Self-Explanatory Financial Planning Models

Donald W. Kosy, Ben P. WiseAAAI 1984

Introduction

This 1984 paper from Carnegie-Mellon University addresses a classic frustration in business: a spreadsheet tells you what happened, but it rarely tells you why. The authors introduce a system called ROME (Reason-Oriented Modeling Environment) that treats financial formulas as a knowledge base to generate natural language explanations for unexpected business results. ROME provides a method for "self-explanatory" financial models that can automatically answer "why" questions about budget variances and performance dips. By combining sensitivity analysis with natural language templates, the system identifies the most significant drivers of change and translates complex algebraic relationships into human-readable narratives.

The Problem: Data Presentation vs. Data Explanation

In the early 1980s, the rise of "financial modeling languages" allowed managers to project complex scenarios, but these systems remained "black boxes" regarding causality. If profit dropped despite an increase in sales, a manager had to manually hunt through rows and columns to find the culprit. The authors identify a gap between data presentation and data explanation. While systems could compute averages or ratios, they couldn't explain the derivation or the significance of a result in the context of a user's expectations.

Methodology: The ROME Architecture

The authors treat the financial model as a directed acyclic graph (DAG) of formulas. To generate an explanation, the system performs several key steps:

  1. Question Interpretation: Using a pattern-matching parser, the system identifies the "focus" (e.g., profit), the "referent" (e.g., last year's profit), and any "presuppositions" (e.g., "even though sales went up").
  2. The $\epsilon$ (Effect) Measure: To determine how much a variable $X$ contributed to a change in $y$, the system calculates: $$\epsilon(X, y) = y_2 - f(Z)$$ Where $f(Z)$ is the value $y$ would have taken if the variables in $X$ had stayed at their old values while everything else changed.

  3. Significance Testing: The system uses a threshold ($\theta = 0.75$) to find the smallest set of variables that account for the majority of the change.

  4. Template-Driven Generation: Once the "culprits" are identified, the system uses linguistic templates to construct sentences like, "Profit went down... primarily because overhead cost went up."
Underlying formulas and assumptions for the ROME financial model.

Case Study: The Battery Company Model

The paper uses a simplified version of the "Battery Company" model, a classic management case study. This model includes five years of data (1972–1973 historical, 1974–1976 projected) and covers variables such as volume, selling price, labor/unit, material costs, overhead, and profit.

Financial data for the Battery Company model (1972-1976).

Key Insights and Results

The authors demonstrated that ROME could successfully handle complex, multi-step queries. A "truthful" explanation (showing all the math) is often a "bad" explanation; a good explanation must distinguish significant effects from insignificant noise.

For the prompt: "Why did profit go down in 73 even though gross sales went up?", the system generated a tiered response:

  1. It identified that overhead cost increases outweighed gross margin gains.
  2. It quantified the conflict: gross margin rose by 19%, but overhead rose by 41%.
  3. It traced the overhead increase specifically to operating expenses and management salaries.

The system also addresses "long-distance" effects, tracking how a variable might be the primary driver of a change at a low level (e.g., shipping costs) even if it seems insignificant to the top-line result.

Limitations and Future Work

The technique works for hierarchical, algebraic models but fails for iterative computations where causality is circular or non-linear. Additionally, the system understands the algebra rather than the business context itself; making the coefficients in an evaluation function "meaningful" to a user remains a challenge. Future work aims to formalize the handling of "why not" questions and improve linguistic modeling.

Summary updated

BibTeX
@inproceedings{kosy1984self,
  author = {Donald W. Kosy and Ben P. Wise},
  title = {{Self-Explanatory Financial Planning Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1984},
}
Figure