Learning to Trust Investment Rules: Holland’s Bucket Brigade for Credit Assignment in a Production Rule Advisory System

Abstract

We present a system that organizes investment knowledge into four knowledge bases—Graham, Buffett, dividend investing, and general market heuristics—each grounded in citable, published thresholds. The system employs Holland’s bucket brigade algorithm to learn rule reliability from backtested outcomes. Multiple rules recommend the same action (e.g., SELL) as they form a coalition; after the outcome is known, credit is distributed proportionally to each rule’s contribution. A three-tier strength model tracks reliability at the per-stock, per-sector, and universal levels, revealing that a rule may work well for technology stocks but poorly for utilities—without explicit per-context rules.We evaluate each knowledge base on 87 stocks across 156 weeks using a market-relative triple barrier method inspired by López de Prado [1], measuring alpha vs. the S&P 500. Graham’s BUY signals achieve 67.8% accuracy—the recommended stock outperforms the market by ≥8% before underperforming by 8% two-thirds of the time. For Graham, the bucket brigade improves actionable signal accuracy (BUY/SELL) from 59.0% to 60.7%. Critically, the symbolic rule representation guards against temporal overfitting: every rule expresses a general financial condition, and walk-forward validation confirms that out-of-sample accuracy matches in-sample performance.