---
id: wiki-2026-0508-outside-thinking
title: Outside Thinking
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Outside View, Reference Class Forecasting, Outsider Perspective]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [decision-making, cognition, forecasting, biases]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: theory
  framework: behavioral-decision-theory
---

# Outside Thinking

## 매 한 줄
> **"매 your project is not special — base rates always win."**. 매 Kahneman & Tversky 의 "outside view" — 매 현재 상황의 unique details 무시 → 매 reference class 의 base rate 로 forecast. 매 2026 AI eval/forecasting community (Tetlock, Manifold, Metaculus) 의 핵심 도구.

## 매 핵심

### 매 inside vs outside
- **Inside view**: 매 plan 의 details 로부터 outcome 추정 ("우리는 매 6주 만에 끝낼 수 있어").
- **Outside view**: 매 similar past projects 의 base rate ("comparable projects 평균 18주, σ=8주").
- **Result**: 매 outside view 가 거의 항상 더 정확 — 매 planning fallacy 회피.

### 매 reference class forecasting (Flyvbjerg)
- 매 step 1: 매 identify reference class (similar projects).
- 매 step 2: 매 collect distribution of outcomes (cost, time, success rate).
- 매 step 3: 매 your project = sample from that distribution.
- 매 step 4: 매 adjust only with strong evidence.

### 매 응용
1. Software estimation: 매 "this PR will take 1 day" → 매 historical median = 4 days.
2. Startup success: 매 "we'll be the exception" → 매 base rate ~10% survive 5y.
3. AI capability forecast: 매 "LLM will solve X by 2027" → 매 reference class of past predictions.

## 💻 패턴

### Pattern 1: Reference class forecaster
```python
import numpy as np

def outside_forecast(reference_class_outcomes: list[float],
                     inside_estimate: float,
                     trust_in_inside: float = 0.2):
    """매 Bayesian blend — 매 prior is base rate."""
    base_rate_mean = np.mean(reference_class_outcomes)
    base_rate_std = np.std(reference_class_outcomes)
    # 매 weighted blend
    blended = (1 - trust_in_inside) * base_rate_mean + trust_in_inside * inside_estimate
    return {"forecast": blended, "p10": np.percentile(reference_class_outcomes, 10),
            "p90": np.percentile(reference_class_outcomes, 90)}
```

### Pattern 2: Estimation poker with history
```python
def estimate(task, similar_tasks_db):
    similar = find_similar(task, similar_tasks_db, k=10)
    durations = [t.actual_duration for t in similar]
    return {
        "p50": np.median(durations),
        "p90": np.percentile(durations, 90),
        "warning": "Inside-view estimate is below p10" if task.guess < np.percentile(durations, 10) else None,
    }
```

### Pattern 3: Pre-mortem — outside view of failure modes
```python
def pre_mortem(project, similar_failed_projects):
    """매 imagine project failed; 매 list reasons from history."""
    failure_modes = []
    for fp in similar_failed_projects:
        failure_modes.extend(fp.post_mortem_causes)
    return Counter(failure_modes).most_common(10)
```

### Pattern 4: Prediction market calibration
```python
# 매 force outside view via market — 매 your private estimate vs market price
def confidence_check(my_p, market_p):
    if abs(my_p - market_p) > 0.20:
        return "RED FLAG: large divergence from outside view"
    return "OK"
```

### Pattern 5: Survivorship bias correction
```python
def correct_for_survivorship(success_stories, full_population):
    survivor_rate = len(success_stories) / len(full_population)
    return {
        "naive_lesson": "Do what successes did",
        "corrected": f"Only {survivor_rate:.0%} survive — failures often did same things",
    }
```

### Pattern 6: LLM as outside view oracle
```python
PROMPT = """For the following plan, list:
1. The reference class (similar past projects)
2. Base rate of success
3. Typical failure modes
4. Why this project might/might-not be representative
"""
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| 매 estimating new project | Outside view first, inside view as adjustment |
| 매 confident in unique advantage | Outside view with small inside-view weight |
| 매 forecasting AI capabilities | Reference class of past predictions |
| 매 startup go/no-go | Compare to founder cohort base rates |
| 매 research timeline | Reference class of similar papers/benchmarks |

**기본값**: 매 outside view first, inside view as 매 small adjustment (≤20% weight).

## 🔗 Graph
- 부모: [[Decision Theory]] · [[Behavioral Economics]]
- 변형: [[Reference Class Forecasting]]
- 응용: [[Forecasting]]

## 🤖 LLM 활용
**언제**: 매 estimation, 매 forecasting, 매 strategic planning, 매 evaluating "we're different" claims.
**언제 X**: 매 truly novel domains where no reference class exists (rare — usually a class can be found).

## ❌ 안티패턴
- **"Our project is unique"**: 매 99% of the time, not unique enough to escape base rates.
- **Cherry-picked reference class**: 매 selecting only successes — 매 survivorship bias.
- **Ignoring distribution**: 매 only using mean — 매 use p10/p90.
- **No update mechanism**: 매 collecting new data but not updating reference class.

## 🧪 검증 / 중복
- Verified (Kahneman 2011, Flyvbjerg 2006, Tetlock 2015).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — outside vs inside view, reference class forecasting |