notque/vexjoy-agent / skills/business/business-ops
business-ops
FlaggedBusiness operations: strategy, technology, growth, competitive intelligence, support, finance, HR, legal, operations, sales, productivity, product management.
references/estimation-techniques.md
View on GitHub---
title: Estimation Techniques — Cone of Uncertainty, Delphi Method, Reference Class Ratios, Planning Fallacy Mitigation
domain: project-evaluation
level: 3
skill: project-evaluation
---
# Estimation Techniques Reference
> **Scope**: Advanced estimation techniques for project evaluation — cone of uncertainty calibration, Delphi aggregation, reference class forecasting with empirical ratios by project type, decomposition patterns, and planning fallacy mitigation. Extends roi-frameworks.md (which covers three-point estimation and T-shirt sizing basics) with deeper methodology and calibration data.
> **Version range**: Framework-agnostic — applicable to software projects, content initiatives, and infrastructure work. Empirical ratios below are from software engineering research; validate against your team's historical actuals.
> **Generated**: 2026-04-09 — planning fallacy research is stable; specific overrun ratios should be replaced with your team's historical data as soon as you have 5+ comparable projects.
---
## Overview
Estimation is a skill that degrades without deliberate practice. Teams that never compare estimates to actuals never improve. Teams that estimate only at the start of projects and never update mid-project confuse commitment with prediction. The techniques here address the two most common estimation failures: the planning fallacy (systematic underestimation of effort and time) and scope uncertainty (not knowing what you're estimating). The cone of uncertainty makes scope uncertainty explicit. Reference class forecasting combats the planning fallacy with data. The Delphi method aggregates expert estimates without anchoring bias. Together they produce estimates that are calibrated — honest about what they know and what they don't.
---
## Cone of Uncertainty
The cone of uncertainty describes how estimation accuracy changes over project lifecycle. From original NASA and software engineering research.
### Cone Calibration by Project Phase
| Phase | Cone Width (high end) | Cone Width (low end) | Typical Accuracy Achievable |
|-------|-----------------------|----------------------|-----------------------------|
| Concept / Initial idea | 4× | 0.25× | "This is a 40-160 hour project" |
| Requirements defined | 2× | 0.5× | "This is a 60-120 hour project" |
| Architecture designed | 1.5× | 0.67× | "This is a 70-105 hour project" |
| Detailed design complete | 1.25× | 0.8× | "This is a 75-95 hour project" |
| Coding underway (50%) | 1.1× | 0.9× | "This is an 81-89 hour project" |
**What this means in practice**: At the concept stage, providing a point estimate (e.g., "this will take 80 hours") is an illusion of precision. The only honest estimate at concept stage is a range: "this could take anywhere from 40 to 320 hours depending on what we find during requirements." Stakeholders who demand point estimates at concept stage are asking for false confidence.
### Cone of Uncertainty Template
```
## Estimation: [Project Name]
## Current project phase: [Concept / Requirements / Architecture / Design / Coding]
### Current phase cone multipliers
High estimate multiplier: ___× (from table above)
Low estimate multiplier: ___× (from table above)
### Base estimate (from three-point or T-shirt sizing in roi-frameworks.md)
Point estimate: ___ hours
### Cone range
High end: ___ hours × ___× = ___ hours
Low end: ___ hours × ___× = ___ hours
### Communication to stakeholders
"Based on current information, this project will take between ___ and ___ hours.
We are at [phase], so the range will narrow to ___ to ___ hours once [next phase deliverable] is complete.
Requesting budget commitment up to the high end until design is complete."
### Phase-gate commitment schedule
| Phase Gate | Date | Commit to | Range Expected |
|-----------|------|-----------|----------------|
| Architecture done | ___ | Budget range | ___×-___× |
| Detailed design done | ___ | Final estimate | ±25% |
| Coding starts | ___ | Hard deadline | ±10% |
```
---
## Reference Class Forecasting with Empirical Ratios
Standard three-point estimation (in roi-frameworks.md) asks estimators to imagine the pessimistic scenario from their own experience. Reference class forecasting bypasses imagination and uses actual historical data from comparable projects.
