I think it is pretty clear to everybody who follows this blog and follows me on social media that I am deeply fascinated by AI and how we can use, for example, Large Language Models in economic research—and I am eager to spread the gospel.
Today, I spent a couple of hours building a simple macroeconomic simulation model in Claude 3.5 Sonnet (see more about that here).
The model is a rather simple AS/AD model for the US economy in its most simple form. Test the model here.
When I had done the model, it struck me that I could test the output (graphs) from the model to see whether Large Language Models like ChatGPT or Claude can do macroeconomic reasoning based on “looking” at graphs of economic data.
In the model, I simulated a cut in interest rates by the Fed. The model, among other things, produced the graphs below.
I then simply copied these two graphs directly into a prompt in ChatGPT 4o and asked it the following question:
“Tell me about this. From an economic-theoretical perspective, what type of shock is this?”
And this is how ChatGPT 4o replied:
The graphs show two time series: GDP growth (%) and inflation (%), each plotted against a time index. Here’s a detailed analysis from an economic-theoretical perspective:
GDP Growth (%)
- Initial State: At the beginning (from -5 to 0), GDP growth is relatively stable at around 2.00%.
- Shock Occurrence: At time 1, there’s a sharp increase in GDP growth, peaking around time 2.
- Post-Shock: After the peak, GDP growth declines and stabilises back to around the initial level by time 8, remaining stable thereafter.
Inflation (%)
- Initial State: Inflation is stable at 2.00% from time -5 to 0.
- Shock Occurrence: A significant increase in inflation begins at time 1, peaking at time 3 and stabilising at a high level until time 6.
- Post-Shock: Post time 6, inflation sharply decreases and stabilises around its initial level by time 8, similar to GDP growth.
Type of Economic Shock
Given the patterns observed in both GDP growth and inflation, the economic shock depicted in these graphs can be analysed as follows:
Supply-Side Shock (Supply Shock)
- Characteristics: A supply shock, such as a sudden increase in productivity or an external factor reducing production costs, can lead to a temporary increase in GDP growth and inflation. This is because the economy can produce more goods and services, thus boosting GDP, while the increased supply might temporarily push prices up.
- Recovery Phase: Once the effects of the shock dissipate, both GDP growth and inflation return to their previous levels, as seen in the graphs.
Demand-Side Shock (Demand Shock)
- Characteristics: A demand-side shock, such as an increase in consumer spending or investment, typically raises both GDP growth and inflation. The initial increase in demand drives up production (GDP growth) and prices (inflation).
- Recovery Phase: Once the demand pressure subsides, both GDP growth and inflation return to their prior states, matching the pattern observed in the graphs.
Conclusion
The graphs suggest a positive demand-side shock due to the simultaneous increase in both GDP growth and inflation. This type of shock might be caused by factors such as increased government spending, tax cuts, or a surge in consumer confidence leading to higher spending. The temporary nature of the shock, with both GDP growth and inflation eventually returning to their baseline levels, supports the identification of a demand-side shock.
Additional Considerations
- Monetary Policy Response: Central banks might respond to such shocks by adjusting interest rates to manage inflation, which can also influence GDP growth.
- External Factors: External economic events, such as changes in global trade policies or commodity prices, could also contribute to these observed patterns.
Overall, the graphs illustrate a typical economic response to a temporary positive demand-side shock.
I know a lot of economists who would answer a lot less correctly than this, and even though I am quite bullish, to say the least, on the abilities of Large Language Models, I must say that I am extremely positively surprised by this. It should be remembered that ChatGPT 4o is not a model designed or trained to be able to do macroeconomic reasoning, but we clearly see here that it can perform this task rather well.
Imagine then when you train a model specifically for macroeconomic reasoning. We are getting quite close to virtual economists—and imagine what that will do, for example, in finance and central banking.
If you want to know more about my work on AI and data, then have a look at the website of PAICE — the AI and data consultancy I have co-founded.