Abstract
FOLIO is an early expert system designed to bridge the gap between qualitative, often uncertain client interviews and the rigid mathematical requirements of portfolio optimization. By combining a heuristic production system for inferring investment goals with a goal programming algorithm for asset allocation, it provides a "best-fit" investment strategy across nine security classes. This hybrid approach allows the system to remain sensitive to the uncertainty implicit in client data while mitigating common human cognitive biases.
Introduction
Modern Portfolio Theory (MPT) provides sophisticated mathematical tools for asset allocation, yet these tools often struggle in real-world applications due to their reliance on precise data. In practice, a client’s risk tolerance, tax status, and investment goals are frequently "incomplete, inaccurate, and sometimes even contradictory." This creates a significant challenge for human advisors, who are skilled at interviewing but prone to forming premature "preconceptions" and ignoring evidence that contradicts their initial hunches.
Conversely, standard optimization algorithms, such as quadratic programming, are precise but often fail when faced with the fuzzy, conflicting constraints of a real human life. FOLIO was developed to serve as an assistant that bridges this gap, providing a mathematically optimized recommendation that is robust to the uncertainty of human-provided data.
System Architecture
FOLIO utilizes a hybrid architecture consisting of three distinct components designed to handle the transition from qualitative data to quantitative results:
- Interview Functions: A simple interface used to collect raw data from the client during the initial consultation.
- Forward-Chaining Production System: This component serves as the "expert" front-end. It utilizes approximately 50 heuristic rules to translate the client's messy interview answers into 14 specific investment goals, such as "preserve capital," "hedges," or "taxable interest."
- Goal Programming Algorithm: Rather than traditional linear programming, which requires a single objective function, FOLIO employs goal programming. This treats investment goals as "targets" rather than hard constraints, allowing the system to minimize the "summed deviation" from these targets using a modified two-phase simplex algorithm.
Notably, the system performs asset allocation rather than individual security selection. It recognizes nine "funds" or classes of securities (e.g., low-risk stocks, municipal bonds, growth stocks). This abstraction allows the system to remain stable even as individual stock prices fluctuate, leaving the final task of "security analysis" to the human advisor.
Bridging Heuristics and Optimization
The most compelling innovation in FOLIO is the use of heuristics as a front-end for mathematical optimization. By using AI rules to handle the "fuzzy" human interface and then passing the resulting "refined" goals to a rigorous mathematical optimizer, FOLIO achieves a balance between human-like reasoning and computational precision.
To prevent the optimization from failing when goals conflict—such as a client desiring both high growth and high dividends—FOLIO utilizes soft constraints via penalty functions. Each goal is assigned a target value and a "penalty" for falling short or exceeding it. This allows the algorithm to find a "relaxation" that best satisfies the most important goals while acknowledging the inherent trade-offs in the client's profile.
Furthermore, by focusing on nine asset classes rather than thousands of individual stocks, the system's knowledge base remains manageable. This design choice ensures that the system does not require constant updates as individual company fortunes change, maintaining its stability and relevance over time.
Implementation and Performance
FOLIO was implemented as a hybrid system to leverage the strengths of different programming environments. The core logic and heuristic rules were developed in Franz Lisp, while the heavy-duty array pivoting required for the simplex algorithm was handled by Pascal "foreign functions."
In performance evaluations conducted on a "lightly loaded DEC 11/730," the system was able to produce an optimized portfolio for a client with three or four accounts in approximately five minutes. Qualitatively, the system successfully demonstrated its ability to resolve multiple, sometimes contradictory, client goals into a portfolio that provides an "optimum fit."
Limitations and Future Directions
Despite its promising results, FOLIO had several limitations at the time of its reporting. It was "as yet untested" in a live production environment, though it was being trialed on existing accounts. The authors also identified a need for a more robust "theory of reasoning about uncertainty" to handle cases where multiple heuristic rules suggest different values for the same goal parameter.
Additionally, the system remains an "assistant" rather than a replacement for human expertise. It still requires a human investment advisor to perform the final step of "instantiating" the recommended asset classes with specific securities. Finally, while aggregate figures change slowly, the system still requires manual maintenance of the characteristics of the nine security funds it recognizes.
Conclusion
FOLIO represents a significant step forward in the application of expert systems to financial management. By integrating heuristic reasoning with goal programming, it addresses the "garbage-in, garbage-out" problem inherent in applying rigid mathematical models to messy real-world data. Its ability to mitigate human bias while providing optimized recommendations makes it a valuable prototype for the next generation of financial decision-support tools.
Summary updated