Thiemo Fetzer | Professor of Economics

Research Module Applied Economics with Production Network Focus

Artificial intelligence and automation are transforming labor markets, the tasks workers perform, the organization of firms, and the returns to different skills. Understanding these transformations is central to designing policies that help workers and firms adapt to technological change while safeguarding shared prosperity.

This research module connects task-based models of structural change with the study of production networks. It examines how production is organized across firms, sectors, and countries, and how modern AI methods can help measure production relationships, supply chains, factor inputs, and the allocation of tasks.

The lectures combine economic concepts, empirical applications, and hands-on work. Students learn to distinguish technological dependencies from observed transactions, construct product-level production links, harmonise descriptions with embeddings, and evaluate model-generated measurements against independent evidence.

Learning objectives

  • Connect task-based models of technological change to production-network representations of firms and sectors.
  • Distinguish technological input requirements from observed transactions and broad product associations.
  • Use structured model output and embeddings to construct and classify candidate production links.
  • Design validation procedures that identify false positives, classification errors, uncertainty, and model dependence.
  • Develop empirical research questions linking network exposure to trade, shocks, factor use, tasks, and labor-market adjustment.

Lectures and materials

Lecture 1 — Production Networks and Economic Dependencies

Production networks describe dependencies among goods, services, capital, energy, knowledge, skills, and tasks. The lecture introduces direct and indirect production links, substitutability, economic criticality, and the measurement limits of transaction, trade, and input-output data.

Lecture 2 — AI-Generated Production Networks

This lecture develops a product-level measurement pipeline. It defines the relevant production relationship, distinguishes forward and backward input-output questions, constrains model responses to a structured schema, identifies false positives and uncertain cases, and validates inferred links against independent information.

Lecture 2 slides — Construction, Validation, and Measurement Concepts

Lecture 3 — Embeddings, Data Harmonisation, and Task Reallocation

Embeddings represent words, descriptions, and documents as dense vectors. Their geometry supports semantic comparison, nomenclature matching, retrieval, and classification. The lecture develops the statistical intuition behind learned vector representations and applies it to data harmonisation, document retrieval, and task-based models of technological change.

  • N-grams, context windows, and the prediction problem underlying embeddings.
  • Softmax probabilities, dot products, cosine similarity, and gradient-based learning.
  • Vector arithmetic, dimensionality reduction, and representational bias.
  • Harmonising product descriptions, survey questions, and job descriptions.
  • Jobs as bundles of tasks and the reallocation of tasks following technological or economic shocks.
Lecture 3 slides — From N-Grams to Embeddings

Supplementary lecture notes develop the connections among embeddings, data harmonisation, production networks, and task reallocation.

Lecture 4 — Network Structure and Empirical Applications

The lecture examines the structure and empirical uses of granular production networks. Topics include centrality, substitutability, topological ordering, shock propagation, global adjustment, trade exposure, and evidence from the Qatar blockade.

Lecture 4 slides — Measurement and Applications from Global Trade

Lecture 5 — Classification and Retrieval Pipelines

The lecture covers local model workflows, structured classification, information retrieval, dataset construction, and strategies for scaling a reviewed labelling exercise. It connects retrieval and classification methods to the construction of research-ready economic data.

Lecture 5 slides — Classification and Retrieval

Technical foundations

Two supporting slide sets provide the machine-learning and language-modelling foundations used throughout the module.

Machine Learning and AI in Economics
Generative Language Modelling with N-Grams

Hands-on notebooks

The practical component treats a production network as a measured object rather than a collection of model-generated statements. The notebooks use small teaching examples so that each modelling and validation choice can be inspected.

  1. Classifying candidate production links: distinguish technological inputs from downstream activity and broad economic association.
  2. Extracting and validating production links: produce schema-constrained candidate links and subject them to a separate review query.
  3. Mapping product descriptions with embeddings: align free-text descriptions with a reference nomenclature using cosine similarity and explicit review thresholds.
Download the Jupyter notebooks

The archive contains the three notebooks listed above as separate .ipynb files.

The notebooks run in Jupyter through the Python (Production Networks) kernel and use local Ollama models. Model outputs are intermediate measurements; they are not observed transactions or causal estimates.

Software setup and local AI

The practical sessions can be run with an open-source editor and a local inference stack. The following videos provide a concise sequence from editor installation to a private coding workflow.

  1. Install VSCodium: Awesome Linux Tools — VSCodium introduces the open-source VS Code distribution and its installation on Linux. Students using macOS or Windows should follow the official VSCodium installation instructions.
  2. Run a model locally: Learn Ollama in 15 Minutes covers installation, model management, the local HTTP API, and Python access.
  3. Connect OpenCode to a local model: OpenCode + Ollama — Local AI Coding Setup demonstrates a complete local coding-agent workflow.
  4. Understand portable self-deployment: Llamafile: Bringing AI to the Masses with Fast CPU Inference explains Mozilla’s single-file approach to local model distribution and inference.

Model choice should be matched to the available memory and hardware. Small quantized models are sufficient for the notebook exercises; larger models generally improve classification quality but require more memory and longer inference times.

Research directions

  • Technology and tasks: connect changes in a sector’s input mix to occupational tasks, skills, employment, and firm organization.
  • Supply-chain exposure: compare technological dependencies with observed trade or transaction links around a disruption.
  • Substitution: identify groups of inputs that perform similar production tasks and test whether substitution occurs after a price or supply shock.
  • Measurement validation: evaluate how prompts, models, classifications, and human review alter the resulting network.

Selected literature