### Empirical Overrun Ratios by Project Type
From software engineering research (Kahneman/Lovallo, Flyvbjerg, and industry surveys):
| Project Type | Median Planned Hours | Median Actual Hours | Typical Overrun Ratio | Range |
|-------------|---------------------|--------------------|-----------------------|-------|
| New feature, known tech stack | 100 | 130 | **1.3×** | 1.1-1.8× |
| New feature, partially known stack | 100 | 160 | **1.6×** | 1.2-2.5× |
| Integration with external API/vendor | 100 | 200 | **2.0×** | 1.3-4.0× |
| Data migration | 100 | 250 | **2.5×** | 1.5-6.0× |
| Security / compliance implementation | 100 | 220 | **2.2×** | 1.4-5.0× |
| Major refactor / rewrite | 100 | 300 | **3.0×** | 1.5-8.0× |
| New product / greenfield | 100 | 350 | **3.5×** | 2.0-10× |
| Infrastructure migration | 100 | 280 | **2.8×** | 1.5-7.0× |
| UI/UX redesign | 100 | 200 | **2.0×** | 1.3-4.0× |
| Performance optimization | 100 | 180 | **1.8×** | 1.2-4.0× |
**Critical note**: These are industry medians. Your team's actual ratios will differ. Build your own reference class table after completing 5+ comparable projects. Your historical data is more accurate than industry medians for your specific team.
### Reference Class Forecasting Template
```
## Reference Class Forecast: [Project Name]
## Project type: [from table above]
### Step 1: Classify project type
Match to closest project type from table:
Project type: ___
Industry median overrun ratio: ___×
### Step 2: Find comparable past projects (your team's data)
If 3+ comparable projects exist in team history, use these instead of industry median.
| Past Project | Category | Planned Hours | Actual Hours | Overrun Ratio |
|-------------|----------|---------------|--------------|---------------|
| ___ | ___ | ___ | ___ | ___× |
| ___ | ___ | ___ | ___ | ___× |
| ___ | ___ | ___ | ___ | ___× |
| Team median overrun ratio | | | | **___×** |
Use team median if N ≥ 3. Fall back to industry median if N < 3.
### Step 3: Apply reference class adjustment
Bottom-up estimate (from three-point or detailed breakdown): ___ hours
Reference class ratio applied: × ___
Reference class forecast: ___ hours
### Step 4: Inside view adjustment (optional, bounded)
Is this project materially different from the reference class? If so, adjust ratio ±20% max.
Adjustment reason: ___
Adjusted ratio: ___×
Final forecast: ___ hours
### Honest communication
Bottom-up estimate: ___ hours
Reference class forecast: ___ hours
Recommended budget: ___ hours (use reference class forecast, not bottom-up)
```
---
## Delphi Method for Team Estimation
The Delphi method aggregates estimates from multiple people without anchoring. Use when a project involves significant uncertainty and has multiple people who could estimate it.
### Standard Delphi Process
```
## Delphi Estimation Session: [Project Name]
### Prerequisites
- 3-5 estimators with relevant domain knowledge
- Written project description available (no verbal briefing that could anchor)
- Estimators work independently — no pre-session discussion of estimates
### Round 1: Individual estimates (silent)
1. Distribute project description to all estimators
2. Each estimator provides: low estimate, most likely estimate, high estimate, and 1-2 assumptions
3. Collect all estimates before anyone sees others' estimates
Round 1 results:
| Estimator | Low | Most Likely | High | Key Assumption |
|-----------|-----|------------|------|----------------|
| A | ___ | ___ | ___ | ___ |
| B | ___ | ___ | ___ | ___ |
| C | ___ | ___ | ___ | ___ |
### Round 2: Reveal and discuss outliers
1. Share all estimates anonymously
2. Focus discussion ONLY on extreme outliers (highest and lowest)
3. Goal: understand which assumptions caused divergence, not to converge by social pressure
4. No one is required to change their estimate
Discussion record:
- Highest estimate assumed: [what extra scope or risk did this estimator see?]
- Lowest estimate assumed: [what simplification did this estimator assume?]
- Key disagreement: [the specific assumption that caused spread]
### Round 3: Revised estimates (with new information)
Each estimator may revise, or maintain their Round 1 estimate.
Round 3 results:
| Estimator | Revised Most Likely | Change from Round 1 |
|-----------|--------------------|--------------------|
| A | ___ | +/- ___% |
| B | ___ | +/- ___% |
| C | ___ | +/- ___% |
### Aggregation
Method 1 (simple): Average of Round 3 most-likely estimates: ___ hours
Method 2 (PERT weighted): (Low_avg + 4 × ML_avg + High_avg) / 6 = ___ hours
Use method 2 when range is wide (High > 2× Low).
