Geoeconomics and LLMs: how to apply them?

Starting from the traditional equation GDP = C + G + I + (X − M), I incorporate qualitative exogenous variables such as the Global Peace Index from the Institute for Economics & Peace to understand the model's interaction with geopolitical factors. My GitHub repo for this first stage: https://github.com/Murgueytio/Models_Econometric

Later, additional qualitative variables will be added from sources like The GDELT Project and the International Institute for Strategic Studies, applying LLMs and Relevance-Based Prediction (RBP). The goal is to evolve econometric modeling and make it more realistic by understanding global dynamics. I rely on techniques learned in the DeepLearning.AI Short Courses led by Andrew Ng.

Four references that guide the work

1. Hayden Van Der Post and Alice Schwartz, in Data Driven Decisions: Advanced Econometric Techniques With Python, put it clearly: "Unlike traditional econometric methods, which often assume linearity and homogeneity in the relationships among variables, machine learning allows for more flexible modeling and better captures the complex, non-linear interactions present in economic data."

2. From my position as a Data Scientist and economist, analyses like Data Economics and Geopolitics of Data interest me: econometrics lets us model the behavior of actors in the data market, and with ML and DL we can quantify the effects of geopolitical decisions on the economy, identify opportunities and mitigate risks. That's why I follow IE University's Center for the Governance of Change closely.

3. There will be biases in the projections, as the World Bank warns in Using Large Language Models for Qualitative Analysis can Introduce Serious Bias (Ashwin, Chhabra and Rao): "…it may be preferable to train a custom model on annotated data rather than use an LLM to annotate."

4. Mark P. Kritzman, in the MIT article This new forecasting model is better than machine learning, on the Relevance-Based Prediction (RBP) approach: "Model-based machine learning looks at historical data to form a prediction, but if circumstances change in the future — if something unprecedented happens — that model is no longer valid and you have to start over." His method measures both unusualness and similarity, using a sub-sample of relevant observations.

What lies ahead

Applying all of this is challenging, but I'm confident we'll reach good results and learn to navigate the new geopolitical dynamics. I acknowledge flaws in the model, but it's improvable. As Ella Peltonen of the University of Oulu (Finland) says: "Researchers realized that, to avoid repeating mistakes, it was necessary to discuss the practical problems of studies and the failed results that don't get published." It's in Illuminating 'the ugly side of science' (Nature).

I'll close with a quote from Steven Levitt in his interview with Paul M. Rand (University of Chicago, October 2023): "Correlation often isn't what we want. We want causality. Causality is harder to establish than correlation. What techniques, in this context, can we do well to get there?"

There's a lot of work ahead.