Abstract
This paper presents an early Expert System (E.S.) designed to automate the complex process of econometric modeling. By integrating a LISP-based reasoning engine with specialized statistical software, the system assists researchers in specifying and estimating production functions through an interactive interface. The approach codifies economic "rules of thumb" and theoretical knowledge, bridging the gap between qualitative economic theory and quantitative statistical analysis.
Introduction
Building econometric models—mathematical representations of economic relationships—is traditionally a labor-intensive task requiring high-level expertise in both economic theory and statistics. Researchers must navigate numerous decisions, such as variable selection (e.g., labor vs. capital) and choosing appropriate functional forms (e.g., "putty-clay" vs. "clay-clay" models).
Jean-Louis Roos identifies a critical role for AI in this domain: acting as an intelligent assistant that guides the modeling process. The goal of the described system is to explain and calculate the factors of aggregate demand for the manufacturing sector, effectively acting as a digital consultant for economists.
System Architecture and Methods
The system is built on a hybrid architecture that distinguishes between qualitative reasoning and quantitative calculation. This separation allows for specialized processing in each domain.
The Reasoning Component (LISP)
The "brain" of the system is written in LISP and utilizes frames—a classic AI data structure—to represent components of a production function like labor, capital, and investment. The inference engine employs forward-chaining logic: as users provide facts, the system uses IF-THEN production rules to derive new information or prompt for necessary follow-up data.
The rule hierarchy is divided into three categories:
- Error-rules: Ensure logical consistency in user inputs.
- Added-fact rules: Drive the logical progression of the model specification.
- Meta-rules: Define high-level search strategies, such as skipping irrelevant questions based on the user's stated interests.
The Econometric Component (TSP)
Once the reasoning engine has populated the necessary frames, it generates a parameter file for the Time Series Processor (TSP). This specialized statistical package performs the actual regressions using a dedicated Fortran library and a database of quarterly macroeconomic data.
Key Insights and Ideas
- Separation of Concerns: Long before modern "LLM-plus-tool" architectures, this system proposed a clear split between a symbolic AI "reasoner" (LISP) and a numerical "calculator" (TSP/Fortran).
- Interactive Explainability: A core feature is the system's transparency. Users can query "WHY" a question is being asked or "HOW" a conclusion was reached, making the modeling process less of a "black box."
- Theoretical Flexibility: The system is not dogmatic; it can evaluate various capital-labor substitution models based on user input, providing a flexible tool for theoretical exploration.
Performance and Results
The system demonstrated significant depth and efficiency for its time:
- Library Depth: The econometric component supports approximately 1,000 different forms of econometric estimations.
- Comprehensive Modeling: A single session can generate up to five consistent equations covering labor, investment, capital, and capacity utilization.
- Efficiency: A complete session, including the econometric estimation, takes roughly 3 minutes.
- Portability: The author noted that a micro-version was planned for the Apple II using the PROLOG language, indicating early interest in desktop accessibility for AI tools.
Limitations and Future Work
Despite its innovations, the system faced several constraints. The natural language interface was a constrained syntactic analyzer rather than a true natural language processor. Additionally, while the system could run models, the component intended to interpret statistical results and provide economic explanations was still in the planning stages at the time of publication. Future work was also envisioned to expand the system's scope beyond production functions to broader macroeconomic theory.
Summary updated