### Residual disagreement
If estimators still diverge by > 50% after Round 3:
- The disagreement IS the finding — the project is not scoped clearly enough to estimate
- Action required: Clarify scope before estimation proceeds
- Do not average through large disagreements; resolve them
```
---
## Decomposition-Based Estimation
When top-down estimation is unreliable (first time doing this type of project), break the project into components and estimate each independently. Research shows decomposed estimates are 30-40% more accurate than holistic estimates.
### Work Breakdown Structure (WBS) Template
```
## WBS Estimate: [Project Name]
### Decomposition rules
- No component should be larger than 40 hours (if it is, decompose further)
- Estimate each component with the person most likely to do the work
- Include explicit non-coding components (design, testing, documentation, review)
### WBS
| Component | Sub-component | Responsible | Low (hrs) | ML (hrs) | High (hrs) | PERT |
|-----------|--------------|-------------|-----------|----------|------------|------|
| Design | UI mockups | ___ | ___ | ___ | ___ | calc |
| Design | API contract | ___ | ___ | ___ | ___ | calc |
| Backend | Data model | ___ | ___ | ___ | ___ | calc |
| Backend | API endpoints | ___ | ___ | ___ | ___ | calc |
| Backend | Integration X | ___ | ___ | ___ | ___ | calc |
| Frontend | Component A | ___ | ___ | ___ | ___ | calc |
| Frontend | Component B | ___ | ___ | ___ | ___ | calc |
| Testing | Unit tests | ___ | ___ | ___ | ___ | calc |
| Testing | Integration tests | ___ | ___ | ___ | ___ | calc |
| Testing | QA pass | ___ | ___ | ___ | ___ | calc |
| Documentation | Runbook | ___ | ___ | ___ | ___ | calc |
| Review | Code review | ___ | ___ | ___ | ___ | calc |
| Review | Security review | ___ | ___ | ___ | ___ | calc |
| Overhead | Coordination | ___ | ___ | ___ | ___ | calc |
| **TOTAL** | | | **___** | **___** | **___** | **___** |
PERT per component: (Low + 4×ML + High) / 6
```
**Overhead budget**: Add 15-20% of total PERT estimate for coordination, unplanned interruptions, and administrative overhead. This is separate from the PERT range.
```python
# PERT calculator for WBS components
components = [
("UI mockups", 4, 8, 16),
("API contract", 2, 4, 8),
("Data model", 4, 8, 20),
# (name, low, most_likely, high)
]
total_pert = 0
for name, lo, ml, hi in components:
pert = (lo + 4*ml + hi) / 6
total_pert += pert
print(f"{name}: {pert:.1f} hours")
print(f"\nTotal PERT: {total_pert:.0f} hours")
print(f"With 15% overhead: {total_pert * 1.15:.0f} hours")
print(f"Reference class check: multiply by your team's overrun ratio")
```
---
## Planning Fallacy Mitigation Techniques
The planning fallacy is systematic. It cannot be eliminated by "trying harder." These techniques structurally counteract it.
### Mitigation Technique Reference
| Technique | When to Apply | Effort Required | Effectiveness |
|-----------|--------------|-----------------|---------------|
| Reference class forecasting | All projects M or larger | Low | High — single highest-impact technique |
| Delphi estimation | Projects with 3+ potential estimators | Medium | High — eliminates anchoring and social convergence |
| WBS decomposition | Any project with unclear scope | High | High — decomposition reduces blind spots |
| Pre-mortem on estimate | Any project before commitment | Low | Medium — catches obvious optimistic assumptions |
| Separate optimistic/pessimistic estimators | Projects with schedule pressure | Low | Medium — explicit adversarial roles |
| History comparison requirement | Before finalizing estimate | Low | High — forces acknowledgement of past performance |
| Commitments after design, not before | Project planning | Low organizational cost, high political cost | Very high — reduces commitment-stage overconfidence |
### Estimation Pre-Mortem Template
Run this on any estimate before it is committed to stakeholders.
```
## Estimation Pre-Mortem: [Project Name]
## Estimate being reviewed: ___ hours / ___ weeks
### Setup
"Assume this project has been delivered and it took 3× longer than estimated.
