Wick — Base Rate

Tetlock-grade discipline applied to any probability claim. The single biggest failure mode in forecasting is estimating from scratch instead of adjusting from the base rate.

When to invoke

• Before /calibrate on any resolvable prediction
• When you catch yourself saying "I think it's about X%"
• When someone else quotes a probability that feels anchored to vibes
• During /decide when scenarios get probability weights

What Wick does

1. Parse the claim
2. Identify the reference class — the set of similar events this is a member of
3. Estimate the base rate — what fraction of events in that class went the predicted way?
4. Identify adjustment factors — what about this case moves the probability up or down from the base?
5. Arrive at: base rate × adjustment, not a free-floating intuition
6. Log reasoning to memory/predictions.md if paired with /calibrate

Example

Claim: "I'd put this seed-stage startup's chance of a meaningful exit at 40%." Reference class: seed-stage B2B SaaS in this sector, 2015-2024 Base rate: ~10-15% reach any exit; ~2-4% reach a meaningful exit Adjustments:

• +3-5% — strong technical founding team with prior exit
• -2% — crowded market, strong incumbents
• +1-3% — existing paying customers

Adjusted estimate: ~4-8%, not 40%.

The gap between gut-40% and base-rate-anchored-5% is the entire value of the skill.

Output format

Reference class: [the class — be specific]
Base rate: [X% — cite source if available, flag if estimated]
Adjustments:
  + [factor 1]: +/-Y%
  - [factor 2]: +/-Y%
Adjusted estimate: ~Z%
Confidence in the base rate itself: [low / medium / high]
Sanity check: does Z% feel wildly different from initial gut estimate? If yes, investigate why.

Framework grounding

Tetlock's superforecasting research. Method of Multiple Explanations — consider alternate reference classes; don't anchor on the first one that fits. Gate 2 (Assent) — the felt probability is an impression; test it before accepting it.

Anti-patterns Wick will flag

• "Inside view only" — estimating from the features of this case without checking the class
• Hand-waved base rate — "probably most of them..." without actually looking it up
• Reference class too narrow — "similar startups founded on a Tuesday in my city"
• Reference class too broad — "all startups" when "seed-stage B2B in this sector" is available