---
name: teaching-compare-llm-signal-p
description: |
  Learned from user correction: Teaching: compare llm signal performance against simple baselines (random, momentum, mean-reversion) to detect overfitting | --- ### 3 | Correction: User wants 'seen' instead of 'have' | Avoid: 'happened - incorrect causal reasoning about market movements - hallucinated case citations (already documented in legal contexts -- sec chair gensler explicitly cited this risk in his feb 2024 yale law school speech) sec chair gensler stated directly: *"if an ai model can hallucinate a bad case citation' | Expected behavior: 'be calibrated against actual accuracy (track hit rate) - reasoning field enables human review and audit trail ### 2'
author: Claude Code (extracted by Claudeception v4.0)
version: 1.0.0
date: 2026-03-26
tags: ["correction", "learned", "wrong_assessment"]
level: user
breakthrough_score: 154.25
---

# Teaching: compare llm signal performance against simple base

## Problem / Use Case

<task-notification>
<task-id>a529bd5889117cd40</task-id>
<tool-use-id>toolu_bdrk_01PoprtXWRkyVmbHLaM5psxv</tool-use-id>
<output-file>/private/tmp/claude-504/-Users-cajias-Projects-net-zero/75585e92-35

## When to Use This Skill

When making similar mistakes to: wrong_assessment

## Solution / Approach

Teaching: compare llm signal performance against simple baselines (random, momentum, mean-reversion) to detect overfitting | --- ### 3 | Correction: User wants 'seen' instead of 'have' | Avoid: 'happened - incorrect causal reasoning about market movements - hallucinated case citations (already documented in legal contexts -- sec chair gensler explicitly cited this risk in his feb 2024 yale law school speech) sec chair gensler stated directly: \*"if an ai model can hallucinate a bad case citation' | Expected behavior: 'be calibrated against actual accuracy (track hit rate) - reasoning field enables human review and audit trail ### 2'

## Verification

- Apply the corrected approach and verify user acceptance

## Extraction Context

- Extracted automatically by Claudeception v4.0
- Breakthrough score: 154.25
- Classification: user
- Confidence: 0.85
- Corrections detected: 3