It is not over budget because of bad luck — the estimate was wrong.
Why was the estimate wrong?"
### Round 1: Reasons estimate might be wrong
| Reason | Probability (Low/Med/High) | Hours Impact | Detection Signal |
|--------|---------------------------|--------------|-----------------|
| Scope expanded during implementation | ___ | +___ hrs | Scope change requests within first 2 weeks |
| Integration complexity higher than expected | ___ | +___ hrs | First integration takes 2× as long as estimated |
| Team member unavailable (illness, competing priority) | ___ | +___ hrs | Any team member goes part-time in first month |
| Key technical assumption proved wrong | ___ | +___ hrs | First proof-of-concept takes longer than 1 day |
| External dependency delayed | ___ | +___ hrs | Dependency misses first checkpoint |
| Quality issues require rework | ___ | +___ hrs | First code review requires >10% rewrite |
### Round 2: Adjustments
Does any High-probability row change the estimate?
- [ ] Yes — revised estimate: ___ hours
- [ ] No — estimate stands with risk mitigations noted
### Risk mitigations added
[List specific actions added to the project plan based on the pre-mortem]
```
---
## Estimation Calibration Tracking
Improve over time by tracking actuals against estimates.
```
## Team Estimation Calibration Log
| Project | Date | Type | Initial Estimate (hrs) | Actual (hrs) | Ratio | Notes |
|---------|------|------|----------------------|--------------|-------|-------|
| ___ | ___ | ___ | ___ | ___ | ___× | ___ |
| ___ | ___ | ___ | ___ | ___ | ___× | ___ |
### Quarterly calibration review
- Average overrun ratio: ___×
- Worst overrun: ___× (project: ___)
- Best accuracy: within ___% (project: ___)
- Estimation technique used for most accurate projects: ___
- Most common source of underestimation: ___
### Calibration update
Apply this ratio to future estimates of the same type:
Team ratio (integration work): ___×
Team ratio (new features): ___×
Team ratio (greenfield): ___×
```
**Target**: After 10 projects, your team's calibration ratio should be between 1.1× and 1.4×. If it is consistently above 1.5×, systematic structural changes are needed (scoping discipline, WBS decomposition requirement, or reference class forecasting mandate).
---
## Patterns to Detect and Fix
<!-- no-pair-required: section header, not an individual failure mode -->
### Point estimate at concept stage
**What it looks like**: "We'll budget 200 hours for this project." — said before requirements exist.
**Why wrong**: At concept stage, cone of uncertainty is 4×. A 200-hour estimate could legitimately be anywhere from 50 to 800 hours. Committing to 200 hours is not an estimate — it is a wish.
**Do instead**: Provide a range at concept stage: "Based on similar projects, this is 100-400 hours. We need requirements before we can narrow it." If stakeholders demand a point estimate, give them the reference class upper end as the planning figure, not the midpoint.
**Fix**: At concept stage, provide a range. "Based on similar projects, this could be 100-400 hours. We need requirements defined before we can narrow this." If stakeholders insist on a point estimate, provide the reference class upper end as the planning figure.
### Averaging through expert disagreement
**What it looks like**: Three estimators give 40, 200, and 350 hours. Team averages to 197 and moves on.
**Why wrong**: 5× divergence between estimators means they are estimating different projects. Averaging does not resolve the disagreement — it obscures it. The 350-hour estimator saw something the 40-hour estimator did not.
**Do instead**: Before averaging Delphi estimates, require all estimates in the same round to be within 2x of each other. If they are not, stop and facilitate a scope clarification discussion. The divergence is the signal worth investigating.
**Fix**: Before averaging Delphi estimates, require that estimates within the same round are within 2× of each other. If they aren't, facilitate a scope clarification discussion first.
### Historical data ignored because "this project is different"
**What it looks like**: Team has a 2.5× overrun history on integration projects. New integration project is estimated without reference class adjustment because "this integration is simpler."
**Why wrong**: "This one is different" is the most common entry point for planning fallacy. The outside view exists precisely to counteract the inside view's over-confidence in project uniqueness.
**Do instead**: Apply reference class adjustment before finalizing any estimate. Name the 3 most similar past projects and their actual outcomes. Reductions to the historical ratio are capped at 20% and require written justification.
