Probability as Uncertainty-Reducing Tool

Probability as Uncertainty-Reducing Tool

Before the 17th century, ‘probable’ could mean an opinion or action reasonable people would approve. Mathematical treatments of probability grew from attempts to analyze games of chance in the 16th century.2 Early work focused on discrete outcomes; later developments, including work by Pierre-Simon Laplace, extended its reach. Andrei Kolmogorov supplied the axiomatic foundation of modern probability theory.3

Bayesian and frequentist approaches interpret probability and statistical inference differently. Bayesian inference represents uncertainty about an unknown quantity with a probability distribution. It combines a prior distribution with a model of the observed data to obtain an updated, or posterior, distribution. Named after Thomas Bayes, the approach uses Bayes’ theorem to relate conditional probabilities. A prior may reflect earlier evidence, expert judgment, or an explicit modeling convention; it is not necessarily a personal hunch. With sparse data, the prior can have a large influence, so its choice needs scrutiny.

Frequentist inference treats probabilities in terms of frequencies under repeated trials or sampling. It uses data and a statistical model to estimate quantities and assess the behavior of a method across repetitions. Suppose 2% of publicly traded companies face securities class action litigation in an average year. That suggests a 2% rate for the group, but an individual company’s risk may differ with its circumstances. Historical frequencies are useful only insofar as the observations and model are relevant to the case being considered.

Both approaches use data and assumptions. Neither is automatically preferable because records are plentiful or scarce. With little information, either approach can be highly uncertain; with stable, informative data, their estimates may be similar. The practical questions are whether the model fits the problem, what assumptions drive the estimate, and how well it performs when checked against evidence. Weather forecasting illustrates the task: observations and models support estimates of future temperature or rainfall, rather than certainty about what will happen.

Business analysts commonly distinguish three kinds of probability:

• Marginal probability concerns an event without conditioning on another variable. An insurer might estimate the probability that an automotive claim exceeds $10,000 across its portfolio.
• Joint probability concerns events occurring together, such as an accident happening in poor weather and involving multiple vehicles.
• Conditional probability concerns an event given specified information, such as the probability of an accident given poor weather or speeding.

A frequency distribution summarizes how often observed values occur, such as the costs of past claims. A probability distribution describes possible outcomes and their probabilities under a model. Either can help an analyst examine the spread of outcomes instead of relying on a single average. Percentiles can make that spread easier to communicate.4 Turning a historical distribution into a forecast still requires assumptions about how well the past represents the future.