelopoulos et al. 2023; Egami et al. 2023) use a small, randomly sampled set of expert "gold-standard" labels to estimate the model's error and correct for it, while still exploiting the full set of predictions.

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

So far, we've documented a pervasive problem for the use of ML tools in economic history. So, what can we do about it?

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

Second, we give a gpt-5.5 model real seventeenth-century English and compare its labels against careful human coding. We that even frontier LLMs can miss period-specific context and irony.

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

We document this with two exercises. First, we take a standard evaluation dataset for emotions and then simulate period-specific writing style. We document that the performance of a BERT classifier significantly decreases the more archaic the language of the text is.

Line chart showing a BERT classifier's accuracy and F1 declining as the same texts are rewritten into progressively older linguistic styles, relative to its performance on modern text.*

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

The effect is not only lower average accuracy but errors correlated with period, language, and region, exactly the structure that turns prediction error into bias.

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

We argue that economic history is unusually exposed to this type of problem. ML tools are typically trained on modern text, yet historical sources often use archaic language that models are less capable of handling.

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

The estimated gap is substantially attenuated relative to the estimate from true gender (11 percentage points). Yet its confidence interval gives no signal that anything has gone wrong.

Coefficient plot of the Stockholm 1920 gender wage gap. The naïve ML estimate is attenuated relative to the "oracle" estimate from true gender; the labeled-only estimate is unbiased but has a wide confidence interval

110

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

Then we can infer gender from names with a naïve Bayes classifier that reaches 98% precision Next, we regress log wages on predicted gender.

110

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

We illustrate this with an example. Consider estimating the gender wage gap in 1920 Stockholm. Let’s assume we don’t know the gender but only the names (we actually have the “true” gender from the records – so we can evaluate the results).

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

Note that model validation doesn't fix this. A model can have high accuracy and still lead to severely biased downstream coefficients.

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

Predictions carry prediction error, and that error is rarely random. In case of structural errors, the consequence is not merely a noisier estimate but a biased one.

110

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

This is conceptually false. The output from ML tools are by construction predictions.

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

Yet a great deal of this work rests on researchers using ML predictions in the same way as real data.

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

Machine learning (ML) has genuinely expanded what economic historians can do: Digitizing, linking, and extracting information from historical text at a scale that was until recently infeasible.

100

Julius Koschnick @juliuskoschnick.bsky.social · 07/07/2026

New working paper With Torben Johansen and @ChristianVedel It's on a methodological issue that we think is crucial for the discipline of economic history: Whenever we use machine learning tools in our workflow, they can introduce structural bias. So, what should we do?

ML tool performance can decline when applied to archaic text

156

Reposted by Julius Koschnick

Christian Vedel @christianvedel.bsky.social · 30/06/2026

Machine Learning is causing massive leaps everywhere. Also in Economic History. There is a conversation we need to have. What ought to be the new scientific standards? Together with two coauthors . Our argument: We should maybe care a bit less about accuracy 📝🧵

41113

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

Hence, our results show that the expansion of education in the seventeenth and eighteenth centuries contributed to the rise of the artisanal ‘upper-tail’, and therefore to the technical changes of the Industrial Revolution.

010

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

We suggest that these effects can account for up to 25% in the compositional changes of apprenticeship skills in Britain during the period

110

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

In contrast, we don’t find an effect on apprentices that did not require literacy and numeracy

100

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

We present both long-run staggered and short-run stacked diff-in-diff results that document that more schooling led to an increase in apprentices requiring literacy and numeracy

110

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

Hence, for school foundations through wills, we establish that the location of an endowment was endogenous, but that the timing of an endowment was purely determined by the timing of death of the donor – at least within a certain time frame

100

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

We then establish causally identified results in a difference-in-differences approach. For identification, we exploit the nature of the 18th century English educational system. Schools were exclusively founded through private endowments that often came in the form of wills.

100

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

These are exactly the ones that would have been needed for implementing the new technologies of the Industrial Revolution

100

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

We first document stylized facts at the cross-sectional level. There we find that non-classical schooling was strongly associated with the presence of apprentices that required reading, writing, or arithmetic upon entry

100

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

In our paper we make a first attempt at resolving this puzzle. We compile new microdata on school foundations during the eighteenth century and match them to ca. 350,000 apprentices

100

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

It seems that all of these facts fit poorly together

100

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

Beyond that, human capital, e.g. occupational skills or natural philosophy has featured prominently in accounts of the Industrial Revolution, including the accounts by recent Nobel laureate Joel Mokyr. And usually human capital and schooling go hand in hand.

100

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

The puzzle deepens when we look at the educational landscape in Britain before the Industrial Revolution. We find that the Industrial Revolution was preceded by two centuries of a “dramatic” expansion of schooling. Did this really have no effect?

210

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

However, classic explanations of the First Industrial Revolution in England have generally concluded that education was irrelevant

110

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

Let’s start by spelling out the puzzle: Education is strongly associated with economic growth over the 20th century and considered as a key factor in the transition to modern economic growth.

100

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

Link to working paper: www.lse.ac.uk/asset-librar...

lse.ac.uk

110

Julius Koschnick @juliuskoschnick.bsky.social · 05/03/2026

New paper alert 🚨 🚨 Education and Skills during the First Industrial Revolution in England Together with co-authors @sdepleijt.bsky.social and @patrickwallis.bsky.social, we set out to solve one of the most intriguing puzzles of the Industrial Revolution

J.M.W. Turner – “Going to School, for Rogers's 'Poems'” and the quantitative landscape of vocational training and education in Britain

1106

Julius Koschnick @juliuskoschnick.bsky.social · 19/12/2025

In the end, what matters is how we produce, organize, and share knowledge. That was true in 1700. It’s still true today.

110

Julius Koschnick @juliuskoschnick.bsky.social · 19/12/2025

More broadly, the paper demonstrates how historical text data and modern NLP tools can be combined to study long-run economic change. This opens up new avenues for research on long-run growth, economic history, and the knowledge economy.

110

Julius Koschnick @juliuskoschnick.bsky.social · 19/12/2025

Taken together, the results provide strong quantitative support for Mokyr’s feedback-loop hypothesis. The new quantitative evidence is crucial since Mokyr’s work provide us with one of the most influential accounts of the origin of modern economic growth.

120

Julius Koschnick @juliuskoschnick.bsky.social · 19/12/2025

Using a difference-in-differences design across technical subfields, I show that exposure to knowledge from the Lexicon technicum led to higher innovation in affected fields over the following decades.

110

Julius Koschnick @juliuskoschnick.bsky.social · 19/12/2025

For causal evidence, I exploit the publication of the Lexicon technicum (1704), the first British scientific encyclopedia, as a shock to access costs to scientific knowledge.

110

Julius Koschnick @juliuskoschnick.bsky.social · 19/12/2025

To address this, I run Placebo tests, first with other unrelated fields and then with all other fields - we see that none other fields produce coefficients of the magnitude of propositional and prescriptive knowledge

110

Julius Koschnick @juliuskoschnick.bsky.social · 19/12/2025

But what about confounders like language or writing style?

110