**Fix**: Reference class adjustment is mandatory before finalizing any estimate. If the estimator believes the project warrants an adjustment, they may reduce the reference class ratio by up to 20% — with a written justification. More than 20% reduction requires review by a second estimator.
### Estimation accuracy measured at delivery
**What it looks like**: Team compares initial estimate to final delivery date — and it's always wrong. No intermediate comparison, no component-level tracking.
**Why wrong**: By measuring only at the end, you cannot improve estimation technique. You don't know whether the estimate was wrong because of scope change, wrong assumptions, or poor technique.
**Do instead**: Capture estimates at three points: initial (concept), post-design, and actual. Compare initial-to-post-design separately from post-design-to-actual. Each comparison diagnoses a different failure mode and requires a different correction.
**Fix**: Track estimates at three points: initial estimate, post-design estimate, and actual. Compare initial-to-post-design (scope discovery accuracy) separately from post-design-to-actual (execution accuracy). They require different improvements.
---
## Detection Commands Reference
```bash
# Calculate PERT estimate for a set of components
python3 -c "
components = [
('Component A', 4, 8, 16),
('Component B', 8, 20, 40),
('Component C', 2, 4, 12),
# (name, optimistic, most_likely, pessimistic)
]
total_pert = 0
total_variance = 0
for name, o, m, p in components:
pert = (o + 4*m + p) / 6
sigma = (p - o) / 6
total_pert += pert
total_variance += sigma**2
print(f'{name}: {pert:.1f}h (σ={sigma:.1f})')
import math
sigma_total = math.sqrt(total_variance)
print(f'Total PERT: {total_pert:.0f}h')
print(f'80% confidence range: {total_pert - sigma_total:.0f}-{total_pert + sigma_total:.0f}h')
print(f'95% confidence range: {total_pert - 2*sigma_total:.0f}-{total_pert + 2*sigma_total:.0f}h')
"
# Reference class ratio calculator from your team's history
python3 -c "
history = [
('Project A', 100, 150), # (name, planned, actual)
('Project B', 80, 200),
('Project C', 120, 130),
]
ratios = [actual/planned for _, planned, actual in history]
avg_ratio = sum(ratios) / len(ratios)
print(f'Overrun ratios: {[f\"{r:.2f}x\" for r in ratios]}')
print(f'Average ratio: {avg_ratio:.2f}x')
print(f'For a 100-hour estimate, budget: {100 * avg_ratio:.0f} hours')
"
# Cone of uncertainty range calculator
python3 -c "
phases = {
'concept': (4.0, 0.25),
'requirements': (2.0, 0.5),
'architecture': (1.5, 0.67),
'design': (1.25, 0.8),
'coding_50pct': (1.1, 0.9),
}
estimate = 100 # replace with your estimate
phase = 'requirements' # replace with current phase
high_mult, low_mult = phases[phase]
print(f'Phase: {phase}')
print(f'Point estimate: {estimate}h')
print(f'Range: {estimate * low_mult:.0f}h - {estimate * high_mult:.0f}h')
"
```
---
## See Also
- `roi-frameworks.md` — T-shirt sizing, basic three-point estimation, story points, and ROI calculation
- `feasibility-scoring.md` — confidence-adjusted feasibility assessment using estimation data
- `skills/business/business-ops/references/risk-assessment.md` — probability-weighted outcomes when estimate uncertainty is high
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| planning-and-task-breakdownaddyosmani/agent-skillsBreaks work into ordered tasks. Use when you have a spec or clear requirements and need to break work into implementable tasks. Use when a task feels too large to start, when you need to estimate scope, or when parallel work is possible. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| screenpipethedotmack/claude-memSearch screen and audio history recorded by Screenpipe and optionally save selected context into claude-mem. Use when the user asks to recall desktop activity, a meeting, or something seen in another app through Screenpipe. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |
| contact-widgetnexu-io/open-designSelf-contained floating chat widget with welcome screen, social links, meeting button, and message input. Single HTML file, zero dependencies. | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |
| meeting-notesnexu-io/open-designMeeting notes page — title bar with attendees, agenda checklist, decisions block, action items table with owners + dates, and a "next meeting" footer. Use when the brief mentions "meeting notes", "minutes", "1:1 notes", "all-hands recap", or "会议纪要". | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |