# Incremental Belief Updating

> Adjust forecasts and beliefs gradually as new evidence arrives rather than making dramatic shifts or anchoring on initial estimates

- Skill: `lev-os/incremental-belief-updating` (Agent Skill)
- Install (CLI): `npx skillmds@latest add lev-os/incremental-belief-updating`
- Raw SKILL.md: https://api.skillmd.com/api/skills/lev-os/incremental-belief-updating/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: lev-os (https://skillmd.com/u/lev-os)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/lev-os/incremental-belief-updating

---


# Incremental Belief Updating

**What**: A forecasting and reasoning technique where you continuously update your probability estimates by small increments as new information arrives, rather than sticking to initial beliefs or making large sudden jumps.

**When to use**: When making predictions or estimates in uncertain domains where new evidence arrives over time—project timelines, market forecasts, risk assessments, or any judgment requiring ongoing calibration.

**Introduced by**: Philip Tetlock and colleagues through the Good Judgment Project (2011-2015), studying superforecasters

## Core Mechanism

**Bayesian-inspired updating**: Each piece of new evidence should shift your probability estimate proportionally to its strength and relevance. Small evidence → small update. Large evidence → large update. No evidence → no update.

**Key contrast**:
- **Anchoring**: Stick to initial estimate despite new data
- **Overcorrection**: Swing wildly with each new headline
- **Incremental updating**: Adjust systematically in proportion to evidence weight

**Why it works**: Reality reveals itself gradually. Superforecasters who update frequently (but not dramatically) calibrate toward accuracy faster than those who anchor or overreact.

## When to Apply

**Use incremental belief updating when:**
- Making predictions in domains with evolving information
- Estimating project timelines, budgets, or resource needs
- Assessing risks or probabilities over time
- Evaluating market conditions, competitive moves, or strategic options
- Tracking leading indicators for future outcomes

**High-value contexts:**
- Forecasting tournaments or prediction markets
- Roadmap planning and deadline estimation
- Risk management and threat assessment
- Investment decisions with ongoing information flow
- A/B test analysis as data accumulates

## Execution Steps

### 1. Set Initial Baseline Estimate
Start with a base rate or reference class forecast. What's the historical frequency of similar outcomes? This is your prior.

### 2. Express Uncertainty as Probability
Instead of binary (will/won't), use probabilities: "60% chance of shipping by Q2." Enables precise updates.

### 3. Identify Information Triggers
What evidence would update your belief? Define upfront: "If X happens, I'll adjust by ~10% toward Y."

### 4. Update Small and Often
When new evidence arrives, adjust your estimate proportionally. Tetlock's superforecasters updated every few days, not just after major news.

### 5. Track Direction and Magnitude
Document: What changed? How strong is the evidence? By how much should I update? Did I update too much or too little?

### 6. Avoid Round Numbers
Don't snap to 50%, 75%, 90%. Precise estimates (63%, 78%) reflect granular thinking and force you to process evidence carefully.

### 7. Review Calibration Over Time
After outcomes resolve, compare your probability trajectory to reality. Were you too slow to update? Too fast? This trains judgment.

## Real-World Applications

**Good Judgment Project**: Superforecasters who updated their predictions frequently (average 5+ times per question) and incrementally outperformed intelligence analysts with classified data.

**Project Delivery Estimates**: Instead of "We'll ship in 6 weeks" (then missing deadline), track probabilities: Start at 40% on-time → Sprint 1 blockers emerge: 30% → Workaround found: 45% → Scope reduced: 65%.

**Market Analysis**: Rather than "The market will crash" or "Bull run forever," track: "Recession probability: 25% → Fed minutes hawkish: 32% → Unemployment data weak: 28% → Credit spreads widen: 35%."

**Incident Severity Assessment**: Initial alert: 60% chance this is customer-impacting → More users reporting issues: 80% → Root cause identified in non-critical service: 40% → Workaround deployed: 15%.

## Key Indicators

**Signs you're doing it well:**
- Probability estimates drift gradually, not jump wildly
- You can articulate what evidence would change your mind
- Updates correlate with information quality, not just recency
- Calibration improves over time (Brier scores decrease)

**Red flags:**
- Estimates haven't moved in weeks despite new information
- Large swings based on headlines without substance analysis
- Can't explain why you're at 70% vs. 65%
- Round numbers (10%, 50%, 90%) dominate estimates

## Common Mistakes

**Under-updating (anchoring)**: Sticking too close to initial estimate despite contradictory evidence. Tetlock found this more common than over-updating.

**Over-updating (recency bias)**: Swinging dramatically with latest news without weighting historical base rates.

**Binary thinking**: "It will happen" or "It won't" leaves no room for calibration. Use probabilities.

**Ignoring base rates**: Starting from intuition rather than reference class frequencies leads to miscalibration.

## Related Frameworks

**Complementary**: Bayesian Reasoning (formal mathematical foundation), Superforecasting (broader methodology), Reference Class Forecasting (establishing priors)

**Contrasting**: Gut Feel (no systematic updating), Prediction Markets (crowd-based), Scenario Planning (qualitative branches)

**Sequential**: Establish base rate → Set initial probability → Monitor information sources → Update proportionally → Track calibration → Adjust update strategy

## Practical Examples

**Feature Launch Timeline**:
Week 1: "70% chance we launch by end of quarter" (based on past launches)
Week 3: Backend integration more complex than expected → 55%
Week 5: Frontend team ahead of schedule → 60%
Week 7: QA finds critical bug → 40%
Week 9: Bug fixed, testing clean → 70%
Result: Launched 2 days into next quarter. Probabilities tracked reality better than binary prediction.

**Hiring Candidate Assessment**:
After resume: 40% chance of hire (base rate for this role)
After phone screen: 60% (clear communication, relevant experience)
After technical interview: 50% (strong on algorithms, weak on system design)
After team fit interview: 70% (excellent culture match, team excited)
After reference checks: 65% (one lukewarm reference, others strong)
Decision: Make offer (hired, worked out well).

**Security Threat Level**:
Start: 5% chance of attack this month (historical rate)
Unusual traffic pattern detected: 8%
Pattern matches known bot signature: 12%
Geo-source is known threat actor region: 25%
Rate limiter successfully blocking: 10%
Final assessment: Elevated but contained risk, maintain monitoring.

## Measurement

**Quantitative signals:**
- Brier Score (measures calibration, lower is better)
- Update frequency (superforecasters average 5-10 updates per question)
- Update magnitude distribution (should be mostly small, few large)
- Correlation between evidence strength and update size

**Qualitative indicators:**
- Can articulate reasoning for each update
- Written logs of what triggered updates
- Diversity of information sources considered
- Willingness to update in both directions (up and down)

## Scoring Criteria

**Practitioner Weight**: 10/10 — Tetlock's research directly studied thousands of real forecasters making real predictions with measurable accuracy; superforecasters validated the approach

**Clarity & Executability**: 8/10 — Conceptually clear, but requires discipline and judgment about update magnitudes; easier with practice

**Proven ROI**: 9/10 — Good Judgment Project demonstrated dramatically improved forecasting accuracy; widely adopted in intelligence community, forecasting platforms

**Novelty**: 8/10 — Validates Bayesian reasoning in practical form; counterintuitive that frequent small updates beat "wait for clarity" approach

**Cross-Domain Applicability**: 9/10 — Applies to any domain with uncertainty and evolving information: business, engineering, investing, research, personal decisions

**Total Score**: 44/50 (Tier 1: Canonical)