SUPERFORECASTING APPLIED TO EVENT MARKETS
Plain English
Superforecasting (from Philip Tetlock's research) is a set of techniques that consistently produce better probability estimates than experts. These techniques apply directly to prediction market trading: start with base rates, update incrementally on new data, avoid dramatic over-updating, and decompose complex questions into smaller ones you can actually estimate.
Going deeper
Superforecasting techniques for prediction markets: (1) Fermi Estimation — break complex questions into estimable components. 'Will CPI exceed 3%?' → estimate the contribution from housing, energy, food, and core services separately, then aggregate. (2) Base Rate First — before analyzing any specific situation, establish the historical frequency. (3) Inside View + Outside View Balance — consider both the specific current situation (inside view) and the historical base rate class (outside view). (4) Granular Probability Updates — update in small increments (5-10%) as new data arrives; avoid dramatic swings unless the new information is truly decisive. (5) Pre-mortem Analysis — before trading, ask 'if I'm wrong, what would that scenario look like?' Forces you to consider disconfirming scenarios. (6) Seek disconfirming evidence — actively look for reasons your thesis is wrong. (7) Track record maintenance — the only way to know if you're a good forecaster is data over many predictions.
Examples
Fermi Decomposition Example
CPI contract: 'Will CPI exceed 3.2%?' Decompose: Shelter inflation (33% weight, currently 3.5%) → ~1.16% contribution. Energy (7% weight, crude oil up 5% MoM) → ~0.35% contribution. Food (15% weight, stable) → ~0.45%. Core services (45% weight, 3.5%) → ~1.58%. Sum: ~3.54%. Contract priced at 40% Yes. Your estimate suggests it should be ~65% Yes. You buy.
Incremental Updating
Fed rate cut contract: Your prior = 30%. New data: Jobs report weak (update +8% → 38%). Fed speaker sounds dovish (update +5% → 43%). Inflation data comes in low (update +7% → 50%). Each new piece of data moves the needle a measured amount — no single data point should move your estimate from 30% to 80% unless it's completely decisive.