Skip to main content

​
Overview

Predictor Hand is an AI forecasting engine inspired by superforecasting principles. It collects signals, builds reasoning chains, makes calibrated predictions, and rigorously tracks accuracy over time. Category: Data
Icon: 🔮

​
What It Does

1

Collect Signals

Gather data from news, social media, financial markets, and academic sources
2

Build Reasoning Chains

Apply base rates, weigh evidence for/against, identify key assumptions
3

Make Predictions

Generate specific, falsifiable predictions with calibrated confidence levels
4

Track Accuracy

Score predictions when they expire, calculate Brier scores, analyze calibration
5

Generate Reports

Deliver prediction reports with accuracy dashboards and meta-analysis

​
Configuration

​
Prediction Domain

DomainFocus AreasExample Predictions
TechnologyProduct launches, adoption, regulations”GPT-5 will launch before Q3 2026”
Finance & MarketsEarnings, macroeconomics, M&A”S&P 500 above 6000 by year-end”
GeopoliticsElections, treaties, conflicts”UK rejoins EU single market by 2028”
Climate & EnergyEmissions, renewable adoption, policy”Solar reaches 30% of US grid by 2027”
GeneralCross-domain trends”Remote work exceeds in-office by 2025”

​
Forecasting Settings

SettingOptionsDescription
Time Horizon1_week, 1_month, 3_months, 1_yearHow far ahead to predict
Data Sourcesnews, social, financial, academic, allWhat to monitor
Report Frequencydaily, weekly, biweekly, monthlyHow often to generate predictions
Predictions Per Report3, 5, 10, 20Number per report

​
Quality Controls

SettingDescription
Track AccuracyScore past predictions when time horizon expires
Confidence ThresholdMinimum confidence to include (low 20%+, medium 40%+, high 70%+)
Contrarian ModeActively seek counter-consensus predictions

​
Activation

​
Basic Setup

openfang hand activate predictor
Configure your forecasting domain: 
openfang hand config predictor \
  --set prediction_domain="tech" \
  --set time_horizon="3_months" \
  --set data_sources="all" \
  --set report_frequency="weekly" \
  --set predictions_per_report="5" \
  --set track_accuracy="true" \
  --set confidence_threshold="medium"

​
Example Workflow

Domain: Technology
Horizon: 3 months
Frequency: Weekly reports with 5 predictions

> Make predictions about AI model releases in Q2 2026
Predictor Hand will:
  1. Collect signals from tech news, social media, company announcements
  2. Analyze base rates (how often do model releases match predictions?)
  3. Build reasoning chains for/against each prediction
  4. Generate 5 calibrated predictions with resolution criteria
  5. Store predictions in ledger with resolution_date = 2026-06-30
  6. When June 30 arrives, research actual outcomes and score accuracy
  7. Update Brier score and calibration metrics

​
How It Works

​
1. Signal Collection

Executes 20-40 targeted search queries based on domain: Technology signals: 
"AI model release 2026"
"GPT-5 launch date"
"Claude 4 announcement"
"AI regulation breaking news"
"LLM adoption metrics"
Financial signals: 
"S&P 500 forecast 2026"
"Fed interest rate decision"
"tech earnings Q1 2026"
"recession indicator"
"analyst consensus"
For each result:
  • web_search → top results
  • web_fetch → extract claims, data points, expert opinions
  • Tag signals:
    • Type: leading/lagging indicator, base rate, expert opinion, data point, anomaly
    • Strength: strong/moderate/weak
    • Direction: bullish/bearish/neutral
    • Source credibility: institutional/media/individual/anonymous

​
2. Accuracy Review

For predictions where resolution_date <= today:
1

Research Outcome

Search for evidence of what actually happened
2

Score Prediction

Correct, Partially correct, Incorrect, or Unresolvable
3

Calculate Brier Score

(predicted_probability - actual_outcome)^2 where outcome is 0 or 1
4

Update Calibration

Check if your 70% predictions are right ~70% of the time
Example: 
Prediction: "GPT-5 will launch before July 1, 2026"
Confidence: 65%
Resolution Date: 2026-07-01

