Extremizing Forecasts

When multiple forecasters independently give similar predictions, their agreement contains valuable information that averaging misses. Moving the combined probability toward 0% or 100% (called extremizing) often yields more accurate forecasts than simple averaging. This happens because each forecaster typically has private information and tends to be cautious—when they all lean the same way, it's likely the truth is more extreme than any individual claims. Try this: When you see agreement among independent experts, consider whether the truth might be more extreme than the consensus suggests.

Extremizing works by adjusting pooled forecasts toward certainty when forecasters show agreement. The process first combines predictions, then applies a transformation function that moves probabilities away from 50% (or the base rate) and toward 0% or 100%. The degree of extremizing depends on how similar the forecasts are—greater agreement justifies more extremizing. Think of it this way: each forecaster has access to some shared evidence and some unique private information. When forecasts cluster together, it suggests the private evidence is also aligned, strengthening the overall signal beyond what any single person observes.

In a business context, imagine five product managers each estimate a 65% chance that a new feature will succeed. A simple average stays at 65%, but their agreement suggests stronger evidence. Each manager has different customer feedback and market data—their convergence indicates multiple independent lines of evidence point the same way. Similarly, in finance, if several analysts independently rate a stock a "buy" with similar conviction, extremizing suggests a stronger buy signal than any individual rating implies.

The key insight is that individual forecasters are typically cautious, especially with limited information. They naturally hedge toward uncertainty. When multiple cautious minds independently arrive at similar conclusions, the accumulated evidence is often stronger than any single analysis captures. However, extremizing only helps when forecasters are truly independent—if they're all reading the same report or influenced by the same leader, their agreement is illusory and extremizing will backfire. Check this: Before adjusting forecasts toward extremes, verify that forecasters worked independently and didn't just share the same information sources.

Research on extremizing traces back to work on aggregating probability judgments, showing that extremized aggregates often outperform simple averages and even individual experts. The key theoretical foundation is that forecasters typically have different information and process evidence somewhat independently, so agreement among them indicates a stronger signal than any single person's assessment suggests.

• Baron et al. (2014) found that extremized aggregation methods consistently outperformed simple averaging across multiple prediction tournament datasets [2].
• Mannes et al. (2014) showed that the "crowd" within a single expert (combining multiple estimates from the same person) benefits from extremizing, suggesting the phenomenon operates even within individuals [3].

The mathematical justification stems from treating each forecaster's prediction as a noisy signal of an underlying truth. When signals are independent and agree, the optimal statistical combination is more extreme than the arithmetic average. The extremizing function typically uses a power transformation or logistic adjustment calibrated to historical performance.

Extremizing assumes forecasters are genuinely independent—correlated errors from shared information can amplify mistakes. Over-extremizing is a real risk; the optimal degree varies by domain and forecaster expertise. Some research suggests extremizing helps most for probability forecasts near 50% but offers less benefit near the extremes where forecasters are already confident. Additionally, extremizing doesn't help when forecasters systematically disagree or have opposing biases—the technique works best with consensus, not controversy.