Conditional Probability

Conditional probability is the probability of an event given specified information. That information need not describe an earlier event. Two events are independent when learning that one occurred does not change the probability of the other; otherwise, they are dependent.

Suppose the last three times we went for ice cream, I chose vanilla. Does that establish that vanilla is my favorite, or that I will always choose it? First check the conditions. Could I choose from 100 flavors each time, or had the people ahead of me already taken all the chocolate?

If earlier choices reduce what remains, the probability of my choosing vanilla depends on those choices. Even when every flavor is available, a few observations do not establish that I will always choose the same one. The point is to identify what information a probability estimate assumes.

The source then turns to fat-tailed distributions. The bell curve is often called a “normal distribution.” It gives very little probability to outcomes far from its center. It has no mathematical hard cap: extreme values remain possible in the model, although their probabilities decline rapidly. Fat-tailed distributions assign comparatively more probability to extremes. They need not all have the same shape, and some lack a finite mean or variance.

Height and wealth illustrate why the distinction matters. Human height has biological constraints; wealth can vary over much larger multiples. Someone may have ten, 100, or 10,000 times another person’s wealth. In a distribution with a heavy upper tail, a few very large observations can strongly affect the average. The most common outcome may therefore be a poor guide to the total consequences.

Consider a hypothetical comparison: last year in your country, 1,000 people died after falling on stairs and 500 died in terrorist attacks. Those counts describe that year. They do not, by themselves, establish which risk has the larger expected toll in the future.

Always be extra mindful of the tails: They might mean everything.

The source uses this example to question comparisons based only on recent death counts.3 A rare event involving many deaths could change the comparison substantially. But identifying a possible catastrophe does not establish its probability, and neither type of risk is adequately described just by assigning it a distribution label. Exposure, the range of plausible outcomes, and uncertainty in the available data all matter.

Planning for tail risk means considering whether an unusually severe event would overwhelm a system. Reserves, redundancy, and limits on exposure can help when the timing and size of such events are uncertain. They do not require pretending to know which extreme event will occur next.

Finally, the source uses “metaprobability” to discuss uncertainty about probability estimates themselves. A forecast can sound precise even when its assumptions are poorly supported. To illustrate overconfidence, the source describes investors advertising annual returns of 20% to 40%, then contrasts those promises with a long-run US market return it summarizes as 7% to 8% per annum. Such comparisons depend on the period, inflation adjustment, dividends, and fees. A useful assessment asks how the forecast was made and how well similar forecasts have performed, rather than relying on the presenter’s confidence.