Actual: GPT-5 launched June 15, 2026
Result: Correct
Brier Score: (0.65 - 1.0)^2 = 0.1225 (good)

​
3. Reasoning Chain Construction

For each potential prediction: 
PREDICTION: GPT-5 will launch before July 1, 2026
CONFIDENCE: 65%
TIME HORIZON: 3 months

REASONING CHAIN:

1. Base rate: OpenAI major releases (GPT-3 → GPT-4 was 16 months)
   - GPT-4 launched March 2023
   - 16-month cycle suggests GPT-5 around July 2024
   - But GPT-4.5 / intermediate releases can delay major versions

2. Evidence FOR (+25%):
   - Sam Altman hinted at "big releases" in Q2 2026 interview
   - OpenAI job postings for "GPT-5 safety team" (leaked, Feb 2026)
   - Compute cluster expansion reported by The Information
   - Typical OpenAI release pattern: announce 2-4 weeks before launch

3. Evidence AGAINST (-10%):
   - No official announcement yet (as of March 6)
   - Safety testing typically takes 6+ months
   - GPT-4.5 Turbo still rolling out features
   - Regulatory scrutiny may delay

4. Net adjustment from base rate:
   - Base: 50% (coin flip without other info)
   - Signals: +25% -10% = +15%
   - Final: 65% confidence

KEY ASSUMPTIONS:
- "GPT-5" refers to a model OpenAI officially calls "GPT-5" (not GPT-4.5, not internal codename)
- "Launch" means public API access (not just announcement or limited preview)

RESOLUTION CRITERIA:
- Check openai.com/blog and OpenAI API docs on July 1, 2026
- If GPT-5 API endpoint is live and publicly accessible → Correct
- If only announced but not launched → Incorrect
- If launched after July 1 → Incorrect

​
4. Cognitive Bias Checks

Before finalizing predictions:
  • Anchoring — Am I fixating on a salient number?
  • Narrative bias — Good story ≠ likely outcome
  • Overconfidence — Are my 90% predictions actually 60%?
  • Base rate neglect — Did I start with historical frequency?

​
5. Contrarian Mode (Optional)

If contrarian_mode = true:
  1. Identify consensus view from collected signals
  2. Search for evidence contradicting consensus
  3. Include at least one counter-consensus prediction per report
Example: 
Consensus: "AI will continue exponential scaling"
Contrarian: "AI scaling will hit diminishing returns by Q4 2026"
Confidence: 35% (minority view, but signals suggest possible)
Reasoning: Chinchilla scaling laws, compute cost curves, plateau in benchmarks

​
6. Report Generation

# Prediction Report: Technology
**Date**: 2026-03-06 | **Report #**: 8 | **Signals Analyzed**: 42

## Accuracy Dashboard
- Overall accuracy: 68% (25 predictions resolved)
- Brier score: 0.18 (lower is better, 0 = perfect)
- Calibration: Well-calibrated (70% predictions correct 72% of time)

## Active Predictions

| # | Prediction | Confidence | Horizon | Status |
|---|-----------|------------|---------|--------|
| P-042 | GPT-5 launches before July 1 | 65% | Jun 30 | Active |
| P-043 | Apple Vision Pro 2 announced WWDC | 80% | Jun 10 | Active |
| P-041 | Llama 4 released | 45% | May 31 | Active |

## New Predictions This Report

### P-042: GPT-5 launches before July 1, 2026
**Confidence:** 65%  
**Resolution Date:** 2026-07-01

**Reasoning Chain:**
[Full reasoning from above]

---

### P-043: Apple announces Vision Pro 2 at WWDC 2026
**Confidence:** 80%  
**Resolution Date:** 2026-06-10

**Reasoning:**
- Base rate: Apple announces hardware at WWDC 40% of time (historically)
- Vision Pro 1 launched Feb 2024, 16-month cycle suggests June 2025
- But Gen 1 adoption slow (reports of <500k units sold)
- Gen 2 would need compelling new features to justify WWDC stage time
- Supply chain leaks show new display tech in production (Kuo, Feb 2026)
- WWDC 2026 theme "Spatial Computing" suggests Vision focus
- Net: 80% confidence (strong signals, fits pattern)

**Resolution:** Check Apple WWDC 2026 keynote for "Vision Pro 2" announcement.

---

## Expired Predictions (Resolved This Cycle)

### P-038: US Congress passes AI regulation by March 1
**Prediction:** 40% confidence  
**Actual:** No bill passed  
**Result:** ✓ Correct (predicted low probability, and it didn't happen)  
**Brier Score:** (0.40 - 0)^2 = 0.16

### P-039: Anthropic raises Series D before Feb 28
**Prediction:** 70% confidence  
**Actual:** Announced Feb 20, $850M  
**Result:** ✓ Correct  
**Brier Score:** (0.70 - 1.0)^2 = 0.09 (excellent)

## Signal Landscape

Key signals this cycle:
- **AI safety regulation:** EU AI Act enforcement begins, US bill stalled
- **Model releases:** Gemini 2.0 Pro launched, Claude 3.5 updated
- **Hardware:** NVIDIA B100 GPUs shipping to hyperscalers
- **Adoption:** Enterprise AI spend up 40% YoY (Gartner)

## Meta-Analysis

Your forecasting strengths:
- **Product launches:** 78% accuracy (strong pattern recognition)
- **Regulatory predictions:** 55% accuracy (high uncertainty, hard to predict)
- **Calibration:** Well-calibrated overall, slight overconfidence on <6 month horizons

Recommendations:
- Lower confidence on short-term regulatory predictions (too many variables)
- Continue current approach for tech product launches
- Consider adding insider network signals for private company predictions

​
Output

FileDescription
prediction_report_YYYY-MM-DD.mdWeekly/monthly prediction report
predictions_database.jsonLedger of all predictions and outcomes

​
Dashboard Metrics

  • Predictions Made — Total predictions ever made
  • Accuracy — Percentage of resolved predictions that were correct
  • Reports Generated — Number of reports delivered
  • Active Predictions — Currently unresolved predictions

​
Prediction Quality

​
What Makes a Good Prediction

Good predictions are:
  • Specific — “GPT-5 will launch before July 1” not “AI will advance”
  • Falsifiable — Clear resolution criteria
  • Calibrated — Honest confidence levels (not always 90%)
  • Timestamped — Exact resolution date
  • Reasoned — Explicit chain of logic

​
Brier Score Explained

Brier score measures prediction accuracy: 
Brier = (predicted_probability - actual_outcome)^2
  • 0.00 — Perfect (predicted 100% and it happened, or 0% and it didn’t)
  • 0.25 — Random guessing (50% confidence on everything)
  • 1.00 — Worst possible (predicted 100% and it didn’t happen)
Goal: Keep Brier score below 0.20 for good forecasting.

​
Tips & Best Practices

Never express confidence as 0% or 100% — Nothing is certain. Use ranges like 5-95%.
For best forecasting:
  • Always start with base rates (historical frequency)
  • Show your work — reasoning chains catch errors
  • Track ALL predictions — don’t selectively forget bad ones
  • Update predictions when new evidence arrives (note updates in ledger)
  • Distinguish predictions (testable) from opinions (untestable)

​
Common Pitfalls

Overconfidence
Most people are overconfident. If you’re above 90% on most predictions, you’re probably overconfident. Narrative bias
A compelling story doesn’t make an outcome likely. Check the base rates. Confirmation bias
Actively search for evidence AGAINST your prediction, not just for it. Anchoring
Don’t fixate on the first number you see. Consider the full range.

​
Advanced Usage

​
Custom Prediction Requests

Predict: Will Rust overtake C++ in the TIOBE index by 2027?

​
Multi-Step Conditional Predictions

If GPT-5 launches in Q2 2026, predict the probability of GPT-6 by Q2 2027

​
Accuracy Analysis

Analyze my prediction accuracy on AI regulation vs product launches

​
Next Steps

Collector Hand

Collect signals for better predictions

Researcher Hand

Deep research on prediction topics