Right Answer, Wrong Reasons: AI Forecasts Costco - Hassan El Mghari & Omar El Mghari, Together AI

An AI model (GPT-OSS-120B) could construct a structured financial forecasting model for Costco's membership fee increase that landed within 4.4% of the actua

Right Answer, Wrong Reasons: AI Forecasts Costco

Abstract

This synthesis examines a controlled experiment testing whether a large language model (LLM) can construct a defensible financial forecasting model, using Costco Wholesale's July 2024 membership fee increase as a natural experiment. GPT-OSS-120B, selected for its pre-event training cutoff, generated structured forecasting formulas across 16 parallel sessions at a total compute cost of $0.19, producing forecasts averaging 4.4% error against realized outcomes. Audit revealed this accuracy was illusory: the model fabricated an executive-tier membership share of 15% against a disclosed figure of 45%, while separately overstating the affected member base by including international members outside the fee increase's scope. These two errors canceled, and correcting only one in isolation degraded accuracy to 28%. The findings demonstrate that engineering-level validation cannot substitute for domain-expert review, motivating a three-lines-of-defense control architecture for agentic financial modeling.

1. Introduction

Large language models are increasingly deployed to build quantitative financial models, yet the standard method of evaluating such systems - comparing predicted outcomes to actual results - contains a latent flaw. A forecast can land close to a realized value not because the underlying reasoning is sound, but because multiple errors in the model's assumptions happen to offset one another. This produces a system that passes outcome-based validation while remaining structurally broken, a failure mode this analysis terms right answer, wrong reasons.

The experiment under review isolates this phenomenon using Costco's July 2024 membership fee increase, the company's first since 2017, which raised the standard tier from $60 to $65 and the executive tier from $120 to $130. Costco presents an unusually clean test case: it generates approximately $300 billion in annual merchandise sales at near-cost margins, with the bulk of its roughly $12 billion in operating profit derived from membership fees rather than merchandise markup. A fee change therefore flows through a single, well-documented accounting channel - deferred revenue recognition, under which membership fees are recognized ratably over the 52-week membership term rather than immediately upon payment.

The central question is whether an LLM, prevented from simply recalling the outcome, can construct a forecasting model whose assumptions are individually correct, and whether automated engineering checks are sufficient to certify such a model without human review. This paper traces the experimental design, the model-building pipeline, the audit that exposed the error-cancellation effect, and the control framework the results imply for deploying AI-built financial models in practice.

2. Background and Related Work

Two practices inform the experimental design. The first is deferred revenue accounting itself: because fee revenue phases in gradually as memberships renew, a valid forecasting model must encode timing explicitly rather than treat a fee increase as an immediate, one-time revenue bump. The second is the three lines of defense framework drawn from internal audit practice, which separates assurance into (1) management controls embedded directly in a process, (2) an independent internal audit function applying domain-specific checks, and (3) external or independent review. The experiment maps this framework onto an agentic workflow: the model-building LLM constitutes the first line, codified domain-specific accounting checks constitute the second line, and a human subject-matter expert constitutes the third.

A critical methodological precondition is data contamination control. Any model trained on data after July 2024 already knows the outcome of the fee increase, making its "forecast" a retrieval task rather than a reasoning task. The experimenters found that contemporary open-weight models, including GLM-5.3 and Chimera-3, already possessed this knowledge, and therefore selected GPT-OSS-120B and Llama-3.3-70B for their earlier training cutoffs, ensuring the forecasting task was genuine rather than retrospective.

3. Core Analysis

3.1 Corpus Construction and Leakage Prevention

The experimenters scraped 214 public Costco filings and decomposed them into approximately 3,000 timestamped facts. These facts were filtered by publication date to retain only information available before the fee-increase announcement, ensuring the model could not draw on post-hoc knowledge even indirectly through disclosed figures published after the cutoff. This filtered corpus was the sole factual input provided to GPT-OSS-120B when constructing its forecasting model.

3.2 Structured Model Construction and Parallel Validation

Rather than asking the model to compute a forecast directly, the experimenters had it output structured formulas and assumptions as JSON - for example, encoding membership fee revenue as recognized "ratably" over a "52 weeks" term, correctly reflecting deferred revenue accounting. This produced 14 formulas alongside a set of explicit assumptions, such as the proportional split between executive and standard membership tiers and between Canadian and U.S. membership populations.

Sixteen parallel sessions of GPT-OSS-120B were run to build independent models, each subjected to basic engineering checks - for instance, verifying that formulas were not self-referential and that a fee increase necessarily raised revenue. Failed runs received error feedback and were permitted up to two rounds of self-correction. Only 5 of the 16 runs passed all checks, at a total compute cost of $0.19. This result highlights a practical implementation strategy: running many cheap parallel attempts and filtering by validation outcome outperformed iteratively refining a single prompt.

3.3 Outcome Accuracy and the Illusion of Validation

The five passing models, once finalized, executed in roughly 10 milliseconds and collectively predicted that Costco's profit would rise by $560 million over two years as a result of the fee increase. Averaged across the five passing builds, forecast error against the real outcome was 4.4%, with individual builds ranging from 0.7% to 5.2%. On outcome accuracy alone, this would appear to validate the model-building pipeline.

3.4 Audit Findings: Error Cancellation

Human audit, prioritizing assumption review over output review, uncovered two compensating errors. First, the model assumed a 15% executive-tier membership share, citing a real but unsupported range of 5-30%, when Costco's actual disclosed figure was 45%. Second, the model applied the fee increase to Costco's entire international membership base rather than restricting it to the U.S. and Canadian population actually affected by the change, overstating the impacted base.

These errors moved the forecast in opposite directions and canceled out, producing the deceptively accurate 4.4% result. When the executive-share assumption alone was corrected from 15% to 45%, error increased from 4.4% to 28%, demonstrating conclusively that the original accuracy reflected luck rather than correct reasoning. The properly recalculated profit impact was approximately $450 million - distant from both the model's $560 million estimate and the much wider $0.5-$6 billion range spanning all 16 original runs. A further diagnostic finding was that one assumption lever, the U.S./Canada membership split, had zero measurable effect on sales and fee outputs, indicating a non-functional or disconnected input within the model.

4. Technical Insights

Several implementation-level findings carry direct relevance for practitioners building agentic financial modeling systems. Structuring model output as explicit formulas and JSON-encoded assumptions, rather than having the LLM perform arithmetic internally, materially improves auditability: every assumption becomes individually inspectable and testable in isolation, which is precisely how the executive-share error was isolated and corrected. Running many cheap parallel sessions (16 runs for $0.19) and filtering by pass/fail engineering checks is a more effective validation strategy than single-prompt iteration, though it is notable that only 5 of 16 runs passed even basic structural checks, underscoring that automated checks catch a meaningful but limited class of errors.

Critically, the engineering checks applied - non-self-referential formulas, directionally correct fee-to-revenue relationships - are necessary but insufficient. They cannot detect a fabricated statistic attached to a real citation, nor can they detect scope errors such as misapplying a fee increase to an unaffected population. Both failures are domain-specific factual errors invisible to purely structural validation, and the wide output range across the 16 raw runs ($0.5B-$6B) versus the narrower range of the 5 "passing" runs illustrates that passing engineering checks substantially narrows but does not eliminate the space of materially wrong models.

5. Discussion

The central implication is that outcome-based accuracy is a poor sole criterion for validating AI-generated financial models, particularly in low-sample, single-event forecasting contexts where compensating errors are plausible. This risk is likely to scale with model complexity: as the number of assumptions embedded in a structured model increases, so does the combinatorial opportunity for offsetting errors to produce spuriously plausible outputs. This suggests that confidence in an AI-built financial model should be grounded primarily in assumption-level audit rather than output-level comparison, reversing the intuitive validation order.

The three-lines-of-defense mapping offers a generalizable control architecture: once a subject-matter expert identifies a domain-specific error such as the 45% executive-tier share, that check can be codified permanently into the second line of defense, progressively reducing the burden on the third line over time. This positions human expert review not as a one-time gate but as a mechanism for continuously hardening the automated checks that precede it. An open question is how many audit cycles are required before second-line checks capture a sufficient share of domain-specific error modes to reduce third-line review to a sampling function rather than a comprehensive one.

6. Conclusion

This experiment demonstrates that an LLM-built financial model can achieve a seemingly strong 4.4% forecast error while containing two significant, uncorrelated factual errors that happened to cancel. Correcting a single error in isolation degraded accuracy sixfold, to 28%, confirming that the apparent precision reflected error cancellation rather than correct domain reasoning. The practical takeaway is direct: engineering-level checks on AI-generated financial models are necessary but not sufficient, and a human subject-matter expert must audit assumptions - not merely outputs - before such models are trusted for decision-making. Organizations deploying agentic financial modeling systems should adopt a three-lines-of-defense architecture, treating human expert review as an active contributor to progressively codified domain checks rather than a terminal rubber stamp.

Sources

• Right Answer, Wrong Reasons: AI Forecasts Costco - Hassan El Mghari & Omar El Mghari, Together